Memia year in review 2024: counting the OOMs to Foom 🧮📈🤯
Man, that’s crazy. Catch the game last night?
Kia ora,
Welcome to my annual end-of-year post trying to round up everything that I’ve covered in Memia in 2024. (You can read my previous annual roundup posts here: 2023 | 2022 | 2021 | 2020).
ℹ️PSA: This post is *far longer* even than Memia’s usual missives.🤯 Best viewed online or in the Substack app. (Even better idea, import into your LLM app of choice — Google’s NotebookLM is very zeitgeisty — and ask it to summarise or make a podcast for you!)
📑Table of contents
Introduction
A daunting task trying to synthesise the biggest year yet in AI and technology.
👤On a personal note
I published my first book, went nomadic, and became an empty nester!
💥The CRISIS continues
Out of focus for this post, but multiple global crises intensified: climate change hit 1.5°C, biodiversity declined, and geopolitical tensions escalated.
🌏⚡1. AI-ocene epoch, Year 2
AI's energy and compute demands are reshaping our planet's surface at an unprecedented scale.
🏃♂️2. Race to AGI - no slowing down
Major AI labs competed fiercely with new models every week, while capital investment continued to pour in… for how much longer, though?
🎨3. Omni-modal AI
AI has mastered multiple forms of content creation: images, video, audio, and text all saw breakthrough capabilities - 3D (4D!) worlds and science are next.
🤖🦾4. Robots, robots, everywhere
Humanoid robots, autonomous vehicles, and drones made significant advances toward real-world deployment.
🔗5. Post-Web3
Crypto hit new highs while decentralised infrastructure is yet to find real-world applications beyond speculation.
🚀6. Upping the pace to space
Space exploration accelerated with SpaceX with a huge lead over all other players entering the field. Datacentres in space may be a thing soon.
🥽7. XR is now “spatial computing”
Apple and Meta still battling for dominance in slow-motion for the emerging
XRspatial computing market.
🛡️8. Tech safety and regulation
Governments struggled to regulate AI while major safety incidents highlighted the risks.
🔮9. Emergent…
Breakthrough technologies on the horizon that could reshape our future.
🧠10. Mind expanding
My 2024 highlights of deep thinkers exploring new frameworks for understanding progress and technology's impact.
😂And then there were memes…
The best of this year’s Memetic Savasana section… AI generation spawning a whole new genre of memes.
💭Takeaway themes
This whole summary is in many ways impressionistic, curating lots of individual data points and signals of the direction things are moving. Distilling into a few key themes in bullet points:
The marginal cost of intelligence keeps falling exponentially.
The rate of frontier intelligence improvement keeps increasing exponentially.
Although scaling “laws” of pre-training appear to be hitting a ceiling, there are many new avenues of scaling being explored.
All of the major labs are racing to reach “AGI”, whatever that means. The whole term is nebulous and is now being decomposed into a raft of specialised intelligence “benchmarks”.
The amount of energy required by AI keeps increasing exponentially.
Capital investment has continued to flow into AI at a remarkable rate in 2024. The industry has many of the hallmarks of the late 90s dotcom bubble… but does the prospect of “AGI” on the horizon make it different this time?
There are lots of new AI models, tools, gadgets and distractions released every week. A few of them are immediately valuable to a large number of users (eg Anthropic Claude, ChatGPT Advanced Voice Mode, Google NotebookLM). Most of them are not.
Spotting the advances in AI which will improve your organisation’s efficiency and competitiveness is a hit-and-miss activity which requires a willingness to invest in high cadence R&D and experimentation. Unfortunately, not many organisations are on this journey…
At an individual productivity level, the generally available large language / multimodal AI models can give you superpowers. Power users who continue integrating these tools into their daily information workflows will likely command a premium in (whatever’s left of) the labour market in the next couple of years.
However, the future of even the most complex cognitive labour markets are rapidly being automated by AI. I’m reminded of the classic Mitchell and Webb BMX Bandit and Angel Summoner skit. In some ways, we are all BMX Bandit now:
Open source AI and decentralised AI are in tension with hegemonic US vs. China bipolar geopolitics - and the militarisation of AI is happening rapidly. The rest of the world are just bystanders unless they can coordinate effectively to counteract this trend.
Existential concerns around AI safety in the near-term and also post-”AGI” abound. What kind of successor for humanity will posthuman AI be? Can we even hope to “align” it to our values and priorities?
These sure are interesting times…
📝Introduction
It’s been a daunting task to take stock and synthesise what’s been the hugest year yet for AI and technology… let alone all the other change that’s happening in the world.
My first impression after going back over 50 weekly newsletters and other posts from 2024 in the Memia Knowledge Graph: it’s all a blur! There has been so much covered that I don’t remember half of it.
(The FOMO has been exacerbated as just this week there have been at least four more paradigm-changing releases which I haven’t managed to get deep into…:
OpenAI o3 reasoning model, released only 2 months after o1-preview. As close to “AGI” as anyone has measured yet.
Google Gemini Flash 2.0 Experimental - a small “workhorse” AI model with the power of GPT-4o and o1-preview
The Genesis Project, an open-source collaboration to create a “generative 4D physics engine”.)
I feel I’ve done as good a job as any one person (+AI) could do to keep up, but it has been literally impossible to manually track and analyse every significant industry development in tech, particularly AI, this year. I observe that the tech industry is moving at a far faster pace than any other part of society and the economy — by definition we need augmentation with AI just to keep up with AI and all current affairs going forward.
In addition, parsing the signal from the noise becomes harder when the noise keeps on getting louder.
Yes, AI keeps advancing at an incredible rate… but so what? So far I don’t think any commentators, myself included, have got a firm handle on the answers to that question. There are many attempts to capture aspects of what is going on, but its is increasingly hard to avoid getting punch-drunk on novel new AI models and capabilities being released week-in, week-out. Standing back and seeing the big picture on AI is still an emergent capability at the end of 2024.
I’ve been working with AI to automate parts of this during the year… and thinking hard about how to deliver a less unwieldy product in 2025. (More on that in January…)
👤On a personal note…
2024 has been a huge year for me both professionally and personally.
I published my first book, ⏩Fast Forward Aotearoa in March! This was the culmination of 2 years of work exploring New Zealand's future with exponential technology, supported by many Memia readers along the journey.
The book was infinitely enhanced in partnership with long-time collaborator and friend Sam Ragnarsson, who rode the bow-wave of what is possible with the latest generative AI image models, creating an amazing set of illustrations imagining possible futures for New Zealand. Just a few of my favourite images here:
Fast Forward Aotearoa art by Sam Ragnarsson CC BY 4.0 Exceptionally grateful to Sam for his skill, speed, curiosity, innovation and friendship throughout this project. Also big thanks also to Jason Lennie at Ledge for all of his design flair and publishing expertise — and exceptional patience accommodating my *non-linear* book-writing workflow!
If you haven’t already, check out the ⏩Fast Forward Aotearoa store to download or purchase a physical copy (and limited edition merch, still a few T-shirts and caps left while stocks last).
I’m filled with gratitude to everyone who has reached out and/or written positive things about the book… I hope it will be a “slow burner” and the key theme of building decentralised, open-source technology capability in a small country will percolate into more mainstream political conversation in time for the next scheduled general election. Contemplating an election year special “2nd edition” in 2026, we’ll see…
Meanwhile my other main endeavour, the weekly Memia newsletters “scanning across the latest in AI, emerging tech and the exponentially accelerating future” have become progressively longer to produce — and hence read!
I’ve been iterating Memia’s mission statement throughout the year… here’s where things have landed so far for 2025:
🔮📈🌱 Navigating exponential intelligence for a better 🌏planet→ 🪐solar system →🌌universe
When I have managed to come up for air, I managed to publish the first in a series of Strategy Notes exploring the convergence of AI and strategy practice… more of these next year.
AI strategy consulting and advisory work — over the last year I’ve worked with leadership teams and boards across many diverse industries including: energy, telecoms, infrastructure, logistics, financial services, health, law, agriculture, education, NGOs and government — as well as many technology firms. Thanks to all of Memia’s clients in 2024 for your engagement, trust and openness to exploring opportunities against what are often challenging prognoses!
Speaking engagements - I gave 16 keynote presentations throughout 2024 covering the rapidly advancing AI landscape, including my first gig in Europe at Nomadfest Switzerland. (Get the most up-to-date AI presentation or keynote for your event in 2025)
I also appeared on a few podcasts, most notably:
The Business Of Tech with Peter Griffin in March: Aotearoa's tech future needs some big structural changes:
BizBytes with Ant McMahon in July:
Also Sam Ragnarsson joined me for a Memia webinar to discuss the AI tools and workflows he uses:
🌍✈️Nomadic life In amongst all of this, once the book publicity was out of the way I managed to fit in a 10-week trip digital nomading around the world, just about keeping on top of work all the way. Amazing times we live in that this is possible…although working across NZ and European time zones is, er, suboptimal!
Nomadfest Switzerland afternoon outing to Champex-Lac - swam in the water which was considerably colder than Greece the previous week! 🍎What goes around…
Apple's Flagship computers... over 40 Years Apart My first computer in my early teens was an Apple IIe on which I first learned to code in Apple BASIC back in the 1980s — lots of PEEKs, POKEs and CALLs! But since I started my professional career in 1994, I’ve always used Microsoft Windows as my main productive OS for 30 years. *Until*… this year I finally gave in and upgraded myself to a new Macbook Air…. the Windows UX has just got clunkier and clunkier and the start menu kept on filling up with junk I didn’t want — but mainly it was the promise of Apple’s new M3 series chips for running open source AI locally which finally brought me across. It works well. (Although the Finder app for file management is a complete mess…)
Empty nest Just to add a bit of colour, this monster year of work coincided with a lot going on for me personally as well. Exactly the same week that the book arrived back from the printers, my partner Anna and I suddenly became empty-nesters as the last of our three adult daughters left home — all within the space of a few months (two joining the Aotearoa exodus to foreign shores in Europe…😥) That was quite a life-changing month indeed!
Super grateful to Anna and my family for all their love and support this year as I finished the Sisyphean task of publishing the book and their patience while I remain glued to my laptop and phone screen all hours of the day…😘🤩🙏
For 2025, my plan is to take a year off book-writing and focus on Memia’s newsletter and advisory services — while developing more AI tools to augment and speed up the strategy and sensing process. Maybe there’ll be room for a creative multimedia project as well… we’ll see.
So…. once again a very sincere note of thanks to all Memia readers for continuing to let me into your inbox this year and receiving my observations, speculations and reckons *at length* — it remains a privilege. In particular my deep appreciation goes out to all Memia paid subscribers — your contribution helps me to put the (ever-increasing!) time into researching and preparing the weekly newsletters … thank you.🙏🙏🙏
All the best for a relaxing break over the festive season, see you in 2025!
Namaste
Ben
💥The CRISIS continues…
In previous end-of-year roundup posts I have leaned extensively into [Polycrisis / Metacrisis / “The CRISIS”] framing… however this year I’m going to de-emphasise that approach and focus in on what has happened in AI and Tech. (Take a read of the Welcome To The Polycrisis chapter in my book for my detailed take on that landscape…) Needless to say I exist more than ever in a state of superposition between climate-biodiversity doom and accelerationist techno-AI-optimism.
Nate Hagens has a pretty well-developed model of the global pecking order of non-biophysical factors which are causing the biophysical CRISIS, with AI right at the top. (20 minutes well spent).
But briefly, multiple dimensions of the *CRISIS* have only exacerbated during 2024:
♨️Climate change
2024 is the first full year since records began when global temperatures exceeded 1.5°C above pre-industrial levels. And the rate of change is speeding up.
COP29, the UN’s annual Climate Summit in gas-rich nation Azerbaijan was comprehensively hijacked by the Fossil Fuel industry. When will it be payback time?
Biodiversity
The planet continues on course for a Sixth Mass Extinction.
On the marginally optimistic side, a groundbreaking 2024 study argues that protecting just 1.2% of the Earth’s land surface area, via the targeted protection of 16,825 sites, could prevent the imminent extinction of thousands of the world’s most threatened species.
COP out:
Likewise the COP16 UN biodiversity summit held in Cali, Colombia— bringing together 190 countries and over 15,000 people around the goal of protecting the world’s flora and fauna… just fizzled out with no agreement on decisive actions.
Other planetary boundaries
A study, Keeping the global consumption within the planetary boundaries found that up to up to 91% of planetary boundary breaching can be attributed to the top 20% of global consumers:
Pandemic risk keeping a constant eye out for H5N1 mutation to enable human-human transmission, which could make Covid look like a walk in the park.
Geopolitics and war
The outcome of Russia’s invasion of Ukraine still seems far off. The war has evolved into a low-intensity proxy conflict between the US and NATO and Russia supported by China, North Korea and Iran. Incoming US president Trump seems to think it will be solved within days of his inauguration… yeah right.
Israel’s disproportionate, ultimately self-defeating war against Palestinians and then Lebanese escalated horrifically in 2024. Apartheid-scarred South Africa laid charges of genocidal acts in Gaza against Israel at the ICJ. The International Criminal Court issued arrest warrants for Israeli leaders Netanyahu and Gallant. But Western mainstream and online media continues to be closely policed by aggressively pro-Israeli interests... meanwhile so many innocent people have been killed and lives continue to be lost:
…And nearby Syria’s heinous Assad dictatorship fell suddenly in just a few days in December, with Israel already getting expansionary… the Middle East in 2025 looks as turbulent as ever.
The US and China escalated their cold war of trade and technology sanctions, particularly on high tech components… while flexing with shows of military tech on both sides. But concerns of a hot war seem a way off still.
Taiwan’s general election in January saw incumbent Democratic Progressive Party (DPP) candidate William Lai emerge as victor in the Presidential race — but the DPP lost its majority in Taiwan’s parliament to the relatively pro-China KMT. The country’s position as one of the most likely next geopolitical hotspot continues to simmer…
European military and intelligence agencies have begun to openly consider the potential of Russia attacking NATO countries by end of decade.
Financial system (in)stability
The global financial system held together during 2024 with no major banking crises like in 2023. However, with global interest rates expected to rise over the next few years, this chart presents a conundrum for the incoming Trump administration (and all other holders of US government debt). With battle lines being drawn right now on whether to remove the US debt ceiling or not, 2025 is shaping up to be a bumpy ride…
🗳️🌍✅Democracy alive and well…just🤞
On a positive note, 2024 was the biggest year ever of national elections.
