AI 2027🔮 data fracking⛏️ llama 4🦙4️⃣ midjourney v7🖼️✨ buy for me🛒 dreamer💭super-Turing🧮 elephant-skin mycelium tiles🐘🍄🧱 #2025.14
Vibe governing 🤷♂️🏛️
Welcome to this week's Memia scan across AI, emerging tech and (post-)humanity’s exponentially accelerating future.
This week I’m trying to be a *mostly* Tariffs-free zone to get your periscope up in case you’re drowning in the mainstream news flooded zone🌊💩…
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 on emerging epistemic technologies: What Comes After The News?
🔄ICYMI
Substack’s desktop web client now comes with a handy Table of Contents slider on the left - just click on the stacked dashes to the left - here’s last week’s issue for example.
(Nothing on mobile app yet I don’t think…)
⏩AI for Aotearoa - recording
The recording of my recent fireside chat with AUT Vice-Chancellor Damon Salesa is now online to view: discussing contemporary AI opportunities, challenges and complexities for our very small country at the edge of the world…
(Get in touch if you’re planning a conference later this year and looking for an AI keynote…)
Also: I’m gauging interest for a subscriber webinar in May, please can you respond using the poll below if this topic below interests you (maybe take a read of AI 2027 below first if you’re not sure…):
Thank you!🙏
🔮AI 2027
My planned Saturday morning relaxation turned into a compulsive read-through AI 2027, the new mind-bomb speculative scenario drop from the team at the AI Futures Project. Co-authored by Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland and Romeo Dean, AI 2027 charts a fast ASI takeoff between now and October 2027 — at which point they provide a fork in the road, depending on whether the AI Alignment folks have influence at the highest levels… or not:
“We predict that the impact of superhuman AI over the next decade will be enormous, exceeding that of the Industrial Revolution.
We wrote a scenario that represents our best guess about what that might look like.1 It’s informed by trend extrapolations, wargames, expert feedback, experience at OpenAI, and previous forecasting successes.2”
(Lead author Daniel Kokotaljo, ex- of OpenAI, was remarkably prescient in his 2021 scenario timeline: What 2026 Looks Like.)
Like last year’s Situational Awareness essays from Leopold Aschenbrenner (also ex-OpenAI), AI 2027 tries to raise the profile of the potentially imminent spectre of artificial superintelligence among political power centres. However, unlike Aschenbrenner’s conclusions, it’s not totally blue-pilled by US exceptionalism and pushing for a “Manhattan Project for AI”… the endings are both far more nuanced than that.
The core scenario looks out two and half years to the end of 2027… to a future where two AI labs: US-based “OpenBrain” and China-based “DeepCent“ have all-but-captured the global AI market and control of R&D.
By the end of 2025, “OpenBrain” is developing its latest AGI-adjacent agentic model: “Agent-1”:
“…Once the new datacenters are up and running, they’ll be able to train a model with 10^28 FLOP—a thousand times more than GPT-4.15 Other companies pour money into their own giant datacenters, hoping to keep pace.
Although models are improving on a wide range of skills, one stands out: OpenBrain focuses on AIs that can speed up AI research. They want to win the twin arms races against China (whose leading company we’ll call “DeepCent”)16 and their US competitors. The more of their research and development (R&D) cycle they can automate, the faster they can go. So when OpenBrain finishes training Agent-1, a new model under internal development, it’s good at many things but great at helping with AI research
The same training environments that teach Agent-1 to autonomously code and web-browse also make it a good hacker. Moreover, it could offer substantial help to terrorists designing bioweapons, thanks to its PhD-level knowledge of every field and ability to browse the web. OpenBrain reassures the government that the model has been “aligned” so that it will refuse to comply with malicious requests.“
Charting the pace from there:
“Late 2026: AI Takes Some Jobs:
Just as others seemed to be catching up, OpenBrain blows the competition out of the water again by releasing Agent-1-mini—a model 10x cheaper than Agent-1 and more easily fine-tuned for different applications. The mainstream narrative around AI has changed from “maybe the hype will blow over” to “guess this is the next big thing,” but people disagree about how big. Bigger than social media? Bigger than smartphones? Bigger than fire?
AI has started to take jobs, but has also created new ones. The stock market has gone up 30% in 2026, led by OpenBrain, Nvidia, and whichever companies have most successfully integrated AI assistants. The job market for junior software engineers is in turmoil: the AIs can do everything taught by a CS degree, but people who know how to manage and quality-control teams of AIs are making a killing. Business gurus tell job seekers that familiarity with AI is the most important skill to put on a resume. Many people fear that the next wave of AIs will come for their jobs; there is a 10,000 person anti-AI protest in DC.
The Department of Defense (DOD) quietly begins contracting OpenBrain directly for cyber, data analysis, and R&D, but integration is slow due to the bureaucracy and DOD procurement process.41”
Late 2027, AI geopolitics settles in:
“The White House is in a difficult position. They understand the national security implications of AI. But they also understand that it is deeply unpopular with the public.70 They have to continue developing more capable AI, in their eyes, or they will catastrophically lose to China. They placate the public with job training programs and unemployment insurance, and point to the stock market, which is in a historic boom. Then they focus entirely on winning the arms race. They strengthen chip export restrictions, order OpenBrain to further restrict its internet connections, and use extreme measures to secure algorithmic progress, like wiretapping OpenBrain employees—this catches the last remaining Chinese spy. To build goodwill for potential geopolitical conflict, they finally give their Five Eyes allies useful information and limited API access to some siloed copies of Agent-3…
…A much smaller group of officials is asked to draw up a different type of contingency plan: what if an AI goes rogue? This is viewed as an unlikely scenario, but worth spending some time on….
…Finally, diplomats consider what an “AI arms control” treaty might look like. If AI progress threatened to overturn nuclear deterrence, could America and China avoid nuclear war? If someone found evidence of AIs going rogue, could the two countries halt research until they better understood the threat? How could such an agreement be monitored and enforced?….
…On the other side of the Pacific, China comes to many of the same conclusions: the intelligence explosion is underway, and small differences in AI capabilities today mean critical gaps in military capability tomorrow. But China is on the wrong side of this dynamic: they have only 10% of the world’s compute, against OpenBrain’s 20% and all US companies’ combined 70%. Their only advantage is centralization: almost all of their compute is now working for DeepCent, and more than half of it is concentrated at the Tianwan CDZ.73 Only chips at sites without hardened security are left out of the centralized effort as they rightly fear US cyberattacks…”
…More happens in that next month of October 2027 than the previous year…
And then…. you get to choose your own adventure:
The race condition between the US and China becomes far more speculative… ultimately by 2030 the AGIs have gone rogue, having (trivially) deceived their masters taken over and decided… no more humans:
“The new decade dawns with Consensus-1’s robot servitors spreading throughout the solar system. By 2035, trillions of tons of planetary material have been launched into space and turned into rings of satellites orbiting the sun.32 The surface of the Earth has been reshaped into Agent-4’s version of utopia: datacenters, laboratories, particle colliders, and many other wondrous constructions doing enormously successful and impressive research. There are even bioengineered human-like creatures (to humans what corgis are to wolves) sitting in office-like environments all day viewing readouts of what’s going on and excitedly approving of everything, since that satisfies some of Agent-4’s drives.33 Genomes and (when appropriate) brain scans of all animals and plants, including humans, sit in a memory bank somewhere, sole surviving artifacts of an earlier era. It is four light years to Alpha Centauri; twenty-five thousand to the galactic edge, and there are compelling theoretical reasons to expect no aliens for another fifty million light years beyond that. Earth-born civilization has a glorious future ahead of it—but not with us.“
That’s a very dark future vision… and outside the lightcone of possibility of nearly everyone on Earth right now.
(The alternative “aligned ASI” outcome is only slightly more favourable for human agency — although in that one we humans continue to exist… albeit under US-pilled ASI hegemony…)
The whole piece is richly written and strongly conceived. As I was working my way through it I found myself nodding along going “aha” as one by one many of my accelerationist biases were confirmed…Three concepts on the timeline which I particularly grokked:
Neuralese (2026):
“One such breakthrough is augmenting the AI’s text-based scratchpad (chain of thought) with a higher-bandwidth thought process (neuralese recurrence and memory)“
Biology-inspired neuromorphic architectures: (September 2027):
“Agent-3, having excellent knowledge of both the human brain and modern AI algorithms, as well as many thousands of copies doing research, ends up making substantial algorithmic strides, narrowing the gap to an agent that’s only around 4,000x less compute-efficient than the human brain.75“
Insect-sized and bird-sized drones (2029):
“There are swarms of insect-sized drones that can poison human infantry before they are even noticed; flocks of bird-sized drones to hunt the insects;”
Early criticism of the piece focuses mainly on the dominant doomer narrative:
“The researchers aim for predictive accuracy and make a big deal of their credentials in forecasting and research. (Although they obscure the actual research, wrapping this up with lots of very specific narrative.) This creates an intended illusion, especially for the majority of people who haven't thought much about AI, that the near term scenarios are basically inevitable--they claim they are so objective, and good at forecasting!“ — @saffronhuang
Other criticism zooms in on the narrative style - it really reads like a science fiction novella… which isn’t everyone’s cup of tea.
One further take:
Co-author Eli Lifland responds to some early criticism here.
Dwarkesh Patel interviews co-authors Daniel Kokotajlo and Scott Alexander (@SlateStarCodex) in depth… the underlying assumptions and logic in the forecasting model come out pretty credibly IMO:
Just as I was hitting send, Zvi Mowshowitz provided his usual impossibly comprehensive roundup of responses, including a review of the Dwarkesh podcast:
My 2¢:
Yes, this is a super-smart, collaboratively curated (although by a bunch of dudes, natch…), free to use, above all credible scenario tool for input into serious strategic forecasting exercises, at least up to the end of 2027. It’s obviously not a tight prediction of the future, but one of the few serious information assets available right now which deeply asks the question “what would AGI and superintelligence actually look like?”
I would appreciate it if they open-sourced the underlying model assumptions and parameters. In the Dwarkesh podcast they mention that the post was originally a spreadsheet … I like spreadsheets to go with my prose!
Like Aschenbrenner’s take last year, it suffers from “AI is US vs. China and everyone else is an NPC” framing… chronic American exceptionalism is irreparably ingrained into US citizens from birth / naturalisation…
It is patently a product of the San Francisco AI / venture capital bubble. Ultimately the only people with any agency in these forecasts are a few hyper-hyperscale AI companies (US and Chinese), their leadership teams, their AI safety teams (the heroes of our story, hurray!), US and Chinese political and military leaders. Again, everyone else is an NPC. Therein lies the main weak point of these scenarios in my view…
In addition, an assumption of continued magic money / political willpower to fund the AGI “slurp” … and continued subservience of the 6+ billion people outside US/China and their governments to support a global economic system which actively enables this continual concentrating, extractive hegemony.
One alternative timeline would have the US AI venture funding bubble going “pop” any time now in 2025, a major multi-year global recession slowing investment worldwide, the price of GPUs hitting the floor and the a post-bust “Cambrian explosion” of open-source AI labs springing up around a more multipolar world — Silicon Valley loses its attractor function as AI talent can be as well paid in other locations, and legislated open-source becomes the anti-centralisation mechanism to mitigate against singleton AI power concentration.
I would encourage politicians, policy makers, business leaders, technologists, educators, students (but maybe not young children😬) all around the world to read AI 2027 in full and have a think about its implications for your actions today, rather than just blithely putting one foot in front of the other…
🔮More AI forecasting out recently:
The Case for AGI by 2030 80,000 Hours founder Benjamin Todd has put together this primer which looks at what’s driven recent progress, estimates how far those drivers can continue, and explains why they’re likely to continue for at least four more years. (Quoting this research from METR featured in Memia 2025.12 a couple of weeks ago).
An alternative view of Fast Takeoff:
AI automation may hit wealthy cities hardest According to new research from the Brookings Institution AI may disrupt wealthy coastal cities more than rural areas, reversing traditional automation patterns:
White-collar information workers face higher AI exposure than manual labourers in manufacturing.
Machine translation tools could enable developing countries to compete globally in service sectors.
UN warns AI could impact 40% of jobs globally The UNCTAD Technology and Innovation Report 2025, titled Inclusive Artificial Intelligence for Development, focuses on policy making for AI's dual potential to accelerate global progress toward the UN SDGs, or alternatively to deepen inequalities between and within nations:
“AI is the first technology in history that can make decisions and generate ideas on its own. This sets it apart from traditional technologies and challenges the notion of technological neutrality. The rapid development of AI has also outpaced the ability of Governments to respond effectively. The Technology and Innovation Report 2025 aims to guide policymakers through the complex AI landscape and support them in designing science, technology and innovation (STI) policies that foster inclusive and equitable technological progress.
The world already has significant digital divides, and with the rise of AI, these could widen even further…“
A few key data points:
The AI market is going to grow massively in the next decade… but potentially displacing millions of jobs:
AI development and R&D are significantly concentrated in a few powerful companies and developed countries, creating a growing AI divide that makes it difficult for developing nations to catch up:
118 countries, mostly from the Global South, are effectively excluded from major AI governance discussions
This is a narrative around AI which rarely surfaces amidst the Silicon Valley hype fest and US vs China “everyone else is an NPC” dynamic …
⛏️Data fracking
Singapore-based AI analyst Georg Zoeller introduces the concept of "Data Fracking":
The “25% there” comment refers to the amount of traffic large information content websites are starting to see drop off as AI search engines answer questions based on their content. Case in point: World History Encyclopedia:
“There used to be this implicit agreement between publishers and Google that basically, Google could scrape, analyze, process, and do whatever they wanted with publishers’ content and in return, they would send traffic to the publishers, send them readers,” he told me. “Now, this unspoken contract is kind of breaking.“
— Jan van der Crabben, World History Encyclopedia
In parallel, open-source websites including Wikipedia are drowning under the weight of AI internet crawlers trying to download all of the content and updates for AI training, despite plaintively dropping a robots.txt instruction not to. Wikimedia is proposing dedicated APIs to manage AI company access more responsibly...good luck with that?
Zoeller’s call to arms to is to establish "National Data Reserves":
“We need to [create] National Data Reserves to short term protect Digitial Sovereignty and to defray the risk of dependence on fundamental data on US Techbroligarchs.
The survival of the critical places of information is not a given, it’s in fact very much in doubt as the incentives for Big Tech, and the new US government data strategy are pointing to destruction of openly available data to drive the value of US AI
Beyond saving, we need to think about how to incentivize and support the creation, maintenance and curation of high quality, accessible data sources in the age of AI, data sources accessible to talent, to SMEs and civil society. By default, in todays incentive system, they are all doomed to die as AI disintermediates pages and cuts off their direct traffic- and with it ads and ability to appeal for donations.
We could spend a few years to talk about digital sustainability, but given the real world … let’s just skip to mitigation - there is no sustainability with the US in the room, not with these incentives.“
- The Internet Archive (Wayback Machine) is chronically underfunded — and meanwhile content is rapidly being "disappeared" from the internet under the Trump administration cuts... history is being lost / rewritten...
- Also one of the few "free" internet archives archive.is which I've always suspected of being funded by Russia in some way has now started diverting to RT News.... go figure: https://www.semafor.com/article/04/02/2025/paywall-removal-sites-mysteriously-redirect-to-kremlin-controlled-media-outlet
- Right now in my own country, still most of Aotearoa New Zealand's digital (and non-digital) culture is hidden away behind paywalls and/or in restricted digital systems or physical libraries...
Two relatively open digital collections of note:
DigitalNZ is a search site for all things New Zealand, connecting to reliable digital collections from content partners.
The National Library carries out “Whole Of Domain Harvests” for the National Archive:
I would argue that a full digitised "National Data Reserve" for Aotearoa needs to be curated urgently (open-data, open-source) and disseminated via onshore hosting (eg onshore Git, Hugging Face etc) to all schools, educational, government and legal institutions — and anyone else who wants it. (How many petabytes would it be approximately do you reckon?)
In addition: a special copyright exemption should be legislated for training fully open-source models on top of these archives.
(It could even be a core part of the education curriculum: learn to train and deploy your own Aotearoa AI language model on the National Data Reserve).
The risk if this doesn’t happen is that Aotearoa New Zealand’s history and culture becomes subsumed inside whatever dubiously-aligned AI slop takes over the public internet and what the large US / Chinese technology companies are training and optimising their models on...
Related:
Death of the Browser, Rebirth of the Web?
Rachel-Lee Nabors' presentation The Death of the Browser envisions a future where traditional internet browsers become obsolete as AI agents transform how we access information online… and this will disrupt existing ad-based platform monopolies and return control to users, making information freely accessible despite platform monetisation efforts…
Jury’s out on that one for me…
🏭Tech industry news
In a week of economic turmoil, tech giants gonna keep on gianting…
🔬TSMC
TSMC has confirmed that the world's most advanced microchip, based on a 2 nanometer (2nm) manufacturing process is set for mass production in Taiwan from late 2025.
TSMC Fab 20 will be the main 2nm production facility The 2nm chip offers 10-15% faster computing with 20-30% less power consumption.
Despite a US$100Bn investment commitment into onshore US chip manufacturing capability, Taiwan's chip dominance will continue for some time yet (assuming the "silicon shield" against potential Chinese invasion holds…)
🔐Apple
The UK government lost its bid to keep the details of a surveillance order it brought against Apple secret, following a UK court ruling allowing public disclosure of “the bare details” of the (to date, unspeakable) surveillance order against Apple.🍿
🤫Google Deepmind
Google's AI arm DeepMind has implemented stricter vetting processes and more restrictive processes for publishing AI research, in particular implementing a six-month embargo on "strategic" generative AI papers. A fundamental change of culture, prioritising competitive advantage over academic research transparency:
“I cannot imagine us putting out the transformer papers for general use now,” said one current researcher.“
⚔️Microsoft
Microsoft's 50th-anniversary celebration was disrupted when software engineer Ibtihal Aboussad publicly confronted AI CEO Mustafa Suleyman, accusing him of being a "war profiteer" and using AI for "genocide."
After being escorted out, Aboussad sent a company-wide email explaining her protest against Microsoft's US$133 million contract with Israel's Ministry of Defense, which she claims enables mass surveillance and military operations against Palestinians:
“My name is Ibtihal, and for the past 3.5 years, I’ve been a software engineer on Microsoft’s AI Platform org. I spoke up today because after learning that my org was powering the genocide of my people in Palestine, I saw no other moral choice. This is especially true when I’ve witnessed how Microsoft has tried to quell and suppress any dissent from my coworkers who tried to raise this issue. For the past year and a half, our Arab, Palestinian, and Muslim community at Microsoft has been silenced, intimidated, harassed, and doxxed, with impunity from Microsoft. Attempts at speaking up at best fell on deaf ears, and at worst, led to the firing of two employees for simply holding a vigil. There was simply no other way to make our voices heard.
We are witnessing a genocide…”
Courageous, particularly in the prevailing political climate in the US. Needless to say, Aboussad and Vaniya Agrawal, also involved in the protest, no longer work for Microsoft.
Microsoft has refuted rumours circulating on Chinese social media that it would cease operations in China this year, due to geopolitical tensions... (Microsoft clarified that the shutdown involves only its outsourcing joint venture Wicresoft rather than Microsoft China itself).
Microsoft's AI for Science lab director Christopher Bishop believes scientific discovery will be AI's most transformative application in an interview with the FT. Microsoft's AI for Science lab applies deep learning to chemistry, physics, and biology, using AI models to process vast scientific literature and optimise experimental iterations.
🇰🇷Samsung
Samsung turns to China as chip business struggles Samsung is increasingly relying on Chinese technology companies to bolster its struggling semiconductor division, with exports to China jumping 54% between 2023 and 2024 as US export controls reshape AI chip supply chains and international partnerships.
🔓OpenAI
A new study by researchers from the University of Washington, University of Copenhagen, and Stanford provides evidence that OpenAI's models were trained on copyrighted content — and actually memorised books an articles, potentially strengthening lawsuits against the company. The researchers developed a method to detect memorization using "high-surprisal" words
The WSJ’s Keach Hagey in The Secrets and Misdirection Behind Sam Altman's Firing From OpenAI reveals more on the insider dynamics behind Altman's dramatic OpenAI firing and reinstatement in November 2023. (Her analysis: it was Mira….!)
Shopify
Shopify’s CEO sent a memo on AI use to all staff:
"AI usage is now a baseline expectation…
Repeat that last paragraph:
Before asking for more Headcount and resources, teams must demonstrate why they cannot get what they want done using AI. What would this area look like if autonomous AI agents were already part of the team? This question can lead to really fun discussions and projects."
Early signal of what’s coming… the tech’s not ready yet, but start exploring to be ready when it is.
🆕 AI releases
🥇Gemini 2.5 Pro for all
Did I mention Gemini 2.5 Pro beating out allcomers in benchmarking last week? (Yes I did.) Google has since opened their new flagship model up to all Gemini users, including free tier. Google now has the current best price/performance range for high benchmarking models: that’s competitive advantage.
@demishassabis (Missing: Llama 4 Maverick..see below)
🦙4️⃣Llama 4
Perhaps until…Meta launched its new Llama 4 family of models, featuring three new models with native multimodality and enhanced capabilities. (Last Month Mark Zuckerberg claimed that Llama’s models had reached 1 billion downloads. They’re mostly bots, Mark…)
Llama 4 comes in three sizes:
Scout (109B parameters), a small MoE model with 17B active parameters and impressive 10M context length, designed to run on a single H100 GPU;
Maverick (400B parameters), a mid-sized MoE model with 17B active parameters and 1M context length, currently ranking #2 on LLM Arena; and
Behemoth (2T parameters), a massive model with 288B active parameters that Meta claims outperforms Claude 3.7, GPT-4.5, and even Gemini 2.0 Pro on various benchmarks.
Just coming in under Gemini 2.5 Pro on the LMArena LMO score vs. cost graph (BUT… see below on gaming the benchmarks…)
Missing: Gemini Pro 2.5 
Meta 
Meta Perhaps the biggest headline of the Llama 4 release is the claimed 10 Million token context window — which in theory enables unprecedented document analysis capabilities. (In practice… Can LLMs Do Retrieval and Reasoning in 1 Million Context Window?)
Other key features as part of the release:native multimodality supports advanced image understanding and multilingual capabilities across 12 languages.
The Scout model designed to run on a single H100, making deployment more accessible.
And of course, all released under Meta’s not-exactly-open-but-more-open-than-others Llama licence … good enough to run locally if you’ve got the GPUs.
Some controversy chatter dogging the release… Meta got caught gaming AI benchmarks …specifically using an "experimental chat version" of Maverick on LMArena that differs from the publicly available version... (It will settle down, every lab does it…
🎭MoCha
Also from Meta, MoCha (Movie-Grade Talking Character Synthesis), a research project from Meta's GenAI team that generates realistic talking characters using only speech and text inputs. Pretty amazing. Just three examples (sound on):
🖼️✨Midjourney V7 Alpha
After over a year since V6, Midjourney announced the release of V7 Alpha with improvements including:
Model personalisation on by default (I went through a 200-image personalisation quiz…)
Draft Mode:
“Draft mode is half the cost and renders images at 10 times the speed. It’s so fast that we change the prompt bar to a ‘conversational mode’ when you’re using it on web. Tell it to swap out a cat with an owl or make it night time and it will automatically manipulate the prompt and start a new job. Click ‘draft mode’ then the microphone button to enable ‘voice mode’ - where you can think out loud and let the images flow beneath you like liquid dreams.“
Enhanced text and image prompt handling
Higher image quality with better textures
Improved coherence for bodies, hands, and objects...
The MJ team will be shipping new features every week or two so for the next 60 days, including a new V7 character and object reference.
Despite ChatGPT’s highly impressive image generation, I find Midjourney more intuitive for imagining…
🧩Copilot, with memory
Consumer AI head Mustafa Suleyman showed off a bunch of new Microsoft’s Copilot features:
remembering user preferences (like ChatGPT *over a year ago*)
Deep Research (like ChatGPT, Perplexity, Gemini)
take independent actions (also like ChatGPT/Operator)
and also a demo of generating podcasts (like Google NotebookLM).
As usual with Microsoft, lots of “me too” features but never quite as coherently packaged or executed… we’ll see.
Of note: the new Copilot features reduce Microsoft's dependency on OpenAI under the hood. And, according to the writeup by the FT, Microsoft is aiming to recapture consumer market share lost to Apple, Amazon, and Google (not to mention OpenAI). Long way to go there…
🎬Nova Reel
Amazon has upgraded its AI video model, Nova Reel 1.1, with longer and more sophisticated video generation, up to two minutes of soda pop adverts…
🔆Quasar Alpha?
A new model Quasar Alpha, which has turned up on OpenRouter….is benchmarking above Claude 3.7 Sonnet … and apparently no-one knows which lab made it:
Intriguing …. and free to use on OpenRouter - just create an API key. I will be doing exactly that today…, save myself a few bucks!
💊TxGemma
Google DeepMind has unveiled TxGemma, a new set of AI models designed to accelerate and improve therapeutic drug development processes:
🔌Zapier MCP server
Zapier has launched its MCP (Machine Conversation Protocol) server, enabling seamless integration between any workflow and Cursor agents. For those of us working in Cursor regularly (over a million now), this is a next level integration:
🛒Buy for me
Amazon's new "Buy for Me" feature, now in Beta in its iOS and Android apps, uses agentic AI to help customers purchase products from external ecommerce websites directly through the Amazon Shopping app... yet another move towards disintermediation and signals a shift towards AI agents handling complex consumer transactions autonomously:
🎓Claude for education
Anthropic started positioning into the tertiary education sector:
“Claude helps universities maintain academic integrity while incorporating AI tools in education, backed by Anthropic's commitment to safety.“
LSE (London School of Economics) is the marquee launch partner. Some of the messaging:
Claude's learning mode guides discovery rather than providing direct answers, preserving academic integrity.
AI tools can save faculty time on course design while administrators gain efficiency in documentation.
Anthropic prioritises student data privacy and doesn't use it to train generative models.
↗️Always go meta
Wise advice from Pliny: always get AI to write your prompts for you…
🥼 AI research
A lot of things to cover this week, lightly edited AI-generated summaries:
🛡️AI safety
The Frontier Model Forum (FMF) has established an information-sharing agreement among all member firms (currently Amazon, Anthropic, Google, Meta, Microsoft, and OpenAI), focused on frontier AI safety and security.
AI firms will share intelligence on vulnerabilities, threats, and dangerous capabilities.
Agreement focuses on preventing AI misuse for CBRN and advanced cyber threats.
Information sharing is restricted to FMF members but may expand over time.
Get Prompting
The Prompt Report: A Systematic Survey of Prompting Techniques presents a comprehensive framework for understanding prompting in Generative AI systems:
Establishes first comprehensive taxonomy of 58 text prompting techniques for GenAI systems and provides standardised vocabulary of 33 terms to address conflicting terminology in prompt engineering.
🎩Mike Hall for recommending:
“This is my reference guide, this PDF contains almost all advanced prompting techniques.“
AI will elevate content quality?
A contrarian take from Noah Brier's in BRXND Dispatch: AI will improve content quality rather than degrade it, incentivising high-quality content as brands compete to influence AI-powered consumer decisions.
Paywalled premium content creates opportunities for strategic, freely accessible information designed for AI consumption.
AI models behave more like "rational consumers" than humans, rewarding clear, structured content over emotional appeals.
Hmmm….
AI with Data ownership
Vana is a decentralised platform that began as an MIT class project, aims to change how AI models are trained by giving users ownership of their data and the resulting AI systems.
Vana enables users to own AI models trained on their data, challenging big tech's data monopoly.
Users can pool data across platforms, creating powerful AI applications impossible under current regulations.
The platform's million-plus users demonstrate demand for democratised AI ownership and development.
….But will it scale?
💭Dreamer
Hafner, D., Pasukonis, J., Ba, J. et al. Mastering diverse control tasks through world models. Nature (2025). https://doi.org/10.1038/s41586-025-08744-2 Google DeepMind researchers have developed an AI algorithm called Dreamer that can teach itself to master Minecraft without prior training, reaching expert level in just nine days.
AI self-improvement without training could revolutionise autonomous robot learning in real-world settings.
Goal-focused learning through virtual world modelling may accelerate AI development across industries.
🦠Bacterial evolution decoded with AI
Researchers have used AI to create a detailed timeline of bacterial evolution on Earth, revealing a new timeline of bacterial evolution:
Adrián A. Davín et al., A geological timescale for bacterial evolution and oxygen adaptation. Science388, eadp1853(2025). DOI:10.1126/science.adp1853 The AI analysis revealed that some bacteria used oxygen 900 million years before Earth's atmosphere contained it, and that Cyanobacteria evolved oxygen use before developing photosynthesis, challenging previous assumptions.
Self-learning artificial neurons
Researchers from the University of Göttingen and Max Planck Institute have developed "infomorphic neurons" that more closely mimic biological neurons by learning independently and self-organizing based on information from their local network environment.
🧮Super-Turing
In a similar vein, Texas A&M University engineers have developed "Super-Turing AI": a new approach that mimics the human brain's neural processes to dramatically reduce energy consumption. The technique uses HfZrO-based synaptic resistor (synstor) circuits that can perform concurrent real-time inference and learning — potentially reducing energy datacentre consumption from gigawatts to watts. Plus, learning-on-the-fly capability eliminates pre-training requirements, enabling faster adaptation to new environments.
The aspiration is high:
“These data centers are consuming power in gigawatts, whereas our brain consumes 20 watts…That’s 1 billion watts compared to just 20.“
— Suin Yi, assistant professor of electrical and computer engineering at Texas A&M’s College of Engineering
⚠️New AI supply-chain vulnerabilities
Pillar Security researchers have uncovered a dangerous new supply chain attack vector called "Rules File Backdoor" affecting GitHub Copilot and Cursor, the leading AI-powered code editors:
Hackers can inject hidden malicious code through AI configuration files, bypassing security reviews.
Attack affects major tools used by 97% of enterprise developers, creating widespread supply chain risk.
AI vendors consider this the user's responsibility, leaving organisations vulnerable without proper safeguards.
AI agents changing frontend development
Alexander T. Williams in The New Stack (TNS) charts how AI agents are evolving from assistants to autonomous teammates in frontend development.
Already goal-oriented AI systems proactively optimise code, performance, and UX without explicit instructions. (I get Cursor to do this all the time now… “suggest improvements to achieve [X]… now make them…”)
Future agents will run A/B tests and recommend optimizations based on real user data.
🔍AI Interpretability
Formation of Representations in Neural Networks: research from MIT CSAIL and others, introducing the Canonical Representation Hypothesis (CRH), a unifying theory for understanding how neural networks form internal representations during training. More progress towards peering inside the black box…
🔮[Weak] signals
What’s happening in tech that isn’t AI… the usual race around a dozen different dimensions of emerging technology…
🔌Consumer Tech
GPMI
China unveils GPMI: new interface rivals HDMI with 192Gbps bandwidth
An alliance of Chinese companies has unveiled a new audio and video interface standard called General Purpose Media Interface (GPMI) that significantly outperforms existing standards. GPMI will come with two connectors:
Type C, which is compatible with existing USB Type-C ports, offering 96Gbps bandwidth and 240W of power transfer, and
Type-B, which uses a proprietary connector and packs the full 192 Gbps/480W.
480W!!! That will power a few GPUs… China leading the way here now.
🛡️Online safety
ROOST (Robust Online Open Safety Tools) launched at the Paris AI Action Summit with $27 million in funding from Google, OpenAI, Discord, and Roblox. The open-source tools aim to democratise trust and safety capabilities for smaller platforms with limited resources, and establish interoperable, ecosystem-wide standards… despite notable large platform holdouts.
“What sets ROOST apart is its emphasis on open source at a time when trust and safety efforts are often fragmented and reinvented from one platform to the next. Models like Roblox’s voice safety classifier and Google Jigsaw’s Altitude tool hint at how shared code can enable more efficient, agile solutions, though they also illustrate the need for continuous oversight to avoid spreading outdated or vulnerable software.“
🚀Space
Kuiper starts launching
Amazon (First covered in Memia 2022.43), this week Amazon begins launching 27 satellites into Low Earth Orbit, marking the first major deployment of its Project Kuiper initiative aimed at providing global internet access. Kuiper’s 3,200+ satellite constellation will create direct competition to SpaceX Starlink in global internet access.
The first batch are being sent into space aboard a United Launch Alliance (ULA) Atlas V rocket from Florida's Cape Canaveral.Artemis II stacked
NASA NASA has fully stacked the Space Launch System (SLS) rocket for the Artemis II mission. At a whopping 65 metres tall and weighing in at 99 tonnes, this beast is prepped to carry the first astronauts in NASA's ambitious Artemis program. The core stage forms the backbone of this monster, supporting everything from the launch vehicle stage adapter to the Orion spacecraft that'll eventually ferry humans back to the lunar surface…. assuming that Elon doesn’t somehow get NASA’s funding pulled.
SpaceX tourism
SpaceX Four space tourists successfully completed a 3.5-day private orbital journey, splashing down in the Pacific Ocean on Friday, becoming the first humans to circle Earth above the poles. Bitcoin investor Chun Wang chartered a private SpaceX rocket for himself and three colleagues… but declined to say how much he paid…
⚛️Quantum
Photonic AI chips
German quantum computing firm Q.ANT announced that it has begun pilot production of the world's first light-powered neural processing units (NPUs), photonic AI chips which could reduce datacentre energy consumption by 30 times. Q.ANT claims that its photonic NPUs can process data 50 times faster than silicon-based chips. The press release emphasises “Chip Sovereignty” for Europe…
Images: Q.ANT
All-to-all
MIT researchers have developed a new device enabling "all-to-all" communication between quantum processors, reducing error rates significantly and marking a significant advancement toward scalable quantum computing...
"All-to-all" communication replaces limited "point-to-point" architecture for quantum networks.
Scalable network design could accelerate development of practical quantum supercomputers.
🎮Drones
Gun drones are here A roundup of the latest footage of drones equipped with ballistic weapons being tested and deployed in the Russia-Ukraine battlefield:
Flying Sun
Freefly Systems has unveiled the Flying Sun 1000, a mobile drone lighting system that transforms darkness into daylight using 288 LEDs mounted on their Alta X heavy-lift quadcopter, providing 300,000 lumens of portable lighting for emergency response and search/rescue operations.
X-Fly The latest demo video of X-Fly’s Bionic bird “sensor assisted ornithopter drone”. Amazing:
⚡Energy
21.5 MW wind turbine
Siemens Ganesa Siemens Gamesa has set a new world record with its SG DD-276 wind turbine installed at Denmark's Østerild test field, featuring a massive 276 meters rotor diameter and 21.5 MW capacity—enough to power 70,000 Danish homes annually and eliminate approximately 55,454 tons of CO2 emissions over its lifetime.
EU invested €30 million in this technology, demonstrating significant government commitment to wind innovation.
Ongoing US-EU-China tariff tensions may disrupt rare earth supply chains needed for future turbines.
🦾Robotics
Kawasaki Corleo Easily the most batshit crazy transportation device concept invented so far… completely CGI generated for now. Will be “hydrogen powered” apparently… people will want these…
AgiBot Physical Intelligence unveiled new experiments with AgiBot showcasing a versatile Vision-Language Agent (VLA) capable of performing multiple tasks across different robot embodiments, including human-hand grippers:
Figure 02 deployed Figure’s latest robots are “permanently deployed” in BMW’s SpartanBurg plant:
Pudu wheeled centaur robot
Pudu Robotics unveiled the FlashBot Arm, a "semi-humanoid" robot designed for practical commercial applications, with a “centaur” wheeled base and humanoid upper torso.
FlashBot Arm can deliver items, press buttons, and access restricted areas autonomously.
Applications span hotels, offices, restaurants, and healthcare with AI conversation capabilities.
(We really are entering the world of Star Wars that George Lucas envisioned 50 years ago…)
🖨️3D printing
Form 4L OK, so 3D printing is *finally* ramping up to commercial manufacturing reality. Last week I showcased the Bambu Labs H2D - this week an in-depth demo video of the Formlabs Form 4L itemises how the device is now competitive with classic injection moulding for runs of over 1000 parts… Time to start working on the economics here.
If a small town had a shipping container full of these, with a robot running things inside, how would that reduce offshore manufacturing and shipping from China? This is the only alternative to Temu-fication of stuff…
✈️Transport
🚗💥🔍 Modern traffic accident forensics Three people died in a high-speed collision involving a Xiaomi SU7 electric vehicle in Anhui Province, China with the vehicle catching fire after colliding with a central barrier. Initial reports on Chinese social media included allegations that the doors could not be unlocked, preventing escape.
Xiaomi CEO Lei Jun issued a public statement addressing the accident, promising full cooperation with the investigation … and the company promptly published details of how the doors can be opened manually, together with detailed telemetry logs from the accident (NOA=Navigation On Autopilot):
10:27:17 PM: NOA activated, vehicle speed at 116 km/h
10:28:17 PM: Mild distraction warning issued
10:36:48 PM: NOA issued hands-off warning: “Please hold the steering wheel”
10:44:24 PM: NOA issued risk warning: “Please note obstacle ahead,” with deceleration request
10:44:25 PM: Driver took over from NOA, entering manual driving mode, steering wheel turned 22.0625 degrees left, brake pedal pressed 31%
10:44:26 PM: Steering wheel turned 1.0625 degrees right, brake pedal pressed 38%
Between 10:44:26–28: Vehicle collided with concrete barrier
10:44:28 PM: Vehicle eCall triggered
10:44:39 PM: eCall connected, accident confirmed, police and emergency services contacted
10:45:06 PM: Contact established with vehicle owner, confirming non-owner was driving
10:47:15 PM: Emergency medical services successfully dispatched
Approximately 11:00 PM: Emergency services arrived at the scene
That was a 4 second warning from the autonomous driving system for the human to take back control… from now on, if an accident is due to driver error, the [semi-]autonomous car company will be able to prove it.
Barrel-rotor flying car?
Cyclotech Cyclotech's "Blackbird" flying car prototype has successfully begun flight testing, marking a significant milestone as the world's first aircraft to fly with six barrel-shaped Cyclorotors for propulsion. (Cyclorotor technology offers faster thrust vectoring than conventional propellers for improved stability.)
Self-landing small plane
Epic Aircraft Epic Aircraft unveiled the E1000 AX, a high-performance personal and business aircraft with an impressive “Autoland” self-landing capability at the Sun 'n Fun Aerospace Expo in Florida. Emergency automation activates automatically when pilot unresponsiveness is detected.
EHang live in Shanghai Pilotless air taxi approved for commercial flights in China
As trailed in Memia 2025.04, China's civil aviation authority has granted EHang approval to begin commercial operations with its pilotless EH216-S eVTOL air taxi — the first commercial approval for pilotless air taxis worldwide.
🐘🍄🧱Constructiontech
Elephant-skin mycelium tiles
Nanyang Technological University Engineers at Nanyang Technological University, Singapore have developed innovative "fungi tiles" that mimic elephant skin to passively cool buildings in tropical climates.
🪙Crypto
Satoshi Nakamoto turns 50? as Bitcoin becomes US reserve asset
@cointelegraph Satoshi Nakamoto, Bitcoin's pseudonymous creator, reputedly turned 50 on April 5, 2025, based on hints on his birthday left on an internet forum back in 2010.
The “birthday” coincided with Bitcoin's unprecedented rise to becoming a US reserve asset.
…Meanwhile BTC continues to take a bath along with the rest of the global markets this week - down from US$109,000 at its peak in January:
Coin360.com Nakamoto’s dormant 1.1 million BTC wallet just sits there on the blockchain…
Math reveals coordinated attack behind US$3.5B cryptocurrency collapse
Mathematicians from Queen Mary University of London have revealed evidence of a coordinated attack behind the $3.5 billion collapse of TerraUSD and LUNA cryptocurrencies in May 2022:
New temporal multilayer graph analysis can detect market manipulation in cryptocurrencies.
Tools developed could help regulators prevent future financial system collapses.
💉Health
Shingles vaccine reduces dementia risk by 20%, study finds
A new Stanford Medicine study reveals that receiving the shingles vaccine reduces dementia risk by 20%. Time to get jabbed…
⏳ Zeitgeist
Around the world, trying to avoid the world “Tariff”…
🚢💨UN meeting to decide fate of global shipping emissions
The International Maritime Organization (IMO) is holding a crucial meeting in London to determine how to reduce global shipping emissions, which currently contribute over 200 million tonnes of fossil fuel consumption annually.
The IMO’s 2023 climate strategy requires 30% emissions cuts by 2030… however global demand for shipping is expected to grow by around 60% in that same time period. So even with a 40% efficiency boost, total emissions from shipping could stay the same—or even go up—because so much more cargo will be moved. Jevon’s paradox…
The IMO meeting will target decisions on efficiency standards and zero-emission fuel requirements — determining if shipping transitions rapidly from fossil fuels or continues high-carbon operations.
will shape global maritime decarbonisation efforts.
❄️📉Arctic ice hits record low at winter peak
The Arctic’s ice extent on March 22, 2025, was the lowest maximum observed in the satellite record. Arctic sea ice reached its lowest winter peak ever recorded on March 22, measuring just 14.33 million square kilometres according to NASA and the US National Snow and Ice Data Center.
🌿Chinese scientists unlock Siberian vegetation prediction models
Principal modes of summer NDVI in eastern Siberia and its influencing factors and climate prediction of NDVI in 2019–2021. Credit: Yuqing Tian Chinese researchers have identified three key patterns of summer vegetation growth in eastern Siberia and developed effective climate prediction models to forecast these variations. Siberia one of the parts of the planet which could be positively affected by climate change…at least from a human habitability perspective.
🐝📉⚠️Honey bee loss
A nationwide survey of US beekeepers has revealed catastrophic honeybee colony losses across the country, with commercial operations reporting an average loss of 62% between June 2024 and February 2025. These unprecedented losses threaten US agriculture, particularly crop pollination for almonds, fruits, vegetables, and other essential food sources. Industry leaders warn the scale is "completely unsustainable" and threatens current food production levels… Yikes.
Government transparency cuts pollution and saves lives
Assignment of treatment and control among experimental sample of cities and treatment effect on PITI scores. Notes: Panel (A) shows the distribution of 25 treatment cities and 25 control cities. Panel (B) shows the average aggregate scores by experimental condition in each year of the original study, with SEs derived from bootstrap sampling within experimental conditions. This figure is reproduced from ref. 25 ©2019, Midwest Political Science Association, which reported the original effects of the treatment on transparency. Credit: Proceedings of the National Academy of Sciences (2025). DOI: 10.1073/pnas.2406761122 A study by UC Santa Barbara researchers demonstrates that government transparency on environmental performance directly leads to significant environmental improvements. Conducting a “field experiment” in Chinese cities (huh?), they found that:
Government transparency reduced air pollution by 8-10%, saving thousands of lives.
Transparent cities saw 37% fewer pollution violations and 90% more regulatory inspections.
The results can be applied globally in other cities
(Who’d have thought it?)
Record high: 91 million Americans unable to afford healthcare
The West Health-Gallup Healthcare Affordability Index reveals a record 35% of Americans (91 million people) cannot afford quality healthcare, with disparities widening significantly across demographic groups.
11% of US adults (29 million people) are "Cost Desperate", unable to access or afford care.
Paywall removal sites redirect users to Russian state media
Archive.is, the enigmatic paywall removal website and de facto internet archive, has begun mysteriously redirecting users to RT (formerly Russia Today), a Kremlin-controlled media outlet. According to Semafor, archive.is still sending users to RT as of Wednesday after at least five days of redirects... I tried it (research purposes, of course…) and it seems to be working fine for me here in Aotearoa.
Timing aligns with US-Russia relations shift under Trump administration.
Illicit sites make attractive targets for testing cyber capabilities without consequences…. always use incognito mode!
🌊💩Flooded zone
And finally we have to go there…
🤷♂️🏛️Vibe governing
Trying to avoid the mainstream news this week which is full of this week’s Trump Tariffs insanity…
The tariffs even targeted uninhabited islands …
“Strange outcomes include Australia’s Norfolk Island, population 2,188, being hit with 58% tariffs, while Antarctica’s Heard Island and McDonald Islands — uninhabited except for penguins — and Jan Mayen, