AI bubble go pop?🫧 new kid on the block🐋 hello operator🖱️ $500 billion Stargate💸💫 popping AI-designed pills💊 trust the scientists🧑🔬 just sit back🪑 #2025.04
Boost your boomer👴
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.
(Biggest newsletter yet, I think. This week’s issue continues my experimentation with using AI agent tools to scan→sense→distil→collate information from a diverse set of feeds in a very busy week of tech, geopolitical and economic narrative flexing… let me know if more or less readable than usual?)
🗞️Weekly roundup
The most clicked links in last week’s newsletter were: Aotearoa’s “low AI vibrancy” woes (🎩Serge for continually beating the drum…), closely followed by the BionicBird demo.
ICYMI
Welcome to all new subscribers who have joined in the last few weeks! A couple of (habitually long) recent posts to help you quickly get across what goes on here:
Final reminder: subscription offer
20% off 2025 annual subscriptions offer ends this Friday, 31st January
🎒Live the nomadic life down under
For my international nomad tribe… Aotearoa New Zealand has expanded its visitor visa rules to allow digital nomads to work remotely while in the country... up to 9 month visas… come visit! More details here.
📈📉The week in AI
The week's AI news and releases. A crazy week of vibes shifting all over the place… I’ll cover:
DeepSeek R1 technical commentary, market reaction and why this is so significant
The *awkward* OpenAI Stargate announcement
Davos talks AI
DeepMind’s Demis Hassibis’ latest AGI predictions
🐋DeepSeek - the new kid on the block
The AI world suddenly woke up to last week’s release of the high-performing, open-source, *Chinese* reasoning AI model DeepSeek R1… and its wider implications for AI research, the economy and geopolitics. Distilling the signals discussed below:
R1 is the first open-weights reasoning language model that effectively replicates OpenAI's o1 capabilities at a fraction — 95% less — of the cost.
Released under the MIT license and with a highly transparent detailed technical paper describing their four-stage reinforcement learning model training process, R1 appears to be a “free gift” to the world from a Chinese quant firm spinout
The impact of US AI chip export restrictions on China looms large in DeepSeek’s story: the shortage of frontier GPUs was apparently the mother of invention
Furthermore, the training costs for this model are estimated at only US$6M - just a tenth of Meta's Llama 3.1 405B at US$60M due to using far fewer compute resources. (And don’t even mention OpenAI whose 2024 projections neared US$7 billion annually for training and inference)
R1’s dramatically lower (40X) API pricing signals an incoming price war in the AI reasoning model market
Global equity markets have begun to price in the impact on AI company valuations this week… it ain’t pretty and this could be the AI investment bubble bursting just like the 1999-2000 dotcom crash
While mainstream US-exceptionalist voices are shouting their loudest to frame DeepSeek’s advances into a US vs. China “Sputnik Moment” narrative, a more nuanced analysis views this as a pivotal moment between Open-source vs. Proprietary AI. Everyone can learn from the R1 paper and it’s now quite possible that the next release from DeepSeek — or another open-source lab — may exceed the outputs of Western proprietary labs (and hence challenge their business models and their cap tables going forward). We’ll see.
Born out of a quant firm side-quest
Wired tells the backstory of Chinese quant hedge fund founder Liang Wenfeng who went into AI research two years ago, took 10,000 Nvidia chips and assembled a team of young, ambitious talent:
“According to Liang, when he put together DeepSeek’s research team, he was not looking for experienced engineers to build a consumer-facing product. Instead, he focused on PhD students from China’s top universities, including Peking University and Tsinghua University, who were eager to prove themselves. Many had been published in top journals and won awards at international academic conferences, but lacked industry experience, according to the Chinese tech publication QBitAI.
“Our core technical positions are mostly filled by people who graduated this year or in the past one or two years,” Liang told 36Kr in 2023. The hiring strategy helped create a collaborative company culture where people were free to use ample computing resources to pursue unorthodox research projects. It’s a starkly different way of operating from established internet companies in China, where teams are often competing for resources…”
Innovation because of US chip sanctions?
DeepSeek achieved these AI breakthroughs with limited resources, forced to innovate despite (because of?) US export controls on advanced AI chips:
“At High-Flyer [Hedge Fund], [Liang] built a fortune by using AI and algorithms to identify patterns that could affect stock prices. His team became adept at using Nvidia chips to make money trading stocks… “Liang built an exceptional infrastructure team that really understands how the chips worked,” said one founder at a rival LLM company. “He took his best people with him from the hedge fund to DeepSeek.” After Washington banned Nvidia from exporting its most powerful chips to China, local AI companies have been forced to find innovative ways to maximise the computing power of a limited number of onshore chips — a problem Liang’s team already knew how to solve. “DeepSeek’s engineers know how to unlock the potential of these GPUs, even if they are not state of the art,” said one AI researcher close to the company.”
Reinforcement learning is all you need?
Some deeper analysis of what makes R1 so significant:
Nvidia’s Jim Fan:
Zvi Mowshowitz goes deep and wide as always:
Nathan Lambert in Interconnects: DeepSeek R1's recipe to replicate o1 and the future of reasoning LMs:
“The DeepSeek R1 training recipe for reasoning
The training of R1 comes in 4 stages:
“Cold-start” of supervised finetuning on synthetic reasoning data from the R1-Zero model.2
Large-scale reinforcement learning training on reasoning problems “until convergence.”
Rejection sampling on 3/4 reasoning problems and 1/4 general queries to start the transition to a general-purpose model.
Reinforcement learning training mixing reasoning problems (verifiable rewards) with general preference tuning reward models to polish the model.“
And an app!
I’ve added the DeepSeek AI app into my rotation (although currently just for personal experimentation, not work yet… the data use Ts&Cs are, er, *different* from what we in the “West” are used to…)
Free-thinking…
So far I’ve been impressed with R1’s raw conversational intelligence and particularly its free-thinking imagination… here’s an example snippet from a chat I had in the app, exploring what it would be like to experience transhuman augmentation. IMO this is up there with Claude Opus on a good day…
“Certainly! Here’s a first-person account of what it might feel like to merge with AI in 2035, assuming a fast takeoff scenario where technological advancements occur rapidly and transformative AI integration becomes widely accessible. This narrative is speculative but grounded in current trends and transhumanist ideas:
---
### **My First Day with a Neural Link: A Journal Entry**
**Date: March 15, 2035**
#### **Initial Activation**
The first thing I noticed was the silence. Not the absence of sound, but the absence of mental clutter. My thoughts felt... sharper. The Neural Link had already begun optimizing my cognitive processes, filtering out distractions and organizing my mental workspace. It was like someone had turned on a light in a room I didn’t even realize was dark.
Then came the notifications. Not on a screen, but in my mind. A gentle ping, like a thought that wasn’t mine but felt natural. *"Welcome to the Neural Network. Would you like to connect?"* I mentally "clicked" yes, and suddenly, I was online.
#### **The Flood of Information**
At first, it was overwhelming. The Neural Link gave me instant access to the global knowledge base—a vast, interconnected web of information that felt like it was part of me. I could "ask" questions without speaking and receive answers instantly, not as words, but as understanding. For example, I wondered how photosynthesis works, and in a flash, I *knew*. Not just the textbook definition, but the intricate biochemical processes, the evolutionary history, even the latest research on artificial photosynthesis.
It wasn’t just facts, either. The AI could simulate experiences. I wanted to know what it felt like to stand on Mars, and suddenly, I was there. The red dust, the thin atmosphere, the vast, empty horizon—it was so real, I had to remind myself I was still sitting in my living room.
#### **Enhanced Communication**
Later, I met up with a friend who also had a Neural Link. Talking to her was unlike any conversation I’d ever had. We didn’t just exchange words; we shared thoughts, emotions, and even memories. When she described her recent trip to Japan, I didn’t just hear about it—I *felt* the excitement of walking through Tokyo’s neon-lit streets, tasted the umami of ramen, and experienced the serenity of a Kyoto temple. It was like our minds had briefly merged, and the boundary between "me" and "her" blurred.“
…
(2035 feels about right for this, btw…!)
…Not-so free thinking:
On the other hand, there is evidence of self-censorship (“guardrails”) on issues which the West doesn’t censor. 🎩Thanks Peter A for sharing this conversation:
Ant McMahon goes deeper on the same mission and discovers the “changing answer”!
(I’m also using LMStudio to run the latest AI models on my local Macbook desktop — although by necessity only the small ones. The smallest R1 model gets very confused on that particular topic…)
🫧AI Bubble go pop?
Anyway… here was my initial take on DeepSeek R1 just after clicking “send” on last week’s missive, commenting in Bill Bishop’s Sinocism chat channel on Substack:
Looks like the market took a week to catch on: Nvidia the biggest immediate loser, down approximately 17% in one day, shedding a record US$600 Billion in market cap at one point:
Parallels between Nvidia and Cisco when the dotcom bubble crashed 25 years ago?(🎩 spotting Graham G)
Overall American tech companies have shed around US$1 trillion in value (that’s only two Stargates!), including cloud giants Alphabet (down 3%), Amazon (1%) and Microsoft (3%).