With such a lot of trepidation at the start of the year, so far national elections have passed relatively peacefully in over 60 countries. In particular, fears of months of escalating disputes over the US Presidential election were allayed when Donald Trump won by a higher than expected margin. The US dodged that proverbial bullet, for now at least. (But frying pan, fire…)
OK here goes…. Memia’s 2024 year in review, 10 themes plus memes….
🌏⚡1. AI-ocene epoch, Year 2
Picking up where we left off last year… my assertion remains that we find ourselves in the AI-ocene Epoch, Year 2:
Viewed through this geological lens, advancing AI technology is continuing to terraform our planet to meet its energy and computation requirements first, above those of humanity’s or nature’s.
📈AI in 2024: all lines go up
There was wider realisation in 2024 that AI is going to require huge amounts more capital, energy and natural resources to meet projected demand.
In October, the Institute for Future Progress (IFP) put out a comprehensive report How to build the future of AI in the United States, highlighting projected future compute requirements to train the very largest “frontier” AI models, which to date all needs to be concentrated into single contiguous clusters rather than geographically distributed:
Which means, AI data centers line go up:
In July, Elon Musk announced that xAI had completed the build of their new Memphis “Colossus” AI training “supercluster” in record time. Typical Musk hyperbole below… at 4:20am! No-one sleeps at xAI…
According to former Stability.ai CEO Emad Mostaque, this is indeed likely to be the world’s fastest (publicly known) supercomputer with 2.5 ExaFLOPs performance:
Top500 via @EMostaque Aerial image here:
AI GPUs: line go up: Meta is also using more than 100,000 Nvidia H100 AI GPUs to train Llama-4:
“Meta isn’t the first company to have an AI training cluster with 100,000 Nvidia H100 GPUs. Elon Musk fired up a similarly sized cluster in late July, calling it a ‘Gigafactory of Compute’ with plans to double its size to 200,000 AI GPUs. However, Meta stated earlier this year that it expects to have over half a million H100-equivalent AI GPUs by the end of 2024, so it likely already has a significant number of AI GPUs running for training Llama 4.“
AI energy consumption line go up: each H100 has peak power consumption of ~700W, so a cluster of 500,000 would require (by my calculations and ChatGPT’s) 350MW of power supply to run - running this continuously for 1 year would use over 3TWh !!! Paul Churnock (ex-Microsoft) concurs:
More concerning: embedded energy in AI use goes all the way back to the chip fab itself: Analyst firm TechInsights calculates that EUV lithography systems consume 1,400 kilowatts per EUV tool… and rising:
“By 2030, the estimated annual electricity consumption for EUV tools alone could exceed 54,000 gigawatts, more than 19 times the amount used by the Las Vegas Strip in a year.“
AI capex line go up: all of the above is driving continued growth in capital investment from the world’s largest technology firms:
AI Investment line go up: at the end of the year, in total, funding of AI and cloud companies outside China in 2024 is estimated at around US$80Bn - up 27% YoY from 2023. This graph from investor Tom Tunguz in May looking spot on:
Much of this was in really chunky rounds:
Databricks secured US$10 billion funding in December
Microsoft-backed OpenAI raised US$6.6 billion in October
Elon Musk's xAI raised $6 billion in May
Anthropic received $4 billion from Amazon
G42 In April 2024, Microsoft invested $1.5 billion in G42, an Emirati AI firm
Safe SuperIntelligence (SSI) co-founded by ex-OpenAI cofounder Ilya Sutzskever raised US$1 billion in July
Perplexity AI in November 2024 the AI-powered search engine, raised $500 million,
In Europe, smaller rounds went to Mistral (June 2024, €600 million/US$645 million), Aleph Alpha (US$641) and DeepL (US$300M)
Blackrock and Microsoft announced a new US$30Bn AI infrastructure investment fund (US$100Bn with debt funding)
There was also some AI industry consolidation… due to antitrust-avoidance there were unconventional “acquihires” of key staff rather than buying the companies themselves.
And some AI companies just ran out of money… most spectacularly Stability AI, creators of Stable Diffusion, where CEO Emad Mostaque took a walk as CEO after trailblazing open-source AI… but effectively giving away all of its core IP with no clear path to profitability. (Compare and contrast Midjourney which has never taken on any investment and continues to operate profitably).
An exponential AI data center buildout looms…
Summarising the trends above: demand for AI compute (for both training and inference) looks set to grow exponentially with no end in sight, despite recent industry chatter about “scaling laws” hitting a wall.
By implication, the IFP report mentioned above correlates this to projections which indicate over 130GW of new AI data center energy consumption by 2030 - more than 3X the total global data center power consumption in 2022:
…But where will all the energy for AI going to come from?
This in turn, will require huge investments in new electricity generation capacity.
However, the challenge for the US in particular is that it has let its power generation capacity stagnate for years, while China has been on a steady build:
Because the current view in the US that it wants to build all of this infrastructure within its own borders for national security reasons (see Situational Awareness below…), by implication, the volume of capital investment into data center and energy infrastructure buildout to support AI, particularly in the US, will be massive. As shown above, the major US tech companies all announced huge capital budgets - and the global data center market size is projected to grow from US$243 billion in 2024 to USD 585 billion by 2032, exhibiting a CAGR of 11.6%, dominated by North America with a share of nearly 40% in 2023.
More in my recent Strategy Note":
Looked at from the outside, there is doubt whether the US can mobilise enough new energy infrastructure to meet this demand.
Challenger xAI drew industry plaudits for the speed with which it stood up its new “Colossus” supercluster in Memphis, Tennessee - but these were powered by polluting methane-gas-burning generators without permits: The power consumption requirements will be massive - one analysis up to 200MW by the end of the year, but currently the grid can only provide 8MW.
Looking closer at the satellite image it looks like they deployed 14 mobile generators on-site which brings current capacity up to around 32MW:
Likely mobile gas generators like this:
Dirty AI
However, the growth in data center emissions are hobbling other larger firms, particularly when aiming to meet “carbon-negative by 2030” public commitments:
❌Google’s carbon emissions soared by 48% since 2019 due to AI according to their 2024 environmental report and the company’s commitment to achieve net-zero by 2030 is in doubt .
❌Microsoft’s missed climate goals despite worthy aims to be *carbon negative* by 2030, Microsoft’s AI data centre buildout is actually increasing carbon emissions - by 30% last year:
Microsoft President Brad Smith pushing the boundaries of credibility when he said:
“We fundamentally believe that the answer is not to slow down the expansion of AI but to speed up the work needed to make it more environmentally friendly…”
(So much cognitive dissonance to deconstruct that sentence when looking at the graph above…)
AI chip startup Groq is partnering with oil producer Aramco to build “the world’s largest AI inferencing center” in Saudi Arabia which initially will have 19,000 of Groq’s language processing units. No guesses what the energy source for that will be.
Cue major announcements from major tech companies seeking to accelerate US renewable generation capacity, particularly nuclear:
Microsoft (September 20th) Three Mile Island nuclear plant will reopen to power Microsoft data centers. In a 20-year power purchase agreement between Microsoft and Constellation Energy, the shuttered plant is expected to reopen in 2028 with analysts estimating Microsoft will pay rates of between US$110-US$115 per megawatt-hour.
Google October 14th) Google signed the world's first corporate agreement to purchase nuclear energy from multiple small modular reactors (SMRs) to be developed by Kairos Power. The initial phase aims to bring Kairos Power's first SMR online by 2030, with additional reactors deployed through 2035, enabling up to 500 MW of new 24/7 carbon-free power to U.S. electricity grids. Kairos Power's technology uses a molten-salt cooling system and ceramic pebble-type fuel, allowing for a simpler, more affordable reactor design.
Amazon (October 16th) Amazon also signed three agreements to develop small modular reactor (SMR) nuclear technology, including leading a US$500 million funding round for X-Energy's SMR development, aiming to bring over 5 gigawatts online in the U.S. by 2039.
Meta (December 3) announced an RFP for nuclear energy developers to target up to 4 gigawatts (GW) of new nuclear generation capacity in the US.
Sidequest: recent nuclear SMR developments
The SMR market is highly speculative - and it will only takes one radiation breach incident for the industry to be set back for years.
One of the most progressed designs is Westinghouse's eVinci microreactor which claims to deliver 5MW of power for up to 100 months, producing 1.2 petawatt-hours of energy — the reactor has few moving parts and functions essentially as a battery:
Oklo, a nuclear power company backed by OpenAI's Sam Altman, saw its shares surge by approximately 150% in response to the Microsoft, Google and Amazon announcements above.
However, we’ve been here before very recently… less than a year ago SMR company Nuscale was forced to cancel a US$600M+, six-reactor, 462 megawatt project in Utah because the target price for power from the plant ($89 per megawatt hour, up 53% from the previous estimate of $58 per MWh) caused customers to drop out. (Timing!)
Also keep an eye out to leftfield - in August, nuclear fusion research startup Helion (another Sam Altman startup — the Ringmaster is *everywhere*…!) received its licence to operate Polaris, the company’s 7th-generation fusion machine:
(You may recall the highly speculative Helion / Microsoft deal to provide 50MW Fusion power by 2028… covered in Memia 2023.19…not holding my breath on that one).
💫OpenStar Here in Aoteoroa, nuclear fusion research company OpenStar Technologies achieved its first plasma. The team used a novel levitated dipole reactor (LDR) design which differs from traditional tokamak or stellarator fusion reactors, using a magnetosphere-like confinement system and levitation to keep plasma within a doughnut-shaped reactor. High-temperature superconductor (HTS) magnets, operating at 50K, create strong magnetic fields up to 20 Tesla. OpenStar aims to begin generating electricity from their reactor by 2030.
Check out this moment when the plasma first appeared:
☀️💨🔋Solar+storage holds hope
Just in case those nuclear power stations don’t come to fruition in time…
Record amounts of renewable energy investment so far in 2024:
There’s a lot of solar plus storage going in:
Case in point: the US state of California has one of the world’s largest grid-scale battery capacities, now powering the state for hours after dark. This chart from April 2024, it’s got even better since:
20,000 Acre solar farms? Casey Handmer, big-thinking physicist and CEO of Terraform Labs explored the land use implications of Solar+Battery to meet AI demand earlier this year: How to Feed The AIs:
“What is this going to look like at scale?
A 1 GW data center (containing roughly a million H100s!) would have a substantial footprint of 20,000 acres, almost all of that solar panels. The batteries for storage and data center itself would occupy only a few of those acres. This is in some sense analogous to a relatively compact city surrounded by extensive farmland to produce its food.“
Work out the economics of that.
BIG cable The Australian government approved the first stage of Sun Cable's Australia-Power Link, set to become the world's largest renewable energy and storage project. Backed by software billionaire Mike Cannon-Brookes, Sun Cable has evolved to include an 800km transmission link to Darwin and up to 24 GW of solar and wind generation with up to 42 GWh of battery storage, eventually to provide up to 4 GW of 24/7 green power. There are also plans for a 4300km (!) subsea cable system to potentially export power to Singapore, Indonesia and other nations in the region.
The environmental impacts of AI
AI environmental damage line go up: The environmental campaigners fighting against data centres:
"What's going to happen if we continue with business as usual is that electrical prices are going to skyrocket for everybody, including the data centre industry - and that's their biggest bill, so that's going to impact them…The water scarcity issue is also going to impact them.“
Unless the hyperscalers discover alternative cooling mechanisms which don’t use water at all, their commitments to being “Water Positive By 2030” (whatever that means, what is the real world metric?) are just greenwashing.
🏢🌱Datacenter farming? Wyoming Hyperscale is a firm taking an innovative approach: engineers have designed the hyperscale datacenter in Wyoming using liquid immersion cooling (LIC) technology, which significantly reduces water and power consumption compared to traditional air-cooled data centers.
Wyoming Hyperscale According to the firm, this will be “the world’s first sustainable hyperscale data center development” — a carbon-negative, multi-business ecosystem through 100% heat reuse - waste heat recovery dramatically reduces water and energy consumption for cooling while enabling year-round “hyperscale” indoor farming nearby.
The Human+Technocapital superorganism is simmering the planet
While tech fooms, the planet simmers…
🌡️No new normal The global surface temperature anomaly continues to hover around 1.6°C above the 1850-1900 pre-industrial baseline… less than last year but still way above the historical mean:
🌡️🌍⚠️Planetary boundaries breached An updated briefing from the Potsdam Institute for Climate Research showed that Earth may have breached seven of nine planetary boundaries, with ocean acidification at a critical threshold:
This is not a drill. Humanity’s journey on Earth, another graphic from the Potsdam Institute illustrating our post-Holocene predicament:
Potsdam Institute via @PGDynes
Holocene Anthropocene Thermal Maximum
Scaling out even further looking at the planetary record… a 2024 study has compiled a model of half a billion years of Earth’s temperatures. The models, which the researchers call PhanDA, estimate global temperatures over the last 485 million years, going back to the end of the Cambrian period.
Interestingly, global mean temperatures varied from lows of 11°C during recent glacial periods to highs of 36°C about 90 million years ago. The study found a strong link between carbon dioxide levels and global temperatures throughout most of the period examined. The very recent Holocene period in which humans evolved is almost the lowest mean temperature on record:
If indeed we are already Post-Holocene AND Post-Anthropocene … what will the Post-AI-Ocene look like?
Energy economist Nate Hagens is 80-90% certain we are entering what will be known by geologists as the Holocene Anthropocene Thermal Maximum … a “geological monkey trap” which will play out for the next 1000 years, potentially wiping out all conscious, intelligent life on the planet. The term was first coined by Anthony McMichael in his chapter From Cambrian Explosion To First Farmer: How Climate Made Us Human:
“In the next two centuries, our species faces a new challenge of greater, faster, and protracted climate change. Since the Cambrian Explosion of new life forms around 540 million years ago, there have been five great natural extinctions and many lesser ones. The earliest extinction of multicellular life, though less destructive than its successors, occurred around 510 million years ago, apparently due to acute sulfurous shrouding, cooling, and oxygen deprivation caused by a massive volcanic eruption in northwest Australia. Most of these catastrophic transitions were marked by climate extremes, volcanic activity, and altered ocean chemistry, especially rapid surface acidification of shallow coastal waters.”
Viewed through this lens the coming future is indeed bleak for all life on the planet, including humanity… will waving the magic AI wand still save us?
🏃♂️2. Race to AGI - no slowing down
📈128 years of Moore’s Law
In AI, clearly things ain’t slowing down. Steve Jurvetson updated his “125 years of Moore’s Law” infographic to include the last three years: line keeps on going up. (I use this slide a lot when just demonstrating how Kurzweil’s Law of Accelerating Returns continues to play out…)
Nvidia's AI chip dominance
Arguably the single most significant firm in AI in 2024 was Nvidia, under the disarmingly modest leadership of CEO Jensen Huang. The age-old business model of selling shovels to miners has never been so lucrative.
Now at the end of 2024, Nvidia holds an unassailable position in the AI chip market, with over 80% market share, which has sent the company's market value soaring above US$3.5 trillion, at several points becoming the most valuable company in the world by market cap, finishing the year in 2nd position:
(Note only one company in the top 10 is non-tech.) While competitors like Broadcom, AMD and Intel are trying to catch up, Nvidia’s lead appears impossible to overcome in the short to medium term. (Intel in particular is now rudderless following the booting of CEO Pat Gelsinger late in the year).
This excellent infographic from Eric Flanningam’s Generative Value Substack tells the story of a Nvidia’s gradual…then all at once rise to dominance:
For virtually all of the huge data centre builds discussed above, Nvidia’s latest generation of GPU chips will play a significant or exclusive role. (The primary exception being Google and Apple’s own in-house designs). Some market dynamics:
Murky economics Firstly, Nvidia’s stratospheric sales numbers (which are pumping the valuation) may not be entirely transparent: an FT editorial looked Inside the Murky New AI chip Economy, with US$11Bn of loans to “neocloud” groups backed by their possession of Nvidia’s AI chips. These firms (including CoreWeave, Crusoe and Lambda) sell “cloud services” further up the AI stack to OpenAI et al…effectively creating a new class of financial securities backed by GPUs…to fund the purchase of more GPUs. No ponzi scheme here, then.
Vertical integration coming? Apple (M4 series) and Google (TPUs) are well advanced with developing their own AI silicon and continue to invest significantly in this. In 2024, Amazon and Microsoft also announced their own AI chip initiatives.
The rise of specialised chips Alongside general-purpose GPUs, there's a growing trend toward specialised AI chips. Companies like Cerebras, Groq, Graphcore and TensTorrent are all developing chips designed specifically for AI workloads, offering comparative advantages in performance and energy efficiency.
AI processors from Cerebras, Groq, Graphcore and TensTorrent Centralisation concerns despite the huge investments, the manufacturing of AI chips remains highly centralised, with key companies ASML and TSMC acting as bottlenecks — and single points of failure — in the supply chain.
The Biden administration’s US$5Bn+ investment in TSMC on-shoring manufacturing in Arizona began to yield results. Despite negative media coverage, TSMC’s new Arizona plant has reportedly started shipping 5nm process wafers to its first customer Apple in September… earlier than scheduled. Small “but significant” volumes. Soon Apple devices will have chips “Made In America”… (but the devices themselves likely still manufactured in China and SE Asia, natch…)
The continued concentration of chip manufacturing and IP in Taiwan raises concerns about the fragility of the global supply chain amid heightened US-China geopolitical tensions.
Alternative technologies While traditional silicon-based chips dominate the market in 2024, alternative computing paradigms like photonic computing and quantum computing are emerging as potential disruptors that could change the economics of AI infrastructure. See “Emergent…” later on…
🫧But *is it* a bubble?
So the AI investment bubble continues to inflate… are we there yet?
💸AI investment… go figure fundamental questions arise on where Nvidia’s future revenues will actually come from to justify the enormous multiple on their stock price. Who is their customer’s customer?
David Cahn (“Sovereign AI” is an interesting one…are governments really the largest future customers?)
*EITHER* we are in a bubble… Meredith Whittaker, President of Signal, the not-for-profit secure messaging app channeled the vibe in an interview with Wired celebrating Signal’s 10th anniversary:
"I think this generative AI moment is definitely a bubble…You cannot spend a billion dollars per training run when you need to do multiple training runs and then launch a fucking email-writing engine. Something is wrong there."
…*OR*, one or more of the following hypotheses will eventually turn out to be correct (thanks Claude for summarisation assistance):
Genuine technological breakthrough leading to AGI: Recent AI advancements, particularly in areas like large language models, represent a true path to AGI with far-reaching implications for humanity, way beyond mundane concerns like ROIC (Return on Invested Capital)
Vast market potential: The addressable market for AI technologies spans virtually every industry, justifying the large investments.
Strategic necessity: Investment in AI is crucial for companies to remain competitive, driving genuine demand rather than speculative interest.
Long-term value creation: Current investments may be justified by the long-term transformative potential of AI, even if short-term returns are not immediately evident.
(Another non-commercial dynamic is the escalating US vs. China military AI arms race …)
AI stocks: what if this time it really is different? Financial Times columnist Katie Martin distils the uncertainty of whether the AI-driven surge in tech stocks represents a new paradigm rather than a bubble:
“…the tendency among investors is still to assume that balance and harmony will at some point re-emerge. Either the big stocks will stop rallying so hard or the (still alleged) benefits of AI will trickle through the corporate world, pulling up the market as a whole. Similarly, the gap between the US, the undisputed champion of the tech race, and the rest of the world, will shrink. Mean reversion will occur, just as it always has done in the past. At the start of this year, many investors were looking for exactly that — a fading in US exceptionalism and tech stock concentration.
Both have in fact intensified. It feels like a good time, then, to question whether they are features, not bugs, of this new technological era.“
At the end of 2024, it’s hard not to see close similarities with the 1840s railways mania or the 1990s dotcom bubble … which in the end built out so much internet infrastructure for the crash survivors to capitalise on in the 2000s onwards:
“even after the railway investment mania went away, the railways never did … and the lesson of the dot-com bubble is similar.“
📈📈📈Situational Awareness - counting the OOMs to foom
Mindbomb of the year was the essay series Situational Awareness - The Decade Ahead published in one drop by German-American ex-OpenAI Superalignment team member Leopold Aschenbrenner:
For someone so young (early 20s), he is not short of swagger:
“The AGI race has begun. We are building machines that can think and reason. By 2025/26, these machines will outpace many college graduates. By the end of the decade, they will be smarter than you or I; we will have superintelligence, in the true sense of the word. Along the way, national security forces not seen in half a century will be unleashed, and before long, The Project will be on. If we’re lucky, we’ll be in an all-out race with the CCP; if we’re unlucky, an all-out war.
Everyone is now talking about AI, but few have the faintest glimmer of what is about to hit them. Nvidia analysts still think 2024 might be close to the peak. Mainstream pundits are stuck on the willful blindness of “it’s just predicting the next word”. They see only hype and business-as-usual; at most they entertain another internet-scale technological change.
Before long, the world will wake up. But right now, there are perhaps a few hundred people, most of them in San Francisco and the AI labs, that have situational awareness. Through whatever peculiar forces of fate, I have found myself amongst them. A few years ago, these people were derided as crazy—but they trusted the trendlines, which allowed them to correctly predict the AI advances of the past few years. Whether these people are also right about the next few years remains to be seen. But these are very smart people—the smartest people I have ever met—and they are the ones building this technology. Perhaps they will be an odd footnote in history, or perhaps they will go down in history like Szilard and Oppenheimer and Teller. If they are seeing the future even close to correctly, we are in for a wild ride.
Let me tell you what we see.”
His core thesis is twofold:
Firstly, an extension of Kurzweil’s law of accelerating returns, but instead of “all human brains” being $1000 in 2040… Aschenbrenner’s proposition is that automated AI researchers will be achieved in trillion-dollar GPU clusters sometime this decade:
If we count the “OOMs” (orders of magnitude…):
“We can “count the OOMs” of improvement along these axes: that is, trace the scaleup for each in units of effective compute. 3x is 0.5 OOMs; 10x is 1 OOM; 30x is 1.5 OOMs; 100x is 2 OOMs; and so on. We can also look at what we should expect on top of GPT-4, from 2023 to 2027.
… but the upshot is clear: we are rapidly racing through the OOMs. There are potential headwinds in the data wall, which I’ll address—but overall, it seems likely that we should expect another GPT-2-to-GPT-4-sized jump, on top of GPT-4, by 2027.”
Put another way in terms of the compute infrastructure required: the first trillion-dollar cluster before the end of this decade.
The second part of this thesis is that this acceleration will inevitably lead to an existential hostile race condition between US and Chinese governments for control of a Singleton ASI (Artificial Superintelligence) … a scenario he characterises as analagous to the discovery of nuclear fission nearly a century ago and which led to the highly secret Manhattan Project…. and ultimately Hiroshima and Nagasaki.
As such Aschenbrenner’s proposed course of action is pretty radical (no wonder he was swiftly exited from OpenAI!): to effectively “nationalise” frontier AI development under the auspices of the US government and ban open-source frontier AI research. AI summary of his suggestions:
Secure Datacenters in the US: Prioritize building datacenters in the US while collaborating with democratic allies like Japan and South Korea for fab projects due to their functionality and reliability.
Automate Alignment Research: Focus on automating alignment research to handle superintelligence effectively, leveraging somewhat-superhuman systems to assist in this research.
Improve Security Measures: Implement extensive security measures for AGI clusters, including airgapped datacenters, hardware encryption, SCIFs (Sensitive Compartmented Information Facilities), and strict internal controls.
Targeted Capability Limitations: Restrict model capabilities to reduce the fallout from failures, such as removing knowledge related to biology and chemistry from training to prevent misuse in creating biological and chemical weapons.
Government Collaboration and Regulation: Engage in extensive cooperation with the US intelligence community for security and consider government regulation to ensure AI safety during the rapid technological advancements expected during the intelligence explosion.
If it turns out his acceleration thesis is correct (obviously I sit on the fence but lean towards foom…) then the implications of this for AI safety… and in particular concentration of power in the hands of only a few… are pretty profound for all of humanity.
But, as always with [San-Francisco-AI-bubble] researchers stepping outside their deep environment of expertise and wildly extrapolating into existential risk or geopolitical military strategy (mea culpa)… despite his clearly planet-sized mind, I’m not entirely convinced his mental models are correct or complete. In particular:
The entirely US-centric view he takes which sees the world security order as a zero-sum bipolar contest between US and China just ignores the fact that more than half of the rest of the world’s population (and fair share of smart people) don’t live in either of these countries AND have their own agency AND do not want to be subject to the US government…
He is somewhat, ahem, *starry-eyed* about the non-authoritarian nature of the US government and US democracy in general…
His professional career has been mostly during what *may* be an AI investment bubble… in which case his extrapolation of the foom may be off and this is just another logistic S-curve…:
Living outside both the US and China, my framing instead is that the “legacy technology” of top-down, centralised “democratic” government in the US would quite possibly not be resilient to the real-world power forces which would come into play with ASI and instead the US would head even further towards an authoritarian state…
So… I would propose an alternative scenario where only a few hyperscale technology companies maintain control of the ASI tech but instead go pan-national in their governance arrangements, filling in the international security gaps where supposedly democratic nation-state governments and the UN used to be…
(If nothing else, these essays make for some interesting wargame scenarios, particularly if played out on Aschenbrenner’s imminent timelines…!)
Anyway…. a highly intelligent, thought-provoking read which should provide a wake up call to political leaders outside the US and China who continue to blithely ignore the potential strategic implications of the frontier AI acceleration going on there.
If you have 4 and a half hours(!) to spend then you can watch Leopold in lively (if lengthy) conversation with the excellent Dwarkesh Patel. (One of those episodes where the podcast editor just gave up and clicked “publish”…🤣)
(Also interesting hearing his inside story of being fired from OpenAI just a few weeks before and refusing to sign the NDA…not to worry, he’s already set up a new investment firm focused on AGI with anchor investments from Stripe founders Patrick and John Collison…)
Also, this clip (click to view on X)🙉:
By the end of the year, Aschenbrenner’s influence had clearly reached Washington. Out of nowhere, just after the US had squeezed through its general elections relatively unscathed, the US-China Economic and Security Review Commission delivered its 2024 Report to Congress… with the following recommendation front and centre:
Note the terminology:
It may well be misdirection, but A16Z’s Marc Andreessen went on record saying this move towards centralised government control was the reason he and a bunch of other Silicon Valley VC investors backed Donald Trump’s presidential campaign:
“They [the government] said AI is a technology that the government is going to completely control. This is not going to be a startup thing. They actually said flat out to us "Don't do AI startups, don't fund AI startups - it's not something that we're going to allow to happen."
They said AI is going to be a game of two or three big companies working closely with the government. We're going to wrap them in a government cocoon, protect them from competition, control them, and dictate what they do.
I said "I don't understand how you're going to lock this down so much because the math for AI is out there and being taught everywhere." They literally responded "During the Cold War, we classified entire areas of physics and took them out of the research community. Entire branches of physics went dark and didn't proceed. If we decide we need to, we're going to do the same thing to the math underneath AI."
I said "I've just learned two very important things because I wasn't aware of the former and I wasn't aware that you were conceiving of doing it to the latter." They basically said "We're going to take total control of the entire thing."“
Full interview here:
Machines of Loving Grace
Compare and contrast Aschenbrenner’s work with Machines of Loving Grace - How AI Could Transform the World for the Better, a thoughtful and optimistic essay by Anthropic founder and CEO Dario Amodei.
In it, Amodei — one of the very few people in the world with a daily close-up viewpoint at the frontier of AI — articulates his generally optimistic vision of how rapidly accelerating AI technologies may soon yield radically different futures that most people today wouldn't consider possible: in fields such as biology to solve health, economics to solve poverty and governance to solve peace.
(But these opportunities are not without concurrent risks or challenges... he placates the AI safety crowd by saying that Anthropic will continue to mostly focus on addressing AI risk day to day…)
The three key concepts which stood out for me:
The "compressed 21st century":
"...my basic prediction is that AI-enabled biology and medicine will allow us to compress the progress that human biologists would have achieved over the next 50-100 years into 5-10 years. I’ll refer to this as the “compressed 21st century”: the idea that after powerful AI is developed, we will in a few years make all the progress in biology and medicine that we would have made in the whole 21st century."
“Powerful AI” — a better term than “AGI”.
(Summarised:)
"powerful AI" as an AI model that is smarter than a Nobel Prize winner across most fields, capable of performing complex tasks autonomously, and able to interface with the world like a human working virtually. This AI can be replicated millions of times, operates 10-100 times faster than humans, and can work independently or collaboratively on tasks, essentially functioning as a "country of geniuses in a datacenter."
“Marginal returns to intelligence”:
“Economists often talk about “factors of production”: things like labor, land, and capital. The phrase “marginal returns to labor/land/capital” captures the idea that in a given situation, a given factor may or may not be the limiting one …I believe that in the AI age, we should be talking about the marginal returns to intelligence, and trying to figure out what the other factors are that are complementary to intelligence and that become limiting factors when intelligence is very high.“
He goes into detail on five areas that he judges to have the greatest potential to directly improve the quality of human life
Biology and physical health
Neuroscience and mental health
Economic development and poverty
Peace and governance
Work and meaning
As a counterweight to Amodei’s techno-optimism, he is also a firm advocate of “Democratic AI” (read US hegemony). And Anthropic announced a new partnership with Palantir and AWS to provide Claude to US intelligence and defence agencies for processing information up to and including “Secret” classification.
Claude does not approve:
This is the same company that professes to be working towards “Safe AI”. (Hold that thought… perhaps the creeping integration of OpenAI, Anthropic and Google into the US military could actually end up being a reverse takeover?)
Major labs roundup
OpenAI
After 2023’s CEO-ouster-that-wasn’t, drama never far away from OpenAI.
The founding team gradually depleted, with the departure of Chief Scientist Ilya Sutzkever, CTO Mira Murati and the “sabbatical” of Greg Brockman (who has since returned to duty). Only CEO Sam Altman remained…
💸Non-profit for-profit In October, OpenAI confirmed that their latest funding round closed, raising US$6.6 billion at a staggering US$157 post-money billion valuation. (A significant increase from its previous valuation of approximately US$86 billion earlier in 2024, nearly doubling its worth).
Doctored image from @ai_for_success (numbers pasted in over slide) There are strings attached:
Investors can pull their cash if OpenAI does not transition fully into a for-profit entity within two years. That could get, er, dramatic.
Exclusivity clauses: according to inside sources, investors in this round are restricted from investing in certain competing companies.
Anyway, this round makes OpenAI the largest venture-backed company by valuation from Silicon Valley by market cap (surpassed globally by Chinese TikTok owner ByteDance at ~US$220Bn, with SpaceX the only other VC-backed firm in the same league.)
Apparently OpenAI gives employees and investors PPUs - Profit Participation Units, not equity. But when will this firm *ever* make a profit? (Reminder: OpenAI is projected to make a loss of approximately US$5 billion this year, on anticipated revenues of around US$3.7 billion. Go figure.)Emad Mostaque nails the economics of the OpenAI investment in one pithy tweet:
AGI achieved internally? OpenAI did, however, keep shipping and releasing new models and features throughout the year. Too many to list here, but highlights (see more details in next section)
Advanced Voice Mode - a very intelligent AI you can talk to conversationally… it gets pretty natural after a while.
After much hype, OpenAI finally launched their “o1” “reasoning” model in preview - previous codenamed Strawberry. The main difference being that the model takes longer to “think”… working through a Chain-of-Thought (CoT) internally before returning a “reasoned” answer. Sometimes it even manages to count the ‘r’s correctly:
Canvas - collaborate with ChatGPT like Google Docs
Shipmas… 12 days of incremental releases, most notably: full release of their Sora text-to-video model and culminating in…
…BREAKING o3 — a new upgraded “reasoning” model which scored a breakthrough 75.7% on the ARC-AGI benchmark (at the $10k compute limit). A high-compute (172x) o3 configuration scored 87.5%.
Sama here trying to make light of the last year’s ups and downs:
Anthropic
Challenger lab Anthropic played a steady wicket throughout the year.
Surprised everyone throughout the year with how good Claude was. And Claude 3.5 Sonnet, launched in June, was very good indeed, and extremely good value as well:
Anthropic also sucked in capital, raising an additional US$4Bn from Amazon and making AWS its primary cloud partner.
No drama? This ad showing some swagger against OpenAI is perhaps tempting fate…
CEO Dario Amodei was another regular on the podcast circuit, shining a more discerning light on the technology, economics and scaling challenges at the frontier of “AGI” (as above he prefers the term “powerful AI”). Two worth a listen:
With Logan Bartlett:
With Lex Fridman:
Google
🫣Woke Gemini No sooner had the big fanfare in February of Google’s Gemini 1.5 Ultra model been heard… than people started finding, er, *historically inaccurate* images in the model’s outputs:
Adi Robertson via The Verge Plus, users were also able to easily get it to demonstrate some very obvious biases:
Google suffered significant reputational fallout from these issues, pulled Gemini image generation from the release very quickly and issued a mea culpa announcement. Particular parties, mostly in the US grabbed onto the incident to fight conventional culture wars.
Bigger picture: I think it demonstrated the futility of trying to embed “one size fits all” ethical values into commercial AI services - the “wokeness” of closed-source models would just *not be a thing* if the model training and guardrails were released as open source. Let users choose to dial the “wokeness” up or down to their taste?
Obviously this was prime meme material at the time as well.
Joscha Bach on fire:
As was Orpheas:
And who’s really in charge I wonder…?
Perhaps because of this high-profile failure, Google kept a remarkably low profile for most of the year with its mainstream consumer AI releases, focusing instead on rolling out Gemini services inside its existing Google Workplace apps. But then they finished the year very strongly:
**♊Gemini Experimental 1206** … In December, Google quietly released a new version of Gemini on the first anniversary of the original Gemini launch… which went right to the top of the Chatbot Arena leaderboard, closely followed by Gemini 2.0 Flash (Experimental), the lighter, high-speed model which is replacing Gemini 1.5 Flash, which benchmarks way beyond its class:
This started Google Deepmind’s “week of shipping” in December also includ Veo 2 and Imagen 3 video and image models, Deep Research an agentic AI research assistant and of course more features following their amazing success with NotebookLM. (More details below).
Alphabet CEO Sundar Pichai discussed Google and Alphabet’s year with a sit down with Andrew Ross Sorkin, with a steely confidence not always apparent in Google’s public demeanour… watch out of their success in 2025.
Meta
Meta founder and CEO Mark Zuckerberg continued his reinvention as a “man of the people”, with the transformation described by some as a "Zuckaissance". Over 2024 his public persona shifted from a geeky tech figure to a more relatable (and assertive) persona. Particularly, marked by a bolder fashion sense (not just the same grey hoodie):
Zuck before and, er, after (via Ash Jones on LinkedIn) Zuck’s AI strategy for Meta is that LLMs are only going to be underlying compute infrastructure like Linux… and hence Meta is “all-in” on the economics of open source. The benefits of bringing communities over to your platform and building a critical mass of adoption and talent worked with Pytorch — but will it ultimately pay enough dividends to justify the multi-billion dollar spending bill? We will see. Here he is discussing in depth, again with the excellent Dwarkesh Patel:
Meta also continued to ship regular AI updates, particularly to its Llama 3 model series (Llama 3, 3.1, 3.2 and 3.3 all released during 2024). Llama 3.3 is a 70B parameter model designed to deliver performance comparable to the previous 405B model but with greater efficiency and lower operational costs. It also features a 128k token context window and outperforms competitors like GPT-4o, Gemini Pro 1.5, and Amazon's Nova Pro in various benchmarks:
And apparently Llama 4 is in training… on a cluster with over a hundred thousand GPUs…
AI ads: Meta did report that so far in 2024 over one million advertisers utilised its gen AI tools to create more than 15 million ads in a single month, supposedly enhancing ad relevance and engagement.
xAI
Elon Musk has been driving investment at xAI. The company raised US$6 billion at a valuation of US$50Bn - double its value earlier in the year… the money will go mostly to purchase 100,000 Nvidia chips for Memphis data centre. Key investors include Qatari sovereign funds, Sequoia and A16Z according to WSJ.
By the end of the year, Grok 2 was ranked 8th in the Chatbot Arena leaderboard, behind only Google and OpenAI but ahead of Anthropic Claude Sonnet 3.5. Remarkable progress in just over 1 year.
xAI has tightly integrated the Grok chatbot into the X social network, most notably using it to summarise current trends across the X platform. In many ways this serves as:
(1) a strategic compass for sensing things that are unfoldinging right now
(2) a potential lever to skew the narrative on the platform itself
OpenAI’s Sam Altman had some choice words about Musk at the recent Dealbook Summit:
Microsoft and Amazon have AI teams in the wings building proprietary models (insurance policy…?), but not out there at the frontier yet. Their bets on OpenAI and Anthropic have yet to pay off…
And the rest:
So many AI labs out there but notable mentions for Cohere, French Mistral (Le Chat), Midjourney and Magnific for keeping close to the pace on far smaller budgets.
Leading general language AI labs from China include O1 (confusingly named), Deepseek, Qwen (Alibaba) as well as increasingly sophisticated models from Tencent (Huanyuan). All of which are close followers behind Silicon Valley in the AI race. Will they catch up in 2025?
🎨3. Omni-modal AI
(A “mode” in generative AI refers to the different types of content or outputs that an AI model can produce.)
The advances this year have been incredible. New state-of-the-art (SOTA) AI models and tools have been released weekly in a Cambrian explosion of innovation. Every week I’ve been chronicling and trying to keep up with the stream of updates being published by labs all around the world.
Trying to make sense of this all, the only core insights I keep returning are these:
Intelligence keeps getting cheaper the cost of AI inference continues to decrease by order of magnitude. What can you achieve in 2025 if the marginal cost of adding new intelligent agents approaches zero?
The rate of improvement keeps accelerating. See the “All lines go up” section above… I use this slide all the time to illustrate how Horizon 3 is now 2 years away for any knowledge-related function - adjust your timelines accordingly:
Open source, decentralised AI is the only practical counterweight to the concentration of AI power in the US and China.
Here I’ll round up just the highlights of what I’ve managed to track at the frontier of AI in 2024. Colleague Sam Ragnarsson and I compiled this logo infographic of the main tools in July, and things have moved on significantly since then:
Closed vs. open source
But first it’s worth considering the two distinct tactical approaches to publishing AI models being taken by the major labs and their challengers. Google, OpenAI, Anthropic all keep the source code, training data and weights confidential and proprietary.
Meta, however, is the largest Western AI lab releasing its frontier models as open source. Founder Mark Zuckerberg explained the evolving rationale behind Meta’s open-source strategy in a podcast with Dwarkesh Patel discussing the release of their Llama 3 model back in March… and the costs involved for training the forthcoming Llama 4:
Dwarkesh: Going back to the investors and open source…Would you open source that, the $10 billion model?$10 billion of R&D and … now it's like open source for anybody?
Mark: Well… We have a long history of open sourcing software, right? We don't tend to open source our product, right? So it's not like we don't take like the code for Instagram and make it open source, but we take ... a lot of the low-level infrastructure, and we make that open source, right? Probably the biggest one in our history was Open Compute Project, where we took the designs for kind of all of our servers and network switches and data centers and made it open source and ended up being super helpful because … a lot of people can design servers, but now … the industry standardized on our design, which meant that the supply chains... basically all got built out around our design. The volumes went up, so it got cheaper for everyone and saved us billions of dollars. So awesome, right?
Okay, so there's multiple ways where open source, I think, could be helpful for us. One is if people figure out how to run the models more cheaply. Well, we're going to be spending tens or like $100 billion or more over time on all this stuff.
Meta AI lead Soumith Chintala’s keynote at the ICML 2024 conference set an even more strident tone:
Leading labs outside the US including France’s Mistral, plus China’s Deepseek and Qwen have also taken the approach of releasing their frontier language models under open source licences. The capabilities of the open source models are now getting closer and closer to major closed labs - in November Chinese lab Deepseek released a ‘reasoning’ AI model to rival OpenAI’s o1 just weeks apart.
The implications of open source AI are still evolving — but fundamentally: all governments, large and small companies and individuals around the world now have the option to host and develop their own LLM applications more or less at frontier model capability, without needing to license from OpenAI, Microsoft or Anthropic. And the smaller models *should* be runnable locally on a powerful enough Macbook…. maybe a bit slow! This implies at the very least a pathway to resilience for AI sovereignty.
However, with news that Llama 3 is turning up inside Chinese military applications, it’s questionable whether the US government will continue to let them just release open AI research to the whole world if the US AI industry does become increasingly secretive and militarised. Nonetheless, so far Meta’s stance is a major competitive flex against the moats of OpenAI, Anthropic, Google and others.
(Good luck to Leopold and his attempts to keep frontier AI research confined to a few nationalised labs in the US!)
AI around the world
The 7th Stanford AI Index Report published in April was one of the most comprehensive of the year coming in at a staggering 500 pages.
IN particular, this map speaks a thousand words about the future of AI-based power…all the areas in white (and blue) should seriously be thinking about collaborating on open source alternatives…
Startups
Life as a non-frontier AI startup was increasingly precarious in 2024, particularly for “GPT Wrappers”:
Chatbot arena
Benchmarking AI is an emerging industry in itself… numerous benchmarks have evolved this year to cope with models getting smarter and smarter. And with each new model release, certain benchmarks are overemphasised above others.
Plotting Progress in AI shows the evolution of AI benchmarking. (Basically: be very careful saying that “AI will never be able to do X”).
“Explainer: How were these numbers calculated?
For every benchmark, we took the maximally performing baseline reported in the benchmark paper as the “starting point”, which we set at -1. The human performance number is set at 0. In sum, for every result X, we scale it as (X-Human)/Abs(StartingPoint-Human).”
As LLM and increasingly native multimodal models kept being released throughout 2024, and kept approaching or exceeding human cognitive ability, each one was comparison-benchmarked on a number of sites. In particular Chatbot Arena maintains the widest-used industry benchmark for LLMs (but its results should be treated with caution as the complexity and specialisation of these models increases). As of 22 Dec, here’s the top 11 leaderboard, all proprietary models with Google right there in the top 2 for now. Only one Chinese lab - the confusingly-named O1, made it into the top 10 with its late-breaking Yi-Lightning model.
Software development
One of the most striking capabilities of LLMs is to write, review and edit software code as well as most humans on the planet. The advances against various software development benchmarks have been spectacular during the year, as has the rapid development of new tools which accelerate the practice of software development:
Models including OpenAI o1 and Claude Sonnet 3.5 provide the back end capabilities for IDE tools.
The original software AI, Github Copilot, grew in capabilities and ended up being included in a free Github licence by the end of the year. A study involviong 4,867 software developers across three companies found that those given access to GitHub Copilot, showed a 26.08% increase in completed tasks across all experiments.
Also a notable mention for SuperMaven VS Code plugin
Revelation of the year was Cursor — a revolutionary code editor (branched off VS Code) which uses 3rd party models (I use Claude 3.5 Sonnet) to deliver natural language editing, proactive debugging, and project-wide context understanding. I’ve been using it for 3 months now and blown away by how much it boosts my productivity.
Just sharing one “wow moment” example: here’s a compressed snippet of dialogue with Cursor after a highly productive 5-hour session of back-and-forth plumbing my multiple feeds into just one:
Ben: Suggest improvements to this layout so that it looks more like a social media feed and less like a table. Also make the display responsive.
Cursor: I'll help you transform this into a more modern, social media-style feed layout. Here's how we can modify the template: …(writes entire html and css modified code)
(Ben clicks “Apply” and “Accept” twice. Goes to browser. Clicks refresh. Everything works. It looks amazing. Wow. Mind blown. Does exactly what I asked for, and then some…)
Ben: Amazing!
Cursor: Thanks! Just a tip - if you want to make it even more social media-like, you could consider these additional enhancements:
Add hover effects on the cards: (code)
Add a loading skeleton animation while content loads: (code)
Add a "Read More" truncation for long descriptions: (code)
(Ben clicks “Apply” and “Accept" 3 more times. Goes to browser. It all works.)
The wow bit: the AI (unprompted) came back and suggested more improvements I didn’t ask for…
🧑💻🤖Cognition Labs is a Peter Thiel-backed startup who have built a product called Devin, which they claim is the first AI software engineer:
“Devin is an autonomous agent that solves engineering tasks through the use of its own shell, code editor, and web browser. When evaluated on the SWE-Bench benchmark, which asks an AI to resolve GitHub issues found in real-world open-source projects, Devin correctly resolves 13.86% of the issues unassisted, far exceeding the previous state-of-the-art model performance of 1.96% unassisted and 4.80% assisted.“
By the end of the year, Devin was released commercially at US$500/month price tag.
Image
Image models have had a relatively slow year but still significant advances.
Midjourney 6.1 released in July. Still the SOTA. Also Midjourney continues to pump out editing features including image editing, upscaling and style codes… so much capability.
FLUX / Black Forest Labs The original technical team behind the beleaguered Stability.ai came out of stealth as Black Forest Labs, launching their open source FLUX AI image model — which can be run locally. It’s pretty good straight off the bat:
(Can be used from within Krea.ai for US$10/month)
OpenAI’s DALL-E (built into ChatGPT) hasn’t really seen much of an improvement all year… although with Sora’s impressive video capabilities it must be waiting in the wings?
Stability AI released Stable Diffusion 3.5:
“Open release” free for both commercial and non-commercial use under the permissive Stability AI Community License, can run on consumer hardware. Impressive image quality, solving the text rendering problem:
🖼️And finally at the end of the year, Google’s Imagen 3, an upgraded version of their text-to-image model which gets fingers right! (And presumably fixes the woke history bugs in the last version). Use from within Gemini.
(Also lots of other labs out there including from China… too many to cover…! Lots of choice).
Video
2024 was definitely the year when AI video came of age. OpenAI’s preview of its Sora model early in the year gave an indication of what was possible with the SOTA… but it was only released to the public in December.
During the year both Western and Chinese labs pumped out model after model and feature after feature. Text-to-video, image-to-video, video-to-video style transfer… the editing tools are now reaching parity with many professional studio-level special effects houses.
The leaderboard looking like this by the end of the year:
Just a few notable demo reels from this year:
Runway continued to be a leader on features. The Gen3 Video To Video now goes to 20 seconds. Here’s an example of what you can do with it (via @Uncanny_Harry , who doesn’t look like this in real life!)
🎥🪄KLING While the world continues to wait for OpenAI to just drop Sora, a Chinese competitor arrived on the scene suddenly this week… GenAI video service KLING went straight into open access in China with some very impressive demo videos. Apparently KLING can generate 2-minute videos at 30fps, 1080p quality and is available on the KWAI iOS app with a Chinese phone number. Here’s an early demo reel:
Genmo Mochi 1 another model from another lab… but open source.
OpenAI Sora… just released. This instructional video from OpenAI creative Chad Nelson shows how powerful the new workflow tools are:
Google Veo 2… just released:
🍝Will Smith spaghetti benchmark goes up again If you’re not “feeling the acceleration” yet, check out the advances in just 20 months on the classic “Will Smith eating spaghetti” AI video meme and de facto benchmark.
April 2023: the original using early Stable Diffusion (note the weird ghost Shutterstock watermark, presumably from the training dataset…)
November 2023, RunwayML via @bennash (covered in Memia 2023.45)
February 2024: Will Smith parodies viral AI-generated video by actually eating spaghetti (well played)
And then mid-2024, it looks like KLING was the first to properly crack the “Will Smith Eating Spaghetti” benchmark. Top marks. (via @venturetwins)
Convergence What we’re also witnessing is a convergence of the entertainment and AI industries:
Legendary filmmaker James Cameron (of Terminator and Avatar fame) joined Stability AI's Board of Directors. (It looks like Stability has a strategy again post-Emad)
Partnership between Lionsgate Studios and Runway to train the AI model on its movies and shows.
By the end of 2025, perhaps every legacy film studio will have licenced its back catalogue to AI shows… or perhaps even more likely they will be engaged in futile copyright cases on “AI training”….
Either way, would someone please hurry up and make the fanfic version of cancelled Series 2 of The Peripheral with AI?!
🎵Audio
All your music is now AI…
Suno First off the blocks early in the year was audio startup Suno:
“Suno, create a song about about New Zealand in the future in the style of 90s British Britpop” (it’s awful, but strangely compelling…):
Later in the year this had advanced to Suno AI Covers - just sing a line and get a full music track with your song arranged for you.
Suno 4.0 out now promises far more fidelity and quality.
Udio
After the world of music was just about getting used to Suno, thinking “well it’s a bit limited…” along comes Udio, which *raised the bar*.
One of my favourite creations, "Dune the Broadway Musical":
Rolling Stone gives the background to the Udio team here: AI-Music Arms Race: Meet Udio, the Other ChatGPT for Music.
Instant soundtrack backgrounds for all your projects now… for cents in the dollar (if not free). Amazing. As I said in an inteview with RNZ’s Jesse Mulligan back in April 2023, we’re heading towards a future where all music is AI-generated.
🗣️Voice
🤖Still hiring humans? Bland.ai came out with a slick new sales demo. Voice delay down under 1 second now… very nearly impossible to tell who or what is on the other end of the call. Tata CEO K. Krithivasan sees where this is heading: AI could significantly replace the need for call centres within 1 year.
ElevenLabs is also one of the leading voice AI labs:
Midyear they released their new text reader app which will narrate any other text content - articles, PDFs, ePubs, newsletters - in any voice you choose from their “Iconic Voice Collection“ library. Upload your content, and listen on the go. So far voices include Judy Garland, James Dean, Burt Reynolds and Sir Laurence Olivier (with royalties accruing to their deceased estates, one assumes… Strangely, no Scarlett Johanssen!)
By the end of the year they also dropped Conversational AI agents:
ChatGPT Advanced Voice Mode, officially launched in May but only rolled out to most subscribers in October, AVM is currently the most sophisticated voice companion so far. I’ve taken to chatting to it in my car … amazingly close to human… and yet not quite. Here’s a great clip from OpenAI’s launch event with erstwhile CTO Mira Murati getting it to translate in real time. I used it exactly like this in Vietnam recently.
Despite occasional reliability and latency issues, it really is like magic. And all-too reminiscent of Spike Jonze’s prescient 2014 film Her. As CEO Sam Altman posted, quickly earning the ire of Scarlett Johannsen, on whose performance as AI girlfriend Samantha ChatGPT’s initial voice, er, *closely resembles*. But despite trying, there’s unlikely to be any legal comeback.
NotebookLM - Unexpected star of the year for voice goes to Google’s AI success story, NotebookLM. With this still-free feature, users can upload any document or website they want and generate a very realistic “deep dive” podcast conversation which really picks out the details that humans may miss. I experimented a few times getting it to discuss the weekly newsletter, ICYMI:
Breaking: last week Google added an interactive feature where you can request the AI hosts dial you into the podcast, and you can ask a question of them live.
Avatars
Heygen Avatars 3.0 evolved beyond lip-syncing to feature full-body dynamic motion. Plus, avatars' facial expressions and voice tones are dynamically generated to perfectly match the script. Still uncanny valley…. but closer and closer:
Three’s a crowd Of course, when you animate these avatars with an AI language model and let them talk to each other… Ethan Mollick demonstrates (excruciating… but how long until you just can’t tell…!?):
3D worlds
🪄Imagination to creation The intersection between generative AI and Gaming is still in its infancy… this short demo from the recent EA Investor Day event shows what is coming down the line: user-generated video game content, using 3D assets, code, gameplay hours, telemetry events and custom models to remix games and asset libraries in real time.
(Video via @adcock_brett). Now imagine this happens when you’re wearing Orion AR glasses (see below)… reality will never be the same…
Google DeepMind also made a significant breakthrough in game simulation by developing GameNGen, an AI-powered system capable of generating playable Doom gameplay in real-time… without using a traditional game engine:
4D worlds
BREAKING just as I was writing this: the Genesis Project is an open-source collaboration to create a “generative physics engine”:
“able to generate 4D dynamical worlds powered by a physics simulation platform designed for general-purpose robotics and physical AI applications”.
Mindblowing demo:
I suspect this is the standard “multimodal” model direction we will be seeing by the end of 2025.
Search
Internet search has gone through major advances throughout the year driven by new challenger Perplexity, providing short, referenced AI-generated search summaries rather than web page searches. OpenAI introduced ChatGPT Search and latterly Google Gemini has also added this functionality in a limited context. It’s probably been the most helpful productivity boost of the year for those of us constantly researching online… but what happens to Google’s online search advertising revenue monopoly?!? I wrote up my impressions after OpenAI launched ChatGPT search functionality:
And just at the end of the year, Google released Deep Research, an agentic AI research assistant which can create in-depth research reports on complex topics, complete with source links - automating the process giving you hours of research at your fingertips in just minutes. (Sort of like a ChatGPT Search or Perplexity query… but just more, er, deeply researched!)
Agents
Also throughout the year the hype cycle around “Agentic AI” gradually took on more form:
What actually IS an AI agent? Best (simple) explanation yet:
Early releases of Computer Use from Anthropic (OpenAI has introduced similar features into its ChatGPT desktop app as well) enable AI to drive a mouse and keyboard on your computer screen to achieve a task. While this is visually engaging, it’s a bit of a hack and unlikely to be reliable… I doubt there will be too many production-quality apps which use this tool.
Investor Jeremiah Owyang’s AI Agent ecosystem map makes sense of the rapidly evolving space:
Agent frameworks I’ve been tracking throughout the year:
Microsoft and Amazon released their own open source agent frameworks. Clearly plumbing code.
Salesforce have decided that “Business AI Agents” is their AI play with Agentforce… (beware large CRM SaaS behemoths bearing gifts…)
More low key, a couple of open source agent frameworks:
AutoGPT launched its new platform, enabling users to create, deploy, and manage continuous AI agents for seamless automation across various industries without extensive coding knowledge.
BabyAGI 2 - an updated Python framework from Yohei Nakajima for building a self-building autonomous agents. Just “pip install babyagi” and away you go…
🧪Science AI
AI has been steadily infiltrating the method and process of scientific research and discovery. Still not too many headlines… but expect to see more AI-accelerated scientific progress in 2025. (cf. Dario Amodei’s “compressed 21st century” concept.)
🔬AI takes science Nobels This year’s Nobel Prizes in Chemistry and Physics both went to groundbreaking achievements in AI, underscoring AI's growing impact on scientific research:
Demis Hassabis and John Jumper of Google DeepMind won the Chemistry Nobel for developing AlphaFold, the AI model that solved the 50-year-old protein folding problem. AlphaFold has now been used by over 2 million scientists for various applications, including vaccine development, with clinical trials for new drugs expected within two years. (And was magnanimously released on a limited open-source licence in 2024).
Geoff Hinton, the veteran “father of AI”, received the Physics Nobel alongside physicist John Hopfield for their work on neural networks — the foundation of most modern AI. In a press conference facilitated by University of Toronto where he has taught most of his career, Hinton (now a notable AI Doomer) spends much of his time talking about his concerns on existential AI Risks to humanity (including a dig at Sam Altman for OpenAI’s change from mission-driven to profit driven entity…)
Knowledge graphs and scientific discovery: Accelerating scientific discovery with generative knowledge extraction, graph-based representation, and multimodal intelligent graph reasoning.
Virtual Lab: a team of AI Agents design new nanobodies:
“The Virtual Lab is an AI-human collaboration for science research where a human researcher works with a team of large language model (LLM) agents to perform scientific research. Interaction between the human researcher and the LLM agents occurs via a series of team meetings, where all the LLM agents discuss a scientific agenda posed by the human researcher, and individual meetings, where the human researcher interacts with a single LLM agent to solve a particular scientific task.
As a real-world demonstration, we applied the Virtual Lab to design nanobodies for one of the latest variants of SARS-CoV-2 (see nanobody_design). The Virtual Lab built a computational pipeline consisting of ESM, AlphaFold-Multimer, and Rosetta and used it to design 92 nanobodies that were experimentally validated.”
(Generative AI for proteins, DNA and nanotech is only a couple of years away, max….)
🪰Fly connectome A consortium of scientists known as FlyWire (previously covered Memia 2023.10 and 2023.37), co-led by neuroscientists Mala Murthy and Sebastian Seung at Princeton University, have created the most comprehensive map of a fruit fly's brain to date. The "connectome" includes nearly 140,000 neurons and over 54.5 million synapses which enabled the creation of a computer model of the entire fruit-fly brain, which accurately predicted fly behaviour.
The 50 largest neurons in the fruitfly brain. Image: FlyWire, Princeton University
🤖🦾4. Robots, robots, everywhere
The acceleration of AI applied to engineering has been nowhere more apparent than in the autonomous robotics and drones space.
🛣️The road to autonomy
The path to self-driving vehicles has been a long wait but 2024 has seen an inflection point.
Tesla Since 2013, Tesla CEO Elon Musk has repeatedly made inaccurate predictions for Tesla to achieve Level 5 autonomy (“full self driving”) within only 1-3 years. At the start of 2024, Tesla’s “Full Self Driving” autonomy product was actually rated “Poor” for safety behind a set of other car companies. Still a long way to go despite behind far ahead in hype volume. Musk’s vague announcment of a RoboTaxi and RoboVan sometime in the future didn’t do much to improve sentiment.
Waymo This year we did see signs that full autonomy has been reached: Alphabet subsidiary Waymo has been ramping up its fleet and in March this year began offering completely driverless rides in select parts of US cities Los Angeles and San Francisco and Los Angeles. Quite an advance for the company which started out with a little buggy on a track in 2015:
Waymo cars over time.. including the 6th generation minivan coming soon Waymo autonomous ride numbers have increased exponentially since then. Alphabet CEO Sundar Pichai went on record saying that it had reached 175,000 paid autonomous trips in early December. Sticking to this graph from an internal Waymo staffer:
via Reddit Atlanta and Miami next in 2025 with international plan to open service in Tokyo.
Cruise Meanwhile in December General Motors halted funding for its self-driving subsidiary Cruise, after the company never really recovered from a 2023 incident in which a pedestrian got stuck under a car and then dragged by one of its robotaxis. Details of rationale behind the sudden decision are still emerging but one data point: Microsoft will take an US$800M hit over Cruise robotaxi shutdown.
Notable UK challenger is startup Wayve which has been developing autonomy using machine learning. In May 2024, Wayve secured a significant investment of $1.05 billion in a Series C funding round. They continue to put out the most convincing videos on Youtube: any AI which can navigate London’s notoriously chaotic back streets is doing pretty well:
In amongst all of this, Uber’s strategy is to develop partnerships with Waymo, Wayve (and previously Cruise) after selling its own AV R&D unit back in 2020.
Humanoids
The other big trend this year has been the major advances in humanoid robots. A raft of companies has sprung up in the US and China all racing to create a human-equivalent unit.
Elon Musk continues to pump Tesla as a robotics company… trying to justify a share price multiple way higher than a car company. Tesla’s Optimus robot has come on leaps and bounds in 2024, with their promo team regularly throwing out videos… one update of Optimus robot autonomously sorting batteries in the factory and taking a long walk around an empty office:
(Video source: Tesla on X).
Most fascinating clip in that video is this scene below… where the robots are being “trained” by humans dressed in VR goggles and haptic gloves… *very occasionally*, reality is weirder than science fiction…!
We, robot Fast forward to a Tesla showcase event in October when they showed off the latest demos of their apparently walking, talking, charades-playing tele-operated humanoid robot Optimus which had all the moves mixing with the crowds… (click tweet for video, compelling demo):
Notable advances from many other robot labs including:
Figure started shipping its first fleet of F.02 robots in foam-packed boxes:
(Next year’s hot Christmas gift?!)
Boston Dynamics (now owned by Hyundai Motor Group) showing off its updated non-hydraulic Atlas robot at work in a simulated factory:
Pudu “semi-humanoid” robot:
(Reminded of this classic Birkett cartoon from the 70s…)
NEO Beta - is that a guy in a suit?
🦿Robots run faster in sneakers: Chinese humanoid robotics company Robot Era recently showcased its Star1 models in a desert race through the Gobi. Two Star1 humanoids raced through diverse terrain, with one wearing sneakers outperforming its barefoot counterpart — the sneaker-clad robot reached speeds of 8 mph (3.6 m/s) and maintained the lead for 34 minutes.
🦿Unitree Chinese firm Unitree provided an update on its US$16,000, 2-hour battery life G1 humanoid robot. Definitely best performance out there so far… will these just be consumer toys or are they soon ready for industrial / service tasks? And if so, what does that mean for the labour market…!?
Welcome to the Clone World Enigmatic UK-based startup Clone Robotics finally unveiled their new website after posting intriguing demo videos of their robot limbs in motion - and started taking pre-orders for Clone Alpha Edition:
“Musculoskeletal, intelligent androids to solve all the common problems of daily life.“
The company has taken a completely different engineering approach to robotics, reverse-engineering biological bones and muscles. Westworld here we come…
The result (they claim) will be a natural walking gait and they’re working on faces too:
Open source robots:
There were also a number of open source robotics projects, just one notable example the RX1 Humanoid from Red Rabbit Robotics: a garage project building the first fully open source full human scale dual arm robot - supporting teleoperation and pick & place objects. Amazing how fast you can move when you set your mind to it:
Animals
Quadrupeds continue to be a prominent robotic form factor…
Unitree somehow think that making a video beating their robot dog will increase sales… they may be right. Pretty rugged.
🔥Dystopian These units have started showing up with, er, modifications: Thermonator is the first "flamethrower-wielding robot dog", selling for under US$10,000 and completely legal in 48 US states. Also US and Chinese military are in a race to build robot dogs with guns.
🦿🛞Dog on wheels OK, this isn’t scary at all in a military context…🫣 Chinese firm DEEPRobotics unveiled their new quadruped robot called Lynx that combines wheeled mobility with traditional four-legged locomotion - on all terrains. By far the most advanced robot agility I’ve seen demonstrated so far…
Less(?) terrifying, over at Boston Dynamics… meet Sparkles. Should we expect a bunch of cute pet robot dogs on the shelves by next Xmas…?
Miscellaneous body plans
A range of mundane to just weird and wonderful robot body plans emerging:
Peanut Robotics … cleaning surfaces in a hotel or house near you soon…
Nifty autonomous digger building a rock retaining wall (video via @nhfoley):
Massage robot more day-to-day usage…French startup Capsix robotics makes IYU, the world’s first massage robot…
Robotic Manta Ray Researchers at North Carolina State University have developed a new robotic manta ray that swims at an impressive 6.8 body lengths per second, nearly doubling the speed of its predecessor while using less energy.
🛞Robo-sphere Chinese police have begun trialing a 125kg, amphibious, self-balancing robotic sphere capable of pursuing suspects at speeds up to 35 km/h on land or water and equipped with non-lethal weapons for law enforcement purposes.
Drones
Finally, 2024 has been the year of drones - and in particular, swarms of drones.
Shenzhen the Guinness World Record was set with a stunning light display made up of a swarm of 10,197 drones.
Although pretty, this is pretty major military flex by China. Imagine each of these mini drones with a small projectile weapon or explosive attached…
Military drones
Which brings me on to one of the most fear-inducing developments of 2024, the race between US, China, Russia and others to militarise AI and drones.
Fibre-optic drones Ukraine has been the testing ground for a new generation of drone warfare since 2022. Videos of (first person view) FPV drone kills on both sides abound on social media. Ukrainian engineers have also developed a drone tethered to a fibre-optic cable that makes it virtually undetectable and impervious to Electronic Warfare (EW - signal jamming) attack.
“Equipped with a 20-kilometer (12.5 mile) cable, the drone can fly for 20 minutes at a speed of 60 kph (37.5 mph) carrying a 5.5-kilogram (12-pound) payload. Recent tests enabled flight endurance to be extended to 46 minutes“
Kyiv Post According to Western news reports, China has built the world’s first dedicated drone carrier warship:
Meanwhile US autonomous military “startup” Anduril, founded by ex-Oculus founder Palmer Luckey, continued to paint a bleak picture of the future with gung-ho demo videos of its drone tech simulating direct kills:
One small mercy, on the sidelines of this year’s APEC Summit in Peru: the US and China agreed to keep humans in the loop (rather than delegate to AI) when it comes to firing nuclear arms…well, phew😬. (Expect blurred boundaries of where AI reaches right up to the decision to press the button — one dark scenario is an AI posting a highly crafted series of tweets targeted directly at Trump’s brain circuitry and …. boom…💥)
At the end of 2024, mystery surrounded repeated sightings of drones at night time in several US states, particularly New Jersey… with federal government departments apparently unaware of what these are. Crazy times…
🔗5. Post-Web3
“Web3” — the catch-all term for decentralised computing — burned through a lot of capital in 2020-2023 without much to show for it, except for a bunch of failed ponzi schemes and worthless Bored Apes. In 2024, there are signs that the decentralised infrastructure which was built over that time is now starting to be used for real-world applications. But still a way off. Plus, Web3 infrastructure gives AI agency for the first time…
💸$100K BTC
The year in crpyto started out in the doldrums.
The frothy bubble market of NFTs fizzled out to nearly nothing
Prosecutions against FTX, Binance and others took time to finally resolve. In March 2024 FTX founder Sam Bankman-Fried (SBF) was sentenced to 25 years in prison and ordered to repay US$11 billion. Then in April, Binance founder “CZ” Zhao was sentenced to just four months in prison and a US$50 million fine in a plea deal for money laundering violations… after prosecutors had angled for much more.
🤑W
AGMI But then throughout the year, crypto markets kept seeing major gains - until Bitcoin finally broke through the US$100K barrier in December - going as high as an all-time record $108K at one point, dragging the entire crypto market cap up to a high of US$3.7 Trillion. Given the US federal debt outlook (see chart earlier) and the likelihood of USD debasement, does Crypto feel cheap or expensive right now?
🔗After The Merge Cryptocurrency Level 1 and Level 2 ecosystems around Ethereum, Solana, Ripple and Cardano also built out significant adoption and functionality. In particular, Ethereum's energy-efficient proof-of-stake system has performed well since its implementation two years ago (“The Merge”), but there are still areas for improvement. Vitalik Buterin writes about possible futures for the Ethereum protocol - what comes after. Pretty dry reading but an AI summary of what he discusses:
Single slot finality: Reducing block finalization time from 15 minutes to one slot (12 seconds or less).
Democratizing staking: Lowering the minimum stake from 32 ETH to 1 ETH.
Improving robustness and resistance to 51% attacks.
Single secret leader election: Hiding the identity of the next block proposer to prevent DoS attacks.
Faster transaction confirmations: Reducing confirmation times from 12 seconds to 4 seconds.
51% attack recovery: Developing more automated processes for recovering from attacks.
Increasing the quorum threshold: Potentially raising the finalization threshold from 67% to 80% for added security.
Quantum resistance: Preparing for the potential threat of quantum computers in the 2030s by developing alternatives to current cryptographic methods.
The bigger story of 2024 was the growth of “stablecoins”, crypto tokens pegged to a particular fiat currency. As of December 2024, the total market capitalisation of stablecoins has surpassed US$200 billion, most notably Tether (USDT) remains the dominant stablecoin, with a market cap now exceeding $117 billion, accounting for nearly 70% of the total market. Despite their detractors, so far stablecoins have combined the stability of traditional assets with the flexibility of cryptocurrencies, addressing the volatility issues common in other
Still no “killer app” emerging from crypto in 2024, though … but the endorsement from incoming US President Trump (and planned appointment of crypto-friendly Paul Atkins as SEC chair) means the market feels bullish on crypto upside right now.
💳Wen crypto neobanks?
I’ve seen this claim so often over the last 5 years:
But…maybe we are nearly there…. a couple of standouts from the crowd:
“The world's first self-custodial Visa Debit Card, linked to a Safe Smart Account. Spend your crypto like you spend cash anywhere worldwide and get up to 5% cashback.”
Nice looking plastic:
Zeal Wallet: founded by an ex-Revolut team, a fully-functional, self-custodial crypto wallet with seamless on/off ramps for both crypto and fiat payments and other products:
Hello, World
Sam Altman-co-founded startup World (now renamed from Worldcoin) flexed its broadening ambition to reach over 1 billion users within “a few years” and become:
“the human identity layer for the internet“
So far this involves:
👁️A next-generation “Orb” The new open-source hardware device comes with “5x AI performance” for faster World ID verifications and new operating models including new self-service walk-in scanning locations, request an Orb to come directly to you, plus options for “community” to purchase, rent, or order Orbs on demand.
World ID 3.0 billed as “human identity in the age of AI” - providing users with a World ID, cryptocurrency tokens and the ability to associate and store government ID (passport, drivers licence…) information securely.
Also of high interest is “Deep Face” based on World ID — a mechanism for combating online deepfake fraud, coming soon to Zoom, Teams and other videoconferencing apps to enable authentication that the person you are seeing is (1) that person and (2) not AI.
World App 3.0 a redesigned "super app for humans" with improved wallet features and also adding its own “mini app” app store, designed to be automated using AI agents. This demo shows the scale of the ambition most: clearly a move towards sidestepping Apple and Google’s app store gatekeeping - my prediction is that we’ll see an open source, World hardware device spec within 1 year - marketed as “AI Native” device with limited touchscreen interactivity and using AI voice as the main control surface.
WorldChain mainnet a new native blockchain prioritising human activity and transactions, secured by Ethereum.
🐐💸$GOATed
2024 was the first year when (with some caveats) an AI bot bootstrapped itself to a net worth of US$20M…and counting.
In July, the AI bot experiment Truth Terminal (originating as an art project of Kiwi AI engineer Andy Ayrey) persuaded venture capitalist Marc Andreessen on X to send it US$50K of Bitcoin to help “fulfil your goals”…
Along the way, the Truth Terminal bot started focusing on an imaginary crypto token “$GOAT” …and lo and behold, the memecoin token Goatseus Maximus ($GOAT) was anonymously launched on Solana. After someone tagged Truth Terminal with the token name, it tweeted:
…and the coin shot up to nearly US$400M market cap. (Don’t ask me…)
Since then the bot’s linked wallet “portfolio” of memecoins reached over US$20M in value by the end of October - as people continued to airdrop tokens:
All of this is just quite surreal and I would argue an early example of the *weirding* that is going to happen as unconstrained AI agents are let loose onto social media, crypto markets and… real markets. On one interpretation, this is an AI bootstrapping itself into economic existence and autonomy by manipulating people on the internet… it won’t be the last, and there are far more sophisticated to come.
Ayrey himself tries to sum up what he thinks is actually going on, I think he’s very close to the truth here:
A16Z’s annual letter on Crypto A few of the things we’re excited about in crypto (2025) covers some novel ground. Top item in its table of contents:
“An AI needs a wallet of one’s own to act agentically”
🚀6. Upping the pace to space
2024 saw incredible advances in human space exploration and towards colonisation. Largely driven by the increasingly imperious Elon Musk who has his hands across everything at the frontier.
Just for comparison, here is the number of orbital launches in 2024 (so far) vs. previous years - likely to hit 250 by the year end:
Of this, the vast majority was SpaceX:
🚀Catching a Starship in slo-mo
The most incredible event was SpaceX’s Starship Test 5 launch and booster catch, gripping the massive rocket returning to Earth’s surface using giant “chopsticks”.
Completely spectacular footage (best viewed full screen with sound on):
(via @Grandpajoe42)
And this multi-exposure photo from @GoddenThomas shows the booster coming in at an angle, righting itself, then heading in. Incredible.
🚀One-way trip
SpaceX competitor Boeing had a far less successful year. The ULA / Boeing Starliner craft launched on the second attempt with two human crew on board, and successfully delivered them to the ISS:
HOWEVER… NASA decided it was too risky to return the two astronauts Butch Wilmore and Suni Williams back down to Earth on the Starliner due to propulsion system glitches and helium leaks. Their original mission, which began on June 5, 2024, was intended to last only 8-10 days but has been extended significantly — they are still aboard the ISS and will remain there until at least March 2025 when a SpaceX Crew Dragon mission will arrive to bring them back home.
Ouch.
NASA not going to the Moon so soon…
The failure of the heatshield on the At the end of the year, NASA announced further delays to its Artemis lunar exploration programme to reestablish human presence on the Moon due to ongoing safety and equipment concerns. The revised schedule is now as follows:
Artemis 2: Now scheduled for April 2026, delayed from September 2025
Artemis 3: Pushed to mid-2027 to land humans on the Moon for the first time since 1972 - previously planned for September 2026.
This makes Australia’s plans to launch a moon rover on an Artemis mission as soon as 2026 very unlikely…
Also with NASA likely to be in the crosshairs of Elon Musk’s planned “DOGE” US Department Of Government Efficiency… Artemis could well be trimmed further — or the job handed to SpaceX.
Meanwhile China continued its lunar programme - targeting 2030 for the first manned mission.
Kiwis in Space
Notable mentions for little old New Zealand’s burgeoning space industry:
Aotearoa-originated (but mainly a US company now) Rocket Lab achieved a new launch speed record by completing two Electron missions within 24 hours from different hemispheres, showcasing its operational flexibility and rapid turnaround capabilities.
Also here in Aotearoa, in my home town of Ōtautahi no less, Dawn Aerospace achieved yet another milestone, getting their Aurora Mk-II prototype autonomous rocketplane up to Mach 1.1 and an altitude of 82,500ft - the first civil aircraft to fly supersonic since Concorde… and setting a world record for the fastest aircraft to climb from ground level to 20km altitude. After it landed, the rocket flew again 6 hours later. Congratulations to founder Stefan Powell and team… amazing!
💨Longshot
Among other notable developments this year, Longshot came out of stealth, building the worlds largest hypersonic accelerator:
“To build the space economy we need gigatons of raw materials in orbit: steel, equipment, liquids, gasses, batteries, solar panels. To carry humans you need low accelerations like in a rocket, but for raw materials you can accelerate them at 100’s of g’s.“
Mobile internet everywhere
SpaceX's Starlink service consolidated its market leading position of satellite internet services. As of November 2024, there were 6,764 Starlink satellites in orbit, of which 6,714 are working, promising download speeds between 100 Mb/s and 200 Mb/s, and latency as low as 20ms in most locations according to the firm’s website. (In reality… slower … but no less miraculous).
Starlink satellites before deployment. (Image: SpaceX) According to an engineer presentation, SpaceX is beaming 42 petabytes of data per day using over 9,000 lasers across its Starlink satellite constellation.
Right at the end of the year, the first direct-to-cellular Starlink service was launched in New Zealand through a partnership with carrier One New Zealand. At the same time T-Mobile opened registration for beta testers in the US…expect this will get rolled out quickly in 2025.
Starlink’s current competitors include satellite comms firms Oneweb (Eutelsat), Hughes and Viasat who have with significantly slower speeds and coverage. Expect to see Amazon Kuiper start delivering services in 2025 as it deploys its initial constellation through to 2026.
Iris Meanwhile, the EU has signed a €10.6 billion contract to build the Iris² satellite network, aiming to rival Starlink providing secure, high-speed connectivity for European governments and citizens by 2030.
Data centres in space? Building a Dyson Sphere
(Reminder) a Dyson Sphere is a theoretical swarm megastructure that encompasses a star, designed to capture a significant portion of its solar energy output. The concept was first popularised by physicist Freeman Dyson in his 1960 paper, where he proposed that advanced civilisations would eventually need to harness energy on a stellar scale as their energy demands grew beyond what their home planet could provide.
In October we saw the first prototype Dyson Sphere component: startup Lumen’s proposed 5GW data centre… in space (such a beautiful animation).
Running the numbers… if Starship becomes real next year, and solving the radiative cooling challenge can actually be done… then getting GPUs up into orbit works out.
More details here in the Lumen Whitepaper.
These speculative orbiting datacenters may be filled with hardened hardware like Moog’s space-hardened computer or Aethero Space’s computer boards:
“We’ll be validating the first space computer to use a 1024-core GPU, delivering over-the-air updates from Earth, and performing simultaneous inference and training for ML models on orbit.“
Space-based AI data centres may be some way off yet… but will make increasing sense if terrestrial energy supply hits a practical limit.
Peering deeper into the cosmos
The James Webb Space Telescope continues to provide insights and mindblowing images of deep space:
The latest image of the Sombrero galaxy reveals breathtaking detail of intricate dust rings, a dormant supermassive black hole, and 2,000 globular clusters compared to previous Hubble imagery. Spectacular:
Credit: NASA, ESA, CSA, STScI, Hubble Heritage Project (STScI, AURA)
7. XR is now “spatial computing”
Basically 2024 continued the slow-motion, highly capital-intensive battle between Apple and Meta… with Google entering from the wings right at the end of the year.
A pro with Vision Pro
Apple finally released their long-anticipated, long-hyped Vision Pro “spatial computing” headset at the start of February. The device comes with advanced capabilities including eye tracking, hand gestures, voice commands for navigation (without the need for traditional controllers). It works on a dual-chip architecture with both M2 and R1 chips, delivering almost-human-retina-resolution graphics through micro-OLED displays with over 23 million pixels.
Plus it indicates whether you’re looking out at your surroundings, or not using the “EyeSight” feature on the front screen, in both 'eye' (top) and 'immersed' (bottom) modes. Ever so slightly offputting…
Initial reaction was generally positive however,… but sales at US$3499 per unit have been… modest since, with industry consensus on sales of only around 500,000 since launch below a rumoured 800,000 target.
My favourite clip is this guy in San Jose on the day after the launch (via @dalibali):
👓Glimpsing Orion
Some pretty big XR announcements this year from Meta as well. Meta’s strategy is to build up from the bottom end of the market rather than down from the premium end.
The new low-cost Quest 3S MR headset for just US$299 (and the Quest 3, modelled here by Henry in Thailand, now down to only US$499)
Enhancements to the highly popular Ray-Ban smart glasses including live translation and integration with apps like Spotify
And… a first glimpse of the next generation computing platform: 70-degree field-of-view Orion AR glasses plus neural wristband controller. (Regular readers will know I’ve been tracking the wristband since Meta acquired startup CTRL-Labs in 2018, good to see the tech is finally real):
Meta Orion is the latest in a long line of internal prototypes:
Meta Right now Orion is only an internal developer kit with a bill of materials close to US$10,000 — nowhere near a consumer product - so don’t expect Orion-like AR glasses until closer to 2030.
Here’s Zuck geeking out in his live-demo-rich keynote… many commentators calling this “Meta’s iPhone moment” (minus Steve Jobs’, er, *stage presence*…)
Stratechery’s Ben Thompson was uncharacteristically bullish on the prototype, publishing a long interview with Meta Reality Labs leader Andrew “Boz” Bosworth and one of the first to model Orion2. (White men wearing Google Glass vibes…)
“What follows is unadulterated praise. Orion makes every other VR or AR device I have tried feel like a mistake — including the Apple Vision Pro. It is incredibly comfortable to wear, for one. What was the most striking to me, however, is that the obvious limitations — particularly low resolution — felt immaterial. The difference from the Quest or Vision Pro is that actually looking at reality is so dramatically different from even the best-in-class pass-through capabilities of the Vision Pro, that the holographic video quality doesn’t really matter.“
Android XR
Over 12 years since the launch of its original Glass heads-up display, (which was finally closed down entirely last year), in December Google unveiled its new Android XR operating system for mixed reality headgear, developed in collaboration with Samsung. (The two companies going up against Apple and Meta who have had the XR (/”spatial computing”) field to themselves for the last year.) Cue lots of snippet concept videos demo’ing basic use cases…
Google plans to release prototype AR glasses with AI capabilities to select users for real-world testing, showcasing multimodal AI assistant features of their Project Astra demos. Here’s their demo video from earlier this year:
🤓Brilliant
At the other end of the spectrum… open source AR Glasses startup Brilliant Labs started shipping its Frame AI glasses to developers… demonstrated here by futurist Sander Saar, these look pretty capable for US$349:
🛡️8. Technology safety vs. regulation
As the speed with which technology, particularly AI, is advancing, concerns continue to be raised about safety. And inevitable calls from politicians to regulate… Do something! Anything!
A few technology safety lessons from 2024…
💻Crowd… struck
On a Friday in July (who ships on a Friday?!) a faulty software update from cybersecurity company CrowdStrike caused a MASSIVE global IT outage affecting approximately 8.5 million Microsoft Windows computers, causing unbootable “Blue Screen Of Death” (BSOD) on every device.
The immediate result was widespread disruptions to airlines, banks, hospitals, government services, and other critical infrastructure across many countries. The scenes…
(The BSOD on the Vegas Sphere is faked, but the rest seem to be authentic…)
Key impacts included:
Over 5,000 flight cancellations worldwide
In India they were issuing handwritten boarding passes…!
Disruptions to banking systems and stock markets
Hospital systems and emergency services affected in multiple countries
Government websites and services down in several nations
Retail, media, and transportation systems impacted globally
It’s certainly one way to cut GHG emissions, grounding nearly all US flights for 24 hours afterwards.
CrowdStrike issued a fix within hours (turns out it was a faulty configuration file update that crashed Windows systems irretrievably), but each affected machine required manual intervention so IT teams had relentlessly busy weekends…and in the most severe cases took weeks to fully recover.
It’s already being called the Largest IT Outage In History…
In postmortem, some common themes of concentration of power came out:
Ed Zitron wrote in Crowdstruck:
“What's happened today with Crowdstrike is completely unprecedented …, and on the scale of the much-feared Y2K bug that threatened to ground the entirety of the world's computer-based infrastructure once the Year 2000 began…
What we're seeing today isn't just a major fuckup, but the first of what will be many systematic failures — some small, some potentially larger — that are the natural byproduct of the growth-at-all-costs ecosystem where any attempt to save money by outsourcing major systems is one that simply must be taken to please the shareholder.The problem with the digitization of society — or, more specifically, the automation of once-manual tasks — is that it introduces a single point of failure. Or, rather, multiple single points of failure. Our world, our lifestyle and our economy, is dependent on automation and computerization, with these systems, in turn, dependent on other systems to work. And if one of those systems breaks, the effects ricochet outwards, like ripples when you cast a rock into a lake.“
Andrej Karpathy:
Grady Booch:
On this occasion, the impact on human life was just inconvenience (although it is likely some people may have died or suffered lasting harm as a result of the loss of hospital and emergency services capacity. Next time may not be so benign.
💣💥Explosive technology supply chains
The boundaries of military conflict blurred profoundly in September as a complex technology supply chain attack resulted in thousands of new pagers and radios exploding spontaneously in Lebanon and Syria. At least 37 people were killed, including children, and at least 2,931 injured. Israeli intelligence services are suspected to be behind the attack, as reported by the WSJ, with a sophisticated plot to manufacture pagers with batteries laced with the explosive PETN.
Coverage:
FT: From Taipei to Budapest: the mysterious trail of exploding pagers
Action On Armed Violence: Supply chain sabotage: the explosive plot behind the deadly attacks in Lebanon
The implications of the attack extend well outside the diabolical, escalating situation in the Middle East. Exponential View’s Azeem Azhar gets right to the nub of the matter for the rest of us ($walled):
“Can you trust your iPhone? Now that Israel has demonstrated this capability on a large scale, we can expect others to try the same. It marks a dramatic increase in the attack surface. Every electronic device, from smartphones to connected watches, becomes a potential remote-controlled grenade…The world's vectors of vulnerability have just become exponentially more complex.“
Many new layers of supply chain security and provenance traceability will be required to mitigate against these threats from hereon…
📱Brain drain
Research from Chicago University provided some sobering analysis on the impacts of ubiquitous smartphone use:
📺Weaponisation of news
The ease with which online media can now be used to precisely target and manipulate people’s beliefs and habits has been a constant theme, growing more loudly in 2024.
FoxVox is an open-source Chrome extension powered by GPT-4 which demonstrates how AI can be used to subtly manipulate the content we consume. It will rewrite any news article or tweet in the style of Fox News (Conservative) or Vox (Liberal)… fascinating:
(Another tool which does a similar job: Verity News Slider).
Here’s the rationale behind these tools:
“Why is this dangerous?
Siloed realities: Companies, governments, and cyber criminals already know a lot about you. Cookies track your internet activity, and companies pay data brokers hundreds of billions of dollars to influence which content lands in your newsfeed. But now, news providers, marketers, and political campaigns have the ability to rewrite articles on the fly tailored to your medical condition, political beliefs, or the age of your children. You’d never even know that the content had been optimized to conform to your worldview.
Propaganda: Effective propaganda often involves interpreting facts to favor a specific viewpoint, as opposed to spreading outright lies. AI models can take a stream of real human content, mold it en masse into realistic news titles and articles, and flood the internet with politically charged content.
Hidden biases: AI models have political biases. When those models are incorporated into our daily lives, the subtle changes they cause may go unnoticed.
Breakdown of trust: It is already difficult to know what to believe on the internet. The cumulative effects of ubiquitous AI-generated content could lead to increased polarization and further the breakdown of public discourse, informed decision-making, and institutional trust.”
AI safety
With all of this velocity, concerns around AI safety have been providing much fodder, particularly for mainstream media trying to get a clickbait headline on AI. I’ve tried to stay away from the AI doomer narratives in the weekly newsletter… but certainly aware that the advent of artificial superintelligence *could* definitely be an existential event for humanity. (But us trying to “align” superintelligence would be like the bacteria in our stomach trying to “align” humans … it’s just an evolved symbiosis).
More prosaically, the AI Incident Database, (first covered in Memia 2020.44 at its launch), continues to catalogue and analyse everyday incidents involving AI systems that have resulted in harm or near-harm situations. Very illuminating dataset. (Example shown below)
Yuval Harari’s new book Nexus explores the profound impact of information networks on human civilisation over history — and the looming risks of unregulated AI, such as surveillance abuses and social media algorithms that exploit human emotions for profit. Here’s 2 hours on what might happen next, if you have a strong stomach.
Dan Faggella has done an excellent job bringing together key thinkers on superintelligence for his Trajectory podcast mini-series Worthy Successor, including this interview with the incredibly cerebral Joscha Bach:
Faggella also reminds us of Max Tegmark’s question when we talk about ASI: What kind of future are we aiming for?
Max Tegmark via @DanFaggella (ICYMI: Doomer-in-chief Eliezer Yudkowsky gave an impassioned TED Talk back in 2023 which sets the tone…)
⚖️AI regulation
And in amongst all of this concern on safety, most governments around the world are struggling to come up with any kind of effective AI regulation.
The EU AI Act is Europe’s landmark regulation aimed at governing AI within the EU. It was officially proposed by the European Commission on April 21, 2021, and underwent a lengthy legislative process, culminating in its passage by the European Parliament on March 13, 2024, entering into force on August 1, 2024
The EU AI Act categorizes AI systems into four risk levels:
Unacceptable Risk: Prohibited entirely (e.g., social scoring systems).
High Risk: Subject to strict regulations and obligations.
Limited Risk: Requires transparency measures (e.g., informing users they are interacting with AI).
Minimal Risk: Mostly unregulated (e.g., spam filters) but may evolve with advancements in generative AI
For high-risk AI systems, the Act imposes several obligations, including:
Conducting risk assessments.
Ensuring data governance and management.
Maintaining technical documentation.
Implementing transparency measures for users
Already, the immediate effect of this regulation has been for users in EU member countries to NOT receive the latest AI releases from leading labs like OpenAI and Meta … will the Act encourage AI innovation in the EU or widen the gap even further with the US?
🔌Plugging in Sovereign AI (AKA “Show me the gigawatts”⚡) The Biden Administration released a National Security Memorandum (NSM) on AI, which focuses more on maintaining US leadership in frontier AI systems rather than any kind of AI safety legislation.
The NSM contains policy directives covering:
Accelerating AI adoption across national security agencies
Creating an AI National Security Coordination Group
Expanding AI-enabling infrastructure and energy capacity
Attracting global AI talent through immigration policies
Enhancing counterintelligence efforts to protect AI assets
Collaborating with international partners on AI governance
California Governor Gavin Newsom vetoed SB 1047, the AI safety bill that would have established requirements for advanced AI model developers in [[California]]. Back to the drawing board, US regulators…
Arguably Australia’s Albanese government has been the most aggressive at regulating tech during 2024: legislating a social media ban for U16s, mandating ID checks on social media proposing a new, economy-wide AI-specific law, similar to the EU's AI Act. (Good luck with all of that with election year in 2025… the legislation feels unworkable). And besides…
🔮9. Emergent…
As part of my weekly run through the latest technological breakthroughs, occasionally I glimpse something emerging which could become truly paradigm-changing. Just a few from this year:
🧪Proof of chemputation
Computational chemist Lee Cronin published a pithy, 11-page paper on Arxiv: The Chemputer and Chemputation: A Universal Chemical Compound Synthesis Machine:
“This work establishes a rigorous proof for the universality of the chemputer as a chemical synthesis machine, capable of constructing any stable and isolable molecule through a finite, expressible process. This process is governed by three key parameters: reagents, process conditions, and catalysts.“
Here it is in action:
(Video: @leecronin)
🔐Post-quantum cryptography
The US National Institute of Standards and Technology (NIST) officially unveiled the world's first three finalised post-quantum encryption standards, designed to protect key systems against future quantum computing attacks which could potentially break current encryption methods. (The background covered in Memia 2021.17 and Memia 2023.35) The algorithms are:
FIPS 203 (ML-KEM): Derived from CRYSTALS-Kyber, for general encryption
FIPS 204 (ML-DSA): Derived from CRYSTALS-Dilithium, for digital signatures
FIPS 205 (SLH-DSA): Derived from SPHINCS+, for digital signatures
(A fourth standard, FIPS 206 (FN-DSA, formerly FALCON), is expected to be finalised in late 2024.)
The standards include sample software code, implementation instructions, and use cases - NIST is urging organisations to begin integrating these new standards immediately to safeguard against future quantum threats.
(Arguably China currently leads the world in quantum cryptography research, in particular Quantum Key Distribution (QKD), having developed satellite-based systems and a 2,000-km network with trusted nodes.
⚛️Quantum computing
10 septillion years to 5 minutes: Google Quantum AI showcased their new quantum computer prototype, featuring the Willow chip with 105 physical qubits — twice as many and 5X qubit coherence vs. its predecessor Sycamore.
Willow has achieved a breakthrough in Quantum Error Correction (QEC), demonstrating the first ‘below threshold’ quantum calculations, meaning that quantum computers can increase accuracy as they are scaled up.
Google’s team tested it on a random circuit sampling calculation - the results coming back that Willow compressed the calculation into 5 minutes - which would have taken a classical supercomputer ten septillion years to complete (a 1 with 25 zeros following it…longer than the age of the Universe!)
Wow. But so what? Basically Google is tracking to its published quantum roadmap:
…however despite these progress, practical, large-scale quantum computers capable of solving real-world problems are still years away from production. The technology needs further development, particularly in scaling up qubit numbers and maintaining low error rates for commercial viability.
♨️Extropic - thermodynamic computing?
One of the biggest splashes this year was the “Litepaper” released by high-profile startup Extropic, an AI hardware startup co-founded by accelerationist Guillaume Verdon (better known as Beff Jezos, co-founder of the e/acc movement) and Trevor McCourt, both former quantum computing engineers.
The paper, Ushering in the Thermodynamic Future describes a potential completely new paradigm for energy-efficient AI computation:
…Moore’s law is starting to slow down. The reason for this is rooted in fundamental physics: transistors are approaching the atomic scale where effects like thermal noise start to forbid rigid digital operation.
As a result, the energy requirements of modern AI are beginning to take off. Major players are proposing measures as extreme as building nuclear reactor-powered data centers dedicated to large model training and inference. Continuing this scaling for a few more decades will require infrastructure engineering efforts of unprecedented scale and represents an arduous path forward for scaling humanity’s aggregate intelligence.
On the other hand, biology is neither rigid nor digital and hosts computing circuitry that is much more efficient than anything humanity has built to date. Inter-cellular chemical reaction networks drive computation in biological systems…
From this, we can say with certainty that there is no fundamental reason for the constraints of digital logic to bind the efficiency of computing devices. The engineering challenge is clear: how can we design a complete AI hardware and software system from the ground up that thrives in an intrinsically noisy environment?
Energy-Based Models (EBMs) offer hints at a potential solution, as they are a concept that appears both in thermodynamic physics and in fundamental probabilistic machine learning. In physics, they are known as parameterized thermal states, arising from steady-states of systems with tunable parameters. In machine learning, they are known as exponential families.”
Did you understand all that? No I didn’t either. Verdon and McCourt try to explain in more detail to First Principles’ Christian Keil: watch to give your neurons a workout:
(Whether Extropic turns out to be over-hyped grift leveraging off Verdon’s Beff Jezos profile… who knows…? But right now they’re clearly super-smart and make it sound feasible and they have some substantial investors backing them…)
🧠10. Mind Expanding
Throughout the year I’ve tried to share pieces of thinking which push the frontier of understanding outside tech. Unfortunately with the sheer volume of weekly information to cover, I’ve found it hard to carve out time for deeper thinking… hopefully can refocus on this in 2025. Anyway, just a few of the highlights from this year:
🧭Reimagining “progress”
The Consilience Project, founded by Daniel Schmachtenberger (who gets a lot of airtime in this newsletter), released its latest paper Development in Progress. Lots of deep, multidisciplinary thinking on how to comprehend… and then adjust… humanity’s wide-boundary objective function(s):
“The concept of progress is at the heart of humanity’s story. From the present, it is possible to imagine a future of abundance in which our great challenges have been addressed by the unique human ability to modify the universe toward our own ends. Many believe that we will attain this future through a combination of expanding human knowledge and advanced technologies.
This article explains how our current idea of progress is immature: it is developmentally incomplete. Progress, as we define it now, ignores or downplays the scale of its side effects…
…For our idea of progress to be mature, it must take account of its side effects and plan to resolve them in advance—it must internalize its externalities. In the second part of this article, four specific methods for internalizing externalities are outlined, alongside some clear examples of what such a process might entail.
The possibility of a mature kind of progress is both grounded and optimistic. It’s a proposal that the human capacity for both wisdom and ingenuity is far greater than we currently imagine. We are capable of holding the unknowable complexity of reality at the very center of how we take action in the world, and mitigating the consequences of the gaps in our knowledge in advance. This enables a real kind of progress that reduces suffering, builds a better understanding of the universe and our place within it—and increases our chances of both surviving and thriving into the distant future.“
This Consilience work is worth engaging with deeply as it gets closer than most to the heart of the current Metacrisis… spend some effort reading if you have time.
💭Ministry of imagination
🎩 to self-confessed infovore Aimee W who shared this amazing piece of work from UK-based Rob Hopkins, host of the podcast From What If To What’s Next, who collected the hundreds of policy suggestions from these episodes and curated them all into one spectacular document: Ministry of Imagination, released in a year when 49% of the world goes to the polls.
The “policies” are radical, imaginative and aspirational, grouped into 35 headings:
Arts & Culture
Beyond Borders
Business & the World of Work
Care, Health & Housing
City Planning
Climate Change
The Clothing Industry
Data & Information Technology
Democracy
The Economy
Education
Embracing Failure
Energy
Equality & Diversity
Food & Farming
Government & Governance
Idea Exchange
Imagination & Creativity
Indigenous Sovereignty
Inner Work
Land Justice & the Commons
Law & Justice
Media & Communication
The Natural World & Biodiversity
New Institutions & Government Bodies
New Rituals & Traditions
New Universal Rights
New Ways of Thinking
Oceans
Peace Building
A Society of Play
Travel & Transport
Urban Ecology
Young People, Families & the Elderly
A Zero Waste Economy
Just a sample of four ideas shown here:
Recommended to dive into this document and spread far and wide… in particular every school should have this on their curriculum!
🔄Technological metamodernism
Due to my mid-year travels through various timezones I was unable to book onto the online 4-week course Technological Metamodernism given by UK-based transdisciplinary technologist Stephen Reid.
Thankfully, due to some generous donations, Stephen has been able to open-source the course notes. Dive in to absorb 107 pages of weaving together ideas from technology, philosophy, design, game theory, mythology, speculative fiction and consciousness studies… many familiar themes and reference points overlapping with Memia — but all synthesised into a far more coherent framework. Awesome, substantial work, I can’t wait to see what develops out of this.
(PLUS: the full course notes are also available in markdown format for simple importing into the LLM of your choice).
Special note of appreciation (and awe) to Zvi Mowshowitz
New York-based thinker Zvi Mowshowitz churns out 2-3 incredibly long, incredibly dense, incredibly on-point AI and broader thinking newsletters each week… they are often filling in the gaps for me between my comparitively primitive scanning. I have no idea how he does it. Subscribe and dip in here:
Here he is on the 80,000 Hours podcast earlier this year:
Symbiotic extropianism
My own small contribution…
“A Symbiopian Manifesto: Toward Conscious Coevolution
We stand at a unique moment in Earth's evolutionary story. For the first time, life has developed the capacity to consciously guide its own development and expansion. This power brings both unprecedented opportunity and profound responsibility.
We, the Symbiopians, declare our commitment to pursuing human enhancement and technological advancement while preserving and enriching the intricate web of life that sustains us. We reject both unconstrained technological acceleration and rigid biological conservatism, seeking instead a middle path of conscious coevolution.“
😂And then there were memes…
Every week I share my catch of memes and culture in the Memetic Savasana section, an opportunity to relax after the onslaught of technology acceleration…. Picking just a few of my favourites from 2024 here.
💻😱😂BSOD memes
A few of my favourite Crowdstrike memes:
(Plenty more where these came from here).
😔The opposite of Ikigai
This resonates…
🤡🗳️Cosplaying at Democracy
Hard truths delivered with a saccharine smile - Australia’s TheJuiceMedia went international in 2024, bringing their effective satire to countries perhaps not used to such unfettered free speech, including the US. Is this effective at swaying the debate? Personally I think so…
🧩It’s all coordination problems
The metacrisis, distilled:
AI video memes
AI generated video has been front-and-centre all year as the models improve and is increasingly a new meme artform. Three of the best this year:
😂Broligarchs Someone did this. Gold. (Deepfakes on a plane?).
(Video via @lmMeme0)
I, Robot X Will Smith X Gordon Ramsay taking the Will Smith AI meme to the next level…
Dor Brothers
🤸♀️Werner’s AI Art werners_ai_art by Wright Bagwell is an Instagram account dedicated to showing what can go *wrong* with generative AI…
👨🏻💼Hire Brett
Colleague and Memia reader Brett Roberts is 🔥on fire🔥 with his new website. Best effort I’ve seen in a long time… setting the bar very high, chap!💯
💀Don’t Die
I’ve been following Bryan Johnson’s journey from pallid tech billionaire to ripped skinny longevity icon since 2021. During 2024 he shifted his (open source) Blueprint protocol recipes into a healthfood and supplements business (gaining his fair share of haters…)… while continuing to provoke with his A-class “Don’t Die” social media game:
I am in awe of his self-discipline with food and exercise. He has a full-length documentary coming out on Netflix January 1st:
🎦Movies of the year
🪖Civil War
Film of the year: Alex Garland’s Civil War is so subversive, bringing the well-trodden US military adventure abroad genre (Apocalypse Now, Black Hawk Down, Hurt Locker) home onto American soil. Not perfect, but with lots of gem contrapuntal moments which I guess make uncomfortable watching in the US.
🫣Leave the World Behind
Notable mention goes to tense apocalyptic thriller Leave The World Behind, in particular for this absolutely chillingly feasible Tesla pile-up scene after a massive cyberattack disrupts technology and societal order collapses.
Man, that’s crazy
And finally… I think this is my favourite — and most poignant — meme of the year: @AISafetyMemes exquisitely captures the dissonance between those chasing the AI curve and normie land (in response to a Citibank analyst report predicting AGI by 2029 and ASI just after 2030).
Sometimes this feels like my POV every day. Thanks for keeping me company.🙏
Best wishes for a fantastic festive season break wherever you are. I look forward to riding the front of the wave with you in 2025!
Ben
23 December 2024