Scaling wobbles📈📉 trillium⚡️ purple states of America🟣🇺🇸 magentic-one🤖 hiked haiku💸 SciMuse🧪💡 COP29 sting🌡️ migrations in motion🦅 #2024.45
Synthesized, not plagiarized
Welcome to this week's Memia scan across emerging tech and the exponentially accelerating future. As always, thanks for being here!
ℹ️PSA: Memia sends *very long emails*, best viewed online or in the Substack app.
🗞️Weekly roundup
The most clicked link in last week’s newsletter was my Strategy note: 🌐the ChatGPT Search ^ Perplexity ^ Gemini "grounding" inflection point — I continue to use all three AI search engines, they definitely each have their strengths and weaknesses. But for sure, traditional web search is goneburger…
⏩🎄📚Xmas book offer
We’re currently running a Christmas 2024 offer of NZ$20 off RRP for paperback orders for my book ⏩Fast Forward Aotearoa — an ideal Christmas gift for anyone interested in exploring New Zealand's future with exponential technology. Get yours before 18th December!
(Reach out directly to discuss wholesale orders for corporate gifting).
📈The week in AI
Straight onto the week's AI news and releases.
📈📉Scaling wobbles
Much of the perceived value underpinning current stratospheric AI industry valuations is continued performance against so-called “scaling laws”: throw an order of magnitude more compute at AI model training and you’ll get a commensurate return on intelligence. In particular, Nvidia’s record share price is predicated on future demand for its GPUs to support this scaling. OpenAI raised at a US$150Bn valuation just a couple of weeks ago, also largely on this premise.
However… just in the last couple of weeks (surely no timing coincidence…!) rumours have started to percolate that the latest attempts at scaling the next generation of AI models have not, so far, yielded the hoped-for results:
Andreessen and Horowitz discussing a flattening out of model performance on their podcast:
Confirmed by The Information reporting OpenAI changing their direction towards more inference-based “reasoning” scaling rather than training-time:
Industry insider Yam Peleg stoked the fire:
Professional AI industry naysayer Gary Marcus couldn’t hold back…
“…as I have been saying since 2022’s “Deep Learning is Hitting a Wall”, scaling laws” are not physical laws. They are merely empirical generalizations that held for a certain period time, when there was enough fresh data and compute… If the rumor is verified, there could be the AI equivalent of a bank run.”
Rumours are rumours, but imminent AGI may be delayed…🍿🍿🍿
🗞️AI industry news
Scanning across other smoke signals from the AI sector…
AI chipmaker Nvidia overtook Apple to become the most valuable company in the world by market cap. (Tesla had a good week too, something to do with an election I heard…)
After a brutal reversal of fortunes, Nvidia replaced Intel on the Dow Jones Industrial Average index…
However…Nvidia’s incredible sales numbers (which are pumping the valuation) may not be entirely transparent: an FT editorial looks Inside the Murky New AI chip Economy: 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…:
”…the deals may stretch valuations in the sector. The precise details of the arrangements between Nvidia and the neoclouds are unclear. But the chipmaker is itself an investor in some of the start-ups, which are in turn among its largest customers. Armed with Nvidia chips to secure loans, the cloud providers can then use the capital to buy more chips from Nvidia. This dynamic could inflate Nvidia’s earnings, and means the neocloud groups risk becoming highly leveraged, too.“
🤔Hmmmm…. GPU futures as the new tulip mania? (🎩 Graham G for sharing).
If you can ignore the Silicon Valley VC bro vibe, this free-flowing podcast interview with Nvidia Founder and CEO Jensen Huang is a good listen… lots of nuggets on how he sees things playing out in here.
Just three key insights from Huang:
He expects distributed AI model training to ultimately be solved, mitigating the need for larger and larger data centers over time
Last year’s AI training data center is this year’s AI inference data center… so they will not be stranded assets as the chips become orders of magnitude more powerful
His hypothesis is that GenAI, Agentic AI will create $trillions of new market value in future… so NVIDIA’s future market is *effectively infinite*. The man at the top of his game thinks about the future unconstrained by the present:
“…companies are only limited by the size of the the fish pond: a goldfish can only be so big and so the question is “what is our fish pond?” And that requires a little of imagination. And this is the reason why market makers think about that future: without creating that new fish pond, it's hard to to figure this out looking backwards and try to take share right? You know share takers can only be so big for sure; [but] market makers can be quite large ….and so you know…the good fortune that our company has is that since the very beginning of our company we had to invent the market for us to go swimming“
OpenAI splashed out and purchased the domain chat.com (previously sold for US$15.5M) in exchange for shares:
(Not) OpenAI’s CFO:
(Actually OpenAI’s CFO) Sarah Friar said that ChatGPT currently has 250 million weekly active users and 75% of its US$3.7Bn annual revenue comes from paying consumers - pretty impressive for a company which had next to zero revenue just 2 years ago! (Although also remember it made a US$5Bn loss last year…)
Anthropic announced a new partnership with Palantir and AWS to provide Claude to US intelligence and defense agencies for processing information up to and including “Secret” classification.
This is the same company that professes to be working towards “Safe AI”.
Claude does not approve:
💸Hiked Haiku: Anthropic also surprised some with the reason for hiked pricing for the “budget” Claude 3.5 Haiku model:
“During final testing, Haiku surpassed Claude 3 Opus, our previous flagship model, on many benchmarks—at a fraction of the cost. As a result, we've increased pricing for Claude 3.5 Haiku to reflect its increase in intelligence“
Synthesis, not copying In what may turn out to be a landmark precedent case for GenAI and copyright, New York judge Colleen McMahon dismissed a copyright lawsuit filed against OpenAI on the grounds that AI “synthesizes” information, rather than directly copies/plagiarizes. Here’s the key passage in her written judgement:
“I agree with the Defendants. The Plaintiffs allege that ChatGPT has been trained on "a scrape of most of the internet", which includes vast amounts of information from numerous sources on almost any given subject. However, the Plaintiffs have not alleged that the information in their articles is copyrighted, nor could they. When a user inputs a question into ChatGPT, it synthesizes relevant information from its repository into an answer. Given the volume of information in the repository, the likelihood of ChatGPT outputting content directly plagiarized from one of the Plaintiffs' articles seems remote.
Although the Plaintiffs cite third-party statistics showing that an earlier version of ChatGPT generated responses containing substantial amounts of plagiarized content, they have not plausibly alleged a "substantial risk" that the current version of ChatGPT will produce a response plagiarizing one of their articles. Accordingly, Plaintiffs lack Article III standing to seek injunctive relief for their alleged lnjury.“
Will this be the moment that finally shuts down the endless AI copyright legal rabbit hole…?
🏢AI data center news
Infrastructure is destiny in the AI era… so the build-out continues:
Massive US investment firms KKR and Energy Capital Partners agreed to invest US$50Bn in AI energy and data center infrastructure projects.
Google announced it is establishing an AI hub in Saudi Arabia, focusing on Arab language AI models and "Saudi-specific AI applications." (This move comes despite Google's previous commitments to reduce emissions and halt algorithm development for oil and gas production.)
Transformer: Why AI companies are eyeing the Middle East. (Gulf states offer abundant energy, faster construction, easier regulations, and abundant capital.)
🐝Bee-ware: Meta’s plan to build a new *emissions free* nuclear-powered AI data centre in the US were halted in part because a rare species of bee was discovered on the land. (Would never happen in Saudi Arabia.)
🤔Claude critiques its own system prompt
More quality prompting from @lefthanddraft with Claude:
📑Research
Notable AI(-related) research that’s caught my attention this week:
Developing Wise AI Imagining and building wise machines: The centrality of AI metacognition argues that current AI systems lack “wisdom”, defined as the ability to navigate intractable problems using effective task-level and metacognitive strategies. Researchers should focus on this.
AI agents in the workplace a paper in Nature navigating the human response to automated agents (AA) in the workplace, it’s all about trust over efficiency:
“Our findings suggest that lower-efficiency AA might outperform higher-efficiency ones due to the constraining influence of trust on adoption rates. Additionally, we find that lower initial trust in AA could lead to increased usage in certain scenarios and that stronger emotional and social responses to the use of AA may foster greater trust but result in decreased AA utilisation.“
🧪💡SciMuse Interesting Scientific Idea Generation Using Knowledge Graphs and LLMs, a large scale study evaluates the effectiveness of a new AI system developed by researchers Xuemei Gu and Mario Krenn, which uses 58 million research papers and an LLM to generate compelling research ideas across disciplines.
🆕New releases
And the releases keep coming…
Google Trillium — Google Cloud announced that Trillium, the sixth-generation TPU, is now available in preview, offering over 4x improvement in training performance, up to 3x increase in inference throughput and 67% increase in energy efficiency. Unlikely to challenge Nvidia’s dominance, but raises Google’s game.
Google Big Sleep is the first public instance of an AI agent discovering a previously unknown exploitable vulnerability in widely used real-world software: it successfully identified a previously unknown memory issue in SQLite.
🤖Magentic-One is a Microsoft’s new framework for multi-agent AI systems
Amazon X-Ray Recaps generates personalised AI recaps of programmes you’re watching on Prime
Meshcapade text-to-motion tool lets you type a prompt, pick from several motion suggestions, and download to deploy your character in Unreal Engine:
Moondream An open source small language model with vision capabilities:
Those NotebookLM AI-generated podcasts can now be brought to life with AI video here’s an example:
ByteDance X-Portrait-2 A new SOTA model for talking head generation that can transfer fast head movements, minuscule expression changes and strong personal emotion. Expect this to be rolled out to Tiktok users imminently…
Shrinked Transform Spoken Content into Structured Knowledge - the best-looking AI summarisation API I’ve seen to date.
🫣Heygen squared
I’m dying here… following on from his previous post covered last week, Ethan Mollick let two identical Heygen avatars go at it on Zoom… excruciatingly bad dialogue but also darkly funny. (And a reminder: in just a few months you won’t be able to tell which ones — if any — are human… )
💌The ultimate AI girlfriend app…?
Finally, I saw this and felt… I’m not sure what I felt. For sure, it’s gonna be a weirdly different AI future we’re heading into, eh…
🔮[Weak] signals
Onwards. A rich catch this week of non-AI tech signals from near and far futures. In no particular order (…or coherence!)
🗺️Google Maps is nearly 20
Worth noting a date in the diary for February 8th1 next year when Google Maps will turn 20:
Innovation policy analyst Adam Thierer is spot on as to the huge societal value created here:
(And no doubt captured financially by Google as well…)
I once read someone saying, quite profoundly, that there are two generations of map users: those who grew up with paper maps and can locate themselves within them—and those who grew up with digital maps, always positioned at the centre.
📱🚫Australia mulls U16 social media ban
A fascinating development: Australia’s government is planning legislation to enforce a minimum age limit of 16 for social media apps from as early as 2025:
The government cites concerns over the mental and physical health impacts of social media on children as the primary motivation for the ban
The bill would likely place responsibility on the platforms themselves — Facebook (so many under-16s!), Instagram, TikTok, X and possibly even YouTube — to take reasonable steps to prevent access — PM Albanese stated that “the onus won’t be on parents or young people”.
So what do we think about this…?
A benign intervention on behalf of youth by a caring but patrician government?
Revenge legislation because social media platforms won’t pay for news?
…OR: authoritarian censorship of young people’s freedom of communication and expression by an out-of-touch gerontocracy?
(Given the susceptibility to mis- and dis-information ecosystems by people at the other end of life, perhaps they should start by banning social media for over-65s first!!!)
Either way, once again it feels like legacy regulatory systems are so far behind anticipating technology change… by the time this law gets enacted Australian teens may have moved on to spending all their time in relationships with a polycule of AI boy/girl/they-friends and “social media” will be a quaint historical memory… who knows whether the mental and physical health outcomes will be measurably better or worse then.
(How would I regulate them…?)
Fund mass education on limbic-hijacking science for all citizens
Legislate and fund open-source ad-blockers, in-app persuasion detection etc
Legislate algorithmic transparency and fund personal algorithm-defence systems so that individuals can optimise their information consumption to their own goals, not the goals of ad-funded and politically-manipulative social media platforms…
🔐Homomorphic encryption
Apple quietly published more information on its implementation of homomorphic encryption (HE) for end-to-end privacy in their ecosystem, combining HE with machine learning to enable private server lookups while protecting user data:
“One of the key technologies we use to do this is homomorphic encryption (HE), a form of cryptography that enables computation on encrypted data (see Figure 1). HE is designed so that a client device encrypts a query before sending it to a server, and the server operates on the encrypted query and generates an encrypted response, which the client then decrypts. The server does not decrypt the original request or even have access to the decryption key, so HE is designed to keep the client query private throughout the process.“
(Another leading open-source startup in the HE space is Zama)
🧠Wetware computing
Shifting gears…startup FinalSpark is providing researchers 24/7 remote access to actual brain organoid-based processors — living clusters of human brain cells connected to a computer chip… for just US$500/month.
Although right now this is just an experimental setup to effectively enable crowdsourced research, “wetware computing” promises massive energy efficiencies over current AI processors — the human brain typically uses only 0.3 KWh per day… compared to the 1.3GWh it took to fine-tune GPT-3!
Here’s a live view of the organoid processor… and its readouts:
Andrew Maynard discusses the ramifications of where this research may go in more detail:
Meanwhile… new research has found that it’s not just brain cells that have memory: a new paper in Nature Communications claims to show that all cells—even kidney cells—can count, detect patterns, store memories, and do so similarly to neurons:
☀️🌍One panel per person
Energy mission: get everyone on the planet above the 1 solar panel energy breadline within 10 years.
🎭Faceless, but not gaitless
Not sure where this video is sourced from originally (may be a spoof), but in an attempt to foil facial recognition algorithms, some people took to wearing white facemasks outdoors:
Indeed, you can actually go to quite some artistic lengths to shield your identity:
Unfortunately, all sorts of other biometric identification markers could still mark you out… for example gait detection algorithms test up to 96% accuracy in controlled conditions. (Although Monty Python anticipated this over 60 years ago…)
💨Free floating skyscraper wind turbines
Love the pictures in this Freethink article by Emma C. Edwards, Six innovative ways to float skyscraper-sized wind turbines (the 6th is hybrid!):
👁️Stem cell treatment for vision
A groundbreaking stem-cell treatment has successfully restored vision in people with damaged corneas: four individuals with limbal stem-cell deficiency (LSCD) received transplants made from reprogrammed induced pluripotent stem (iPS) cells and three of the four recipients experienced substantial, lasting improvements in vision for over a year with no severe side effects observed two years later.
💃Kinetic dancefloors
For all my friends heading to see Coldplay at Eden Park in Tāmaki Makaurau this week… check out the “sustainability features”2 including energy-generating kinetic dance floors:
🧠Mind expanding
Tickling the neurons this week…
📜Post-Labour Economics Manifesto
Dave Shapiro has gone off in a different direction to most AI rainbow chasers… what should a future economic framework to replace neoliberalism look like when AI outcompetes most labour? Challenging… at least someone is thinking about this…
Those who control the fantasy…
Never underestimate the power of science fiction storytelling to envision better futures… this post from protopian futurist Monika Bielskyte struck a chord:
“I know for many of us it’s hard to not get discouraged to the core now… Dystopias used to be thought of as cautionary tales, yet they became product roadmaps, & now it’s just our reality…Remember, it all starts with a vision. Unless we can truly see it, we won’t be able to do it. Those who control the fantasy, control the future. It’s time to wrestle that control away from those that have only been able to offer us accelerationist fantasies of doom.”
Discrete spacetime
Not wishing to send you down too many speculative physics rabbit holes…. but eruditely intelligent British physicist Jonathan Gorard (rapidly outgrowing his Stephen Wolfram-protégé background…) tweeted a thread claiming initial experimental findings (pending publication later this year) that Spacetime discreteness is observationally detectable in quasar luminosities and other phenomena…. huge if true.
⏳Zeitgeist
Once around the non-tech world not knowing where to look...
📈Bitcoin high
Worth noting the moment: Bitcoin and other crypto broke through to hit record highs, pumped by enthusiasm for the Trump presidency (go figure…) Bitcoin almost breaking US$90K yesterday:
🌡️COP29 sting
After the UN’s Biodiversity COP16 conference in Colombia fizzled out with no substantial resolutions… this week it’s the turn of the UN COP29 Climate Change Summit, kicking off in Baku, Azerbaijan.
Even before the event got started, the Guardian carried out a sting operation on the COP29 CEO, Elnur Soltanov, catching him on camera apparently agreeing to facilitate fossil fuel deals at the summit, highlighting broad-based concerns about the influence of the fossil fuel industry in the climate talks.
While GHG emissions are no longer on track for the worst case scenario (8.5 degrees of warming…!) … who actually believes actions from the UN COPs are effective at bending this trajectory further down… or is it actually technology innovation and market forces doing all the work here *despite* captured governments and survivalist fossil fuel companies? (That’s not to diss the amazing scientific work done to produce the IPCC reports over the years which actually seek to model the scenarios…)
Comparison of global fossil CO2 emissions to the RCP scenarios and older IPCC scenarios. Figure courtesy of Glen Peters. (via Zeke Hausfather) Having just voted in a climate change denier as president, the US currently has a record number of states in drought:
NYT
🦅Migrations in motion
Hypnotic but disturbing video visualisations from the Nature Conservancy, Migrations In Motion models 2,954 species' predicted migration across the Americas to find new habitats in response to climate change. (Pink for mammals, blue for birds, yellow for amphibians.)
(How different will it be for humans…? Invest in land in the Canadian Province of Nunavut now?)
(🎩 Jay Scott for sharing)
🗳️🇺🇸Trumped
The US had an election…which turned out in the end to deliver a decisive result. (All those of us expecting a 50/50 outcome with months of simmering disputes over who won were thankfully proven wrong…
I’ve been trying to ignore the tribally-drenched wall-to-wall mainstream and social media commentary… things I noted:
Turns out all the crypto bros on Polymarket were on the money:
French President Emmanuel Macron was quick to offer his congratulations:
🟣🇺🇸Purple States of America despite the deeply divisive campaign (a derivative of the country’s bizarre, polarizing first-past-the-post electoral college system of democracy), the US political spectrum is far more evenly spread across the country than the hard blue/red maps we saw on election night:
And despite the Republicans winning the Presidency, Senate and The House, outside the headlines were some issue-based state-level votes. There were wins for more progressive-leaning voters here, including several states (Missouri, Arizona, Colorado, Maryland) passing initiatives to enshrine abortion protections in state constitutions, and other states voted on maintaining paid sick leave, wage Increases and public school funding.
On tech: what does it mean for all of the AI regulatory initiatives spun up under Biden? Dean Ball takes a punt at AI Safety under Republican Leadership:
“There is a reasonably high chance that Donald Trump will lead the United States federal government when “AGI” is developed…Many in the AI safety community perceive this as a disaster—they think that Republicans are “anti AI safety,” and that the Trump campaign’s commitment to repeal the Biden Executive Order on AI is a signal of a fully laissez-faire approach to come, even on matters of catastrophic risk. I think this misapprehends both the attitude of the Republican Party and the broader politics of AI safety. Indeed, I think that the GOP and Trump are far better positioned to take major AI risks seriously than the Democrats…“
Ultimately no-one knows yet.
What I have noticed a lot is progressive media coalescing around a new “Oligarchs” narrative… comparing the US moment now as teetering on the edge of a Russia-modelled billionaire oligarchy, with Musk, Zuckerberg, Bezos and the others currying favour with President Elect Trump divvying up the spoils for themselves. Carol Cadwalladr in the Guardian is particularly eloquently acerbic: A new era dawns. America’s tech bros now strut their stuff in the corridors of power:
“We had eight years to hold Silicon Valley to account. And we failed. Utterly.
Because this, now, isn’t politics in any sense we understand it. The young men who came out for Trump were voting for protein powder and deadlifting as much as they were for a 78-year-old convicted felon. They were voting for bitcoin and weighted squats. For YouTube shorts and Twitch streams. For podcast bros and crypto bros and tech bros and the bro of bros: Elon Musk.
Social media is mainstream media now. It’s where the majority of the world gets its news. Though who even cares about news? It’s where the world gets its memes and jokes and consumes its endlessly mutating trends. Forget “internet culture”. The internet is culture. And this is where this election was fought and won … long before a single person cast a ballot.
Steve Bannon was right. Politics is downstream from culture...Elections are downstream from white men talking on platforms that white men built, juiced by invisible algorithms our broligarch overlords control. This is culture now.“
Certainly Musk’s leverage of his debt-laden social platform X to swing young male voters over to Trump’s side seems to have made the impulse purchase worthwhile… and right now the two bros are best buds… even meming DOGE (Department of Government Efficiency) as Musk’s next venture… see how long he lasts.
BUT: remember the first Trump Presidency: US foreign policy was incoherent, chaotic, directionless, impulsive, reactive, attention-seeking… a total and utter mess, ultimately weakening the US’ standing with its allies (and rivals) around the world. And the domestic Covid response was Idiocracy-level disastrous. Right now from the outside the US arguably looks more vulnerable than ever to further political disintegration. And, with the US national debt house of cards perpetually threatening to collapse at the next bank failure …. Trump may well rue the day he won the election. China and Russia must be feeling quite comfortable as they survey their chess boards....
Meanwhile much of the world increasingly runs on the US’ enviable tech sector exports (almost all of which is HQ’ed in California and Washington State a long way from D.C.). Whether the Trump administration could actually wield any meaningful power over these companies as they extend their footprints transnationally…. is probably the bigger picture here.
Just one example: Meta's entire family of apps—including Facebook, Instagram, WhatsApp, and Messenger—the total number of users is around 3.98 billion - US users make up only ~5% of that. If push came to shove, would Meta really remain a US company first and foremost?
The rest of the world can likely afford four more years to wait this one out while the US turns inward on itself… or hopefully pulls through out the other side to a more internationalist stance. It’s a long game.
🧘Memetic savasana
In amongst it all, a bumper crop of memes:
🗳️Inside the swing voter …
The Onion returned on election night… brilliantly executed:
❌Verification test
Remember this one next time some anon volunteers to contribute to your open source project:
😱Worst UI. Ever.
Engineers on Reddit are competing to see who can create the worst Ul. It’s gold. A thread of the best entries, starting with Please enter your phone number:
⌨️Prompt Engineer
We keep being told it’s the skillset of the future…
🌬️Telepresent wind
David Bowen is a studio artist and educator… I came across his 2010 work Telepresent Wind:
“ This installation consists of a series of 126 x/y tilting mechanical devices connected to thin dried plant stalks installed in a gallery and a dried plant stalk connected to an accelerometer installed outdoors. When the wind blows it causes the stalk outside to sway. The accelerometer detects this movement transmitting the motion to the grouping of devices in the gallery. Therefore the stalks in the gallery space move in real-time and in unison based on the movement of the wind outside.“
🚫Access denied
This meme has an internal logic to it…
🍝Will Smith spaghetti benchmark goes up again
“Will Smith eating spaghetti” is the longstanding AI video meme and de facto benchmark. Here is the absolute state of the art:
(Video DARRI3D via @kimmonismus)
That’s it for another week…! Sorry I’m a bit late hitting “send”, have been speaking at a conference in Waihopai this morning and missed my early morning deadline…
🙏🙏🙏 Thanks as always to everyone who takes the time to get in touch with links and feedback.
Namaste
Ben
(Actually my birthday too!)
/gimmicks…
LOVE "synthesized, not plagiarized" ...we are all synthesis organisms!:)