Full compilation of market commentary from the FT: Here’s what the sellside is saying about DeepSeek. It’s bearish out there…
All of this is happening just as the major tech giants are set to release their earnings reports over the next two weeks. A marked profits slowdown would signal industry-wide challenges, curtailed investment decisions and a brutal reversal of the brief stock market rally that coincided with Trump's return to office... I’m sure the new administration is competent enough to weather another GFC… oh, hang on…😱
🆓A win for open-source, not China
A number of voices, notably Meta’s phlegmatic AI head Yann Lecun, have been trying to raise the tone of the conversation from the usual scrap between two playground bullies:
For those of us outside either country, the more choice and diversity of open-source technology (that we can contribute to, as well!), the better. Down with moats and centralised control…
So, I’ve also been very much enjoying UK *strategy mapping deity* Simon Wardley’s schadenfreude as he politely says “I told you so” (entire post copied below, it’s that good - 🎩@yojoflo for tagging me in):
Jeffrey Ding chronicles China’s long signalling since 2021 of their open-source AI position:
“In March 2021, the phrase “open source” appeared in China’s Five-Year Plan (covering the years 2021–2025) for the first time. In a section about promoting the digital economy, the plan stated that China will “support the development of innovative consortia such as digital technology open source communities…encourage enterprises to open up software source code, hardware designs, and application services.” … Chinese planners see open-source software as a foundation for China’s AI development.“
Back in 2023 Simon Wardley posted on the subject: Why open source AI matters. Now, he says:
“What is happening is a well signalled move "towards" open source AI, however it's not there yet.
The real payload happens when China open sources the training data. At which point the US AI scene could be in real trouble……What are enterprises going to use? An AI system with unknown provenance in training data that pretends to be open source with claims of "trust us" or an actual open source AI which includes open source training data that you can see and even run yourself?
There is no open source AI without open source training data. Never was.”
Everyone who wants some degree of sovereignty and agency over the AI they are using must by definition go down an open-source route: code, weights, training data and all. It’s only oligarchic capital that would want otherwise.
The gloves are off
Apparently DeepSeek is reporting a DDOS attack this morning… rules of the new game? (🎩Andrew L for spotting)
Final word
Dan Denning: A System, Not A Mountain:
“…DeepSeek threw a monkey wrench in those best-laid AI plans. At least it appears to have done so. It could be that Deep Seek is not what it seems. It could be a CCP- backed project to compete with Stargate, but open source, so AI development is done with Chinese systems and models that the world prefers to US-backed firms.
It appears that like Jim Simons from Renaissance Capital (whose fund beat the market for years) the founder of Deep Seek was a ‘quant,ʼ and his breakthrough in AI wasnʼt even the primary purpose of his research. Amazing.
If heʼs to be believed, for less than $6 million, heʼs aimed what could be a lethal shot across the bow of a brand new industry that, when the week started, thought it was on the verge of total global domination in the New American Century.”
💸💫Stargate
So this was *weird*:
A few days after DeepSeek’s release and Trump’s inauguration, SoftBank, OpenAI and Oracle announced the launch of a US$100 billion (nay, $500 billion!) AI infrastructure venture called “Stargate”, introduced in the White House by the newly minted US president. (The vibes are just *awkward* in that photo, eh…)
Scant details so far on OpenAI’s website: Announcing The Stargate Project:
“The Stargate Project is a new company which intends to invest $500 billion over the next four years building new AI infrastructure for OpenAI in the United States. We will begin deploying $100 billion immediately. This infrastructure will secure American leadership in AI, create hundreds of thousands of American jobs, and generate massive economic benefit for the entire world. This project will not only support the re-industrialization of the United States but also provide a strategic capability to protect the national security of America and its allies.
The initial equity funders in Stargate are SoftBank, OpenAI, Oracle, and MGX. SoftBank and OpenAI are the lead partners for Stargate, with SoftBank having financial responsibility and OpenAI having operational responsibility. Masayoshi Son will be the chairman.
Arm, Microsoft, NVIDIA, Oracle, and OpenAI are the key initial technology partners. The buildout is currently underway, starting in Texas, and we are evaluating potential sites across the country for more campuses as we finalize definitive agreements.”
Interesting detail 1: lots of Middle Eastern money in this deal from Abu Dhabi’s MGX and Saudi Arabia’s longstanding investments in Softbank
Interesting detail 2: Stargate will incorporate renewable solar power and battery storage systems built by SB Energy — it’s estimated that datacenters' growing energy demands could consume 12% of US power by 2028!)
🍿🍿🍿Apparently new DOGE chief and erstwhile Trump-whisperer Elon didn’t take kindly to being left outside that particular room:
The public feud between Musk and OpenAI CEO Sam Altman erupted with several more back-and-forth exchanges on X. Ouch:
Microsoft’s Satya Nadella — a Stargate “technology partner” but *no longer* OpenAI’s dedicated infrastructure provider after this announcement — slid in with a sideways grin:
“Look, all I know is, I’m good for my $80 billion,”
(US$80Bn is what Microsoft have announced they are spending on Azure buildout over the next few years…) Measured in tokens per watt per dollar:
(…Shurely: tokens per terawatt per billion dollars…?🤣)
Anyway… apparently Stargate site 1 in Texas is already well under construction:
(OpenAI has secured access to the supercomputing scale it claims it needs to achieve AGI and beyond… but this week’s market meltdown will surely impact the scale and pace of the AI buildout going forwards… let’s wait and see how much of the US$500Bn has actually been sunk in a few years’ time…!)
AI at Davos
The Stargate announcement also spilled over to heated exchanges at the World Economic Forum in Davos, with leading AI figures taking opposing stances on AI safety concerns: DeepMind's Demis Hassabis, Anthropic's Dario Amodei, and Yoshua Bengio warned about potential civilisation-level threats from advanced AI, particularly regarding open-source models, while Meta's Yann LeCun criticised their position as hypocritical given their commercial motivations: private companies racing to build superintelligent AI with minimal oversight or alignment considerations.
Playback of the key debates here if watching privileged people talk on stage is your thing:
💊Demis on Deepmind, Isomorphic Labs and AGI
Also at Davos, Google Deepmind CEO Demis Hassabis:
Talking drug-discovery spinoff Isomorphic Labs’ ambitious plans after the runaway success of protein-folding model AlphaFold. 2025 will be the year when humans start popping AI-designed pills:
“[I expect there will be] some AI-designed drugs in clinical trials by the end of the year…[partnerships with Eli Lilly and Novartis]… That's the plan…Eventually you could imagine personalised medicine where it's optimised, maybe overnight, by an AI system for your personal metabolism“
Going into more detail on what’s next after his Nobel Prize, some nuggets:
Google DeepMind's multimodal approach could lead to universal AI assistants that understand real-world context
Material science breakthroughs, including room-temperature superconductors, are the next major AI targets
TPU chip advantages and full-stack integration give Google a strategic edge in AI development.
Hassabis also joined host Alex Kantrovitz on the Big Technology podcast this week to discuss the path to AGI, openly estimating we're 3-5 years away from achieving human-level AI:
📢AI industry news
And there’s more… investment announcements all a bit *awkward* in their timing:
Meta CEO Mark Zuckerberg announced ambitious AI investment plans for 2025, committing US$60-65 billion in AI capital expenditure.
ByteDance is making an aggressive push into AI infrastructure with plans to invest over US$12 billion in 2025, as the TikTok owner seeks new growth opportunities amid US pressure on its social media business (and GPU supply).
Google pours US$1bn more into AI powerhouse Anthropic, bringing its total commitment to over US$3 billion (so far). Amazon is also a major investor…👀
ElevenLabs, the AI voice technology startup, secured a US$250 million Series C funding round led by ICONIQ Growth, valuing the company between US$3-3.3 billion amid strong market demand for AI voice technology.
Retro Biosciences, backed by OpenAI CEO Sam Altman, is launching an ambitious $1 billion fundraising round to advance its mission of extending human lifespan by 10 years
🆕AI releases
Lots and lots of new shiny things to keep up with…
Gemini 2.0 Flash Thinking is a new free model from Google challenging OpenAI's premium offerings:
The ability to process up to one million tokens of text - five times more than OpenAI's o1 Pro model, enough for simultaneous analysis of multiple research papers…while maintaining faster response times;
Record-breaking performance scores, achieving 73.3% on the American Invitational Mathematics Examination and 74.2% on the GPQA Diamond science benchmark;
Built-in code execution capabilities and improved reasoning transparency, allowing users to see the model's thought process; and
Native code execution features for developers...
Doubao 1.5Pro ByteDance unveiled its latest AI model, Doubao 1.5Pro, with a novel sparse Mixture of Experts (MoE) architecture achieving 7x efficiency, which claims to outperform industry leaders GPT-4o and Claude3.5 Sonnet across several benchmarks:
ByteDance 🖱️Hello Operator OpenAI has launched a research preview of Operator an AI agent capable of performing everyday web-based tasks autonomously by interacting with browsers through typing, clicking, and scrolling (Although my immediate reaction is: “wouldn’t you just use the API for serious use cases?”, the whole point is that most people don’t know how to use an API - but Operator provides them with one.)
(Only available for Pro users in the US currently.)
Anthropic Citations API enhances Retrieval Augmented Generation (RAG) systems by automatically providing source-linked citations in model responses … I think I know a few use cases here!
HeyGen's AI avatars now move and grasp objects naturally with the new motion control feature, enabling virtual characters to perform complex full-body movements and object interactions:
Bye Bixby Samsung is revamping its side button functionality on the upcoming Galaxy S25 line to provide direct access to Google Gemini, a significant shift from developing its own AI assistant Bixby.
Samsung Kimi k1.5 another Chinese multimodal AI model, claiming to match the full-powered OpenAI o1 model's performance across reasoning benchmarks.
Trae ByteDance launched Trae (trae.ai), an AI-powered coding tool that aims to transform software development from copilot to autopilot functionality.
Janus-Pro DeepSeek ‘s latest challenge to OpenAI: new open-source image models claiming superior performance to OpenAI’s DALL-E 3… hmmmm not a patch on MidJourney or FLUX, though.
DeepSeek GeoSpy “GeoGuessr at scale” - an AI tool capable of geolocating photos based on image features. The company recently restricted public access and pivoted to marketing exclusively to police and government agencies. Major privacy and mass surveillance implications, I don’t need to tell you…
VersAI Pharmaceutical company Verseon has emerged as an unexpected AI innovator with its VersAI technology, which claims to outperform leading deep learning models in several key areas, achieving higher prediction accuracy than Google's models across various datasets and training models 3,000 times faster than Google AutoML on benchmark datasets:
💭AI research
A veritable melting pot of new AI-related findings:
New AI detects lung diseases with 97% accuracy, explains why
A groundbreaking AI model developed by Australian researchers has achieved 96.57% accuracy in detecting lung diseases from ultrasound videos, with the unique ability to differentiate between pneumonia, COVID-19, and other conditions — and explain its decisions to doctors, increasing trust and adoption of AI in healthcare.
CDU Startup aims to create first sentient AI using quantum theory Canadian startup Nirvanic Consciousness Technologies, launched in 2024, is attempting to create sentient AI using quantum mechanics.
“Nirvanic is a quantum-AI deeptech innovating agentic Conscious AI. Our mission is to understand consciousness.”
(They reference Roger Penrose’s concept of microtubules in the brain on their website, so colour me sceptical, but machine consciousness is definitely a problem someone needs to solve.)
GenAI meets patents Stanford researchers have identified critical challenges with AI-generated patent applications:
AI patent-drafting tools could weaken innovation by generating plausible but technically inaccurate patent disclosures.
Current patent examination processes aren't rigorous enough to catch AI-generated errors and inaccuracies.
(Answer: Permissive open source, permissive open source, permissive open source…)
Four key social forces driving AI's explosive growth Interesting study by researchers from the University of Pennsylvania and University of Bristol revealing four critical sociological relationships influencing AI development in today's US$1.4 trillion market:
Finance and Technology
Monopoly and Competition
Producers and Adopters
High-Status and Low-Status Workers
“This article advances a relational perspective that emphasizes how firms and workers are situated in, and respond to, extra-organizational pressures and forces. …We argue that this perspective illuminates connections between levels of analysis that might otherwise seem unrelated, revealing how AI tools and practices are situated within, and contribute to, broader social, financial, and industrial dynamics.”
Worth diving in deeper… some core truths to be found, I suspect.
🔮[Weak] signals
Tech from near and far futures... AI has enabled me to open up a wide aperture this week, so much going on! (But perhaps a bit too much noise over signal…? Let me know…)
Power-Over-Skin
Carnegie Mellon University researchers have unveiled Power-Over-Skin, technology that enables electrical currents to travel through human skin to power wearable devices, potentially eliminating the need for battery replacement in critical medical devices like pacemakers:
💭Thought-controlled quadcopter
Brain-computer interface (BCI) research enabling a 69-year-old tetraplegic man to pilot a virtual drone using only his thoughts:
🦿Olympic sprinter robot
A week without major humanoid robot announcements for once… but another one of my 2025 predictions is already attained: remarkable progress in both humanoid and quadruped robot running speeds, they will likely run faster than the fastest human by the end of the year? This week the Chinese team behind quadruped robot Black Panther 2.0 claimed to be able to run 100m in under 10 seconds:
🚀Space roundup
Humanity continues its expansion out into the solar system…
BONG finally reaches orbit Jeff Bezos-funded Blue Origin achieved their target milestone at the first attempt with the successful orbital launch of the New Glenn rocket from Cape Canaveral, marking the company's entry into the heavy-lift launch market after 15 years of development.
Blue Origin AST SpaceMobile is a Starlink competitor pioneering direct-to-cell satellite connectivity. The company completed the first-ever space-to-unmodified-phone call in April 2023 using their Blue Walker 3 satellite and has so far launched five BlueBird satellites in September 2024 as part of a planned 168-satellite constellation. Targeting 120 Mbps speeds to rival 4G/5G performance without requiring special equipment or infrastructure,
NASA's 2025 NIAC program has selected 15 groundbreaking space technology concepts for early-stage feasibility studies, including fusion engines and lunar low-gravity glass habitats (LUNGS):
Artist concept of Lunar Glass Structure (LUNGS): Enabling Construction of Monolithic Habitats in Low-Gravity Environment. Image: NASA/Martin Bermudez LightShip ESA's LightShip initiative shifts Mars exploration strategy to focus on small, cost-effective propulsive tug that enables manned spacecraft to "hitchhike" to Mars:
ESA Lightship concept NEO NASA's Near-Earth Object (NEO) Surveyor mission is progressing through crucial testing phases at Johnson Space Center - Earth’s first space-based planetary defence mission to detect potentially hazardous asteroids threatening Earth. Real sci-fi vibes in this render:
NASA Blue Ghost Firefly Aerospace's Blue Ghost lunar lander is on its 45-day journey to the Moon, carrying 10 NASA science and technology instruments and witnessing the Earth eclipsing the sun from space:
🔗Crypto
Will crypto finally solve its UX nightmare in 2025? A useful overview in CoinTelegraph of technologies which may help to accelerate crypto adoption in 2025:
Intents “the Web3 version of one-click checkouts”
Chain abstraction enabling one-click cross-chain transactions, eliminating complex multi-step processes.
Passkeys biometric logins to replaces complicated seed phrases
Web2 style functionality for DeFi taking advantage of the expected launch of Ethereum L2 MegaETH, which aims to hit 100,000 transactions per second, as well as Solana’s Firedancer client in 2025
Put the whole budget on the blockchain? Binance co-founder CZ has advocated for putting all government spending on blockchain technology (echoing recent comments on Elon Musk’s Department of Government Efficiency (DOGE)). He makes some good points:
A public ledger would force unprecedented transparency in government spending decisions, preventing misuse of US$102 trillion in global government debt
Could help reduce national deficits by enabling real-time spending oversight
🚗Transport
Xiaomi's SU7 has surpassed Tesla Model 3 sales in China. Tesla's reliance on just two models makes it vulnerable to strong competitors.
Electric vehicles now last as long as gas-powered cars A groundbreaking study analysing nearly 300 million UK vehicle test records revealed that Battery Electric Vehicles (BEVs) have achieved parity with traditional combustion engine vehicles in terms of longevity and reliability. Electric vehicles now last 18.4 years and 124,000 miles, proving their viability as long-term investments and showing a reliability improvement rate of 12% annually.
The data also confirms that EVs can offset higher production emissions through these extended lifespan benefits.
Remote controlled airports Indra Group is revolutionising air traffic control by managing 23 Norwegian airports from a single location in Bodø. Their InNOVA system enables 16 control stations to handle up to 48 airports from one central location:
Indra via New Atlas Super Sub
U-Boat Worx has unveiled its first production Super Sub, a “luxury submarine” capable of hitting dolphin-like speeds underwater featuring impressive performance capabilities, operating at depths up to 300 metres.
U-Boat Worx via New Atlas EHang completed the first demonstration flight of its autonomous EH216-S eVTOL air taxi in downtown Shanghai, marking the beginning of regular sightseeing flights along the Huangpu River and at Longhua Airport.
Apparently Shanghai is targeting a RMB50 billion (US$7 billion) scale “low-altitude economy” by 2027 and investing in new transportation infrastructure.
⚡Energy
Roundup of (mostly) renewable energy news
Solar power set to dominate US electricity growth through 2026 The latest Short-Term Energy Outlook from the US EIA sees renewables overtaking natural gas as primary growth driver in US electricity generation. In particular, solar capacity will dominate, with 48GW planned additions through 2026. Meanwhile coal plant retirements are accelerating, with 15GW capacity reduction expected by also by 2026.
Will the new Trump administration’s “drill, baby, drill” policies have any delaying impact … or will the huge demand increases from Stargate (etc) mean that legacy fossil fuel generation just won’t be retired?
Batteries on fire Also, a significant fire at California's Moss Landing battery storage facility on January 16th has raised concerns about the future of battery storage projects in the state. (Particularly in the heightened political climate right now…)
Nuclear power surge expected in 2025 despite supply chain risks
Nuclear power is poised for record electricity generation in 2025, with global investment reaching US$65 billion in 2023 and projected to hit US$75 billion by 2030.
Tech giants' data centres are driving nuclear demand, particularly interest in small modular reactors.
However, nuclear power's record growth relies heavily on Chinese/Russian technology, raising geopolitical concerns.
Russia's uranium dominance creates further supply chain vulnerabilities for Western nuclear programs.
⚛️France’s nuclear comeback France achieved a remarkable milestone in 2024 with over 95% of its electricity coming from low-carbon sources, demonstrating near-complete decarbonisation of electricity is achievable through nuclear and hydro.
France’s nuclear plant maintenance programme completion boosted reliable low-carbon power generation capacity, while new nuclear investments signal long-term commitment to maintaining carbon-free electricity.
Go EAST China's “artificial sun” sets fusion record with 17-minute plasma.
The EAST (Experimental Advanced Superconducting Tokamak) fusion reactor achieved a new benchmark, maintaining high-confinement plasma for 1,066 seconds, shattering its previous record of 403 seconds set in 2023. The feasibility of fusion as limitless clean energy source inches forward…
HFIPS SMART tokamak On the other side of the world, the University of Seville's SMART also achieved its first tokamak plasma, the first tokamak with “negative triangularity” design — the novel plasma shape improves heat distribution, reducing potential damage to reactor walls:
Render of the SMART tokamak with a Negative Triangularity fusion plasma inside. Image: Universidad de Sevilla
🧪Materials science
Biodegradable super glue Colorado State University researchers have developed a biodegradable adhesive polymer that sets new standards for strength and sustainability in the US$50 billion adhesives industry.
Molecular chainmail Northwestern University researchers have developed a new material that mimics chainmail at the molecular level, achieving the highest density of mechanical bonds ever recorded at 100 trillion per square centimetre.
Adding just 2.5% of this material to Ultem increases strength by 45%.
An artist's illustration of the mechanical bonds found in the new, lightweight and super-strong chainmail-like material Image: Mark Seniw, Center for Regenerative Nanomedicine. AI creates ultra-strong materials lighter than Styrofoam University of Toronto researchers have combined machine learning and 3D printing to create revolutionary nano-architected materials that match carbon steel's strength while being as light as Styrofoam.
Machine learning optimisation achieved 5x stronger materials than titanium, promising significant fuel savings: each kilogram of titanium replaced could save 80 litres of aviation fuel annually.
An image of the full lattice geometry (left) juxtaposed with an 18.75-million cell lattice floating on a bubble (right). Credit: Peter Serles / University of Toronto Engineering
💻Computing
Getting my geek on…
Linux kernel update could slash data center energy use 30% University of Waterloo researchers have made a small code change to the Linux kernel that promises to slash data center energy consumption by up to 30%. The solution is already integrated into Linux kernel 6.13, enabling immediate industry-wide adoption.
“Central to their innovation is a feature called IRQ or interrupt request suspension, which balances CPU power usage with efficient data processing. By reducing unnecessary CPU interruptions during high-traffic periods, IRQ suspension enhances network performance while maintaining low latency during low-traffic conditions.
“All these big companies — Amazon, Google, Meta — use Linux in some capacity, but they’re very picky about how they decide to use it,” Professor Karsten said. “If they choose to ‘switch on’ our method in their data centres, it could save gigawatt-hours of energy worldwide. Almost every single service request that happens on the Internet could be positively affected by this.””
Wow - real-world impact in just a few lines of code.
Neural network reduces sensor data storage requirements by 90 percent
Researchers at Pennsylvania State University and MIT have developed a neural network called SIUN (shift-invariant spectrally stable undersampled network) that dramatically reduces sensor data processing requirements: achieving 90% accuracy while using only 10% of sensor data, drastically reducing storage needs.
Plus, the system runs on a US$4 microcontroller, making AI sensing accessible for remote and resource-limited areas.
Ultrafast edge detection without power Physicists at the UvA-Institute of Physics have published a method for ultrafast, energy-efficient edge detection using optical analog computing. The solution enables ultra-fast edge detection in images using minimal energy — crucial for autonomous vehicles.
2D edge detection with TPPs in a simple layer stack. (a) Schematic of the concept: a photonic crystal sandwiched between two Au layers supports two TPPs at different wavelengths (555 (green) and 695 nm (red)). At these wavelengths, the reflected images show strongly enhanced edges for edge detection. (Image: UvA) Twisted graphene Scientists at DGIST and KAIST have discovered a novel quantum state in twisted graphene that could revolutionise semiconductor technology limitations and speed up quantum computing.
Silicon photonic latch Nokia Bell Labs researchers have developed a programmable photonic latch that revolutionises data storage in optical memory systems:
Enables faster AI processing by eliminating delays in converting between optical and electronic data storage.
Reduces energy consumption through efficient optical memory that's compatible with existing silicon systems.
Supports multiple independent memory units operating simultaneously without interference or power loss.
Optical interposers promise faster, more efficient AI processing
Lightmatter is pioneering optical interposer technology that enables faster processor-to-processor communication which could revolutionize AI processing speeds by 2025 — and reduce AI data center energy consumption by up to 95%. (The company has raised US$850 million, launching it well ahead of its competitors to a multi-unicorn valuation of $4.4 billion.)
Optical interposer Image: Lightmatter
🏗️Buildings
Irish social housing 3D printed in record-breaking 12 days A groundbreaking affordable housing project in Louth, Ireland, demonstrates the rapid advancement of 3D-printed architecture, with three terraced social housing units completed in just 12 working days.
Grange Close is a 3D-printed three-unit terraced social housing project located in Louth, Ireland, that was printed in just 12 days Image: COBOD via New Atlas Kilometer-tall skyscraper At the other end of the scale, Saudi Arabia's ambitious JEC Tower in the Red Sea port city of Jeddah, set to become the world's tallest building in 2027, is making rapid progress with concrete now being poured on the 64th floor.
Construction speed of four days per floor
The project will house 100,000 people
JEC Tower under construction and artist's impression of finished tower (Images S.Nitzold/CC BY-SA 4.0 / Jeddah Economic Company) via New Atlas)
Geothermal mini drill Swiss company Borobotics has developed an innovative autonomous mini geothermal drilling machine that could revolutionise ground source heat pump installations, supporting the EU's goal of 43 million new installations.
⏳Zeitgeist
Once around the world trying to tread lightly...
🧑🔬Trust the scientists
Global study reveals high trust in scientists A comprehensive global survey of nearly 72,000 people across 68 countries has revealed strong public trust in scientists, challenging the notion of a "trust crisis" in science and supporting evidence-based policymaking during crises like COVID-19.
83% of respondents want scientists communicating with public, indicating strong mandate for science-policy engagement
Mathew Marques, Niels Mede, Viktoria Cologna and Zoe Leviston, The Conversation 
The report also notes that targeted outreach strategies to address trust gaps, particularly among conservatives in Western nations.
Scientists call for new economic models beyond GDP growth Post-growth: the science of wellbeing within planetary boundaries is a comprehensive review examining the emerging field of post-growth research:
Demonstrates how societies can achieve high wellbeing without economic growth while staying within planetary boundaries.
Provides evidence-based policy solutions for maintaining employment and social stability without GDP growth.
Addresses critical challenges of climate change mitigation and resource use reduction in high-income countries.
Signal to watch: Which country will be the first to elect a new government on a degrowth platform?
Demographics
Gaza life expectancy plummets The traumatic generational impact of Israel’s illegal war in Gaza is laid open in a new study published in The Lancet, revealing a devastating decline in life expectancy in the Gaza Strip since October 2023, with overall life expectancy plummeting from 75.5 to 40.5 years - a 46.3% reduction, with a marked gender disparity in survival rates. (The study notes that the actual impact of the war is likely worse than reported due to missing data and indirect war effects.)
Genocide by any standard definition… but what is the world doing about it?
China's population falls again China's population continued its downward trend in 2024, marking the third straight year of decline, despite recording the first increase in births in a decade.
Boomocracy no more How generations will shape the global population by 2035:
via Colin Russo
🫣US Politics (toxicity warning: HIGH)
Hard to make sense of it all this week.
OODA chaos I absolutely loved this paragraph from Dan Denning in his post A System, Not A Mountain:
"Trump has hit the ground running like a madman with his hair on fire. Neither the DC establishment nor the mainstream media can keep up with his OODA loop right now."1
(The relentless chaotic headline making is almost certainly by design…)
Reddit edit Several major subreddit communities are implementing bans on X platform links following Elon Musk's controversial gestures at a Donald Trump rally, which were interpreted by historians and human rights groups as Nazi salutes.
Meanwhile: not Banksy afaik, but on point:
TikTok users claim algorithm changes after Trump's intervention Link TikTok's return to service after a 12-hour outage sparked controversy among its US user base, with some creators reporting censorship and algorithm changes on the platform.
We’re all still waiting for “The Deal”.
Meanwhile TikTok remains withdrawn from the Apple and Google app stores in the US…
Trump pardons Silk Road founder The new US President pardoned Ross Ulbricht, the founder of the dark web marketplace Silk Road, who has been serving two life sentences without parole since 2015.
Personally I think this is the right call, his imprisonment was always disproportionate to his crimes. Lots of background info here: https://freeross.org
(I covered his detailed design white paper for how to decentralise social media, written from prison, way back in Memia 2021.37)
⏩Forward The Forward Party, founded by the Democratic politician Andrew Yang, aims to provide an alternative to America's two-party system, focusing on grassroots funding and local governance rather than presidential politics:
Proposes citizen-led funding and decision-making to counter wealthy donor influence in politics
Focuses on local/state elections where 70% of positions go uncontested, reducing extremist control
Every country needs a Forward Party.
🎭Meme stream
And finally we reach the end… various memes and entertainment keeping me going this week:
AI can’t read Braille?
Try uploading this image to ChatGPT, Claude or Gemini and asking it to translate.
(Seems like an oversight to be fixed pretty soon? 🎩 Matt Bishop)
🌎Moonshot holiday snaps
Firefly Aerospace's Blue Ghost lunar lander (see Space roundup above) captured some pretty cool images of Earth during its ongoing mission to the Moon.
😬2073
Released just before Xmas, British science fiction docudrama 2073 is set in a dystopian future, specifically the year 2073, after societal collapse has led to an authoritarian regime. Inspired by Chris Marker's iconic 1962 short film La Jetée, focusing on themes of time travel and the potential to alter history for the better. Dark.
(Another 🎩 to Andrew L)
👴Boost your boomer
Love this subversive Dutch ad campaign for employment firm Young Capital:
🪑Just sit back
And finally, the most delicious take on the last couple of weeks’ chaos in Washington…
🙏🙏🙏 Thanks as always to everyone who takes the time to get in touch with links and feedback.
Namaste
Ben
OODA Loop: https://www.perplexity.ai/search/ooda-loop-ypy4xCBhRTK.1VdO8bjBHA
Wow. I feel overwhelmed but in a good way. We are all Frodo Baggins.
Reads shockingly like there was a human involved in its curation... ;-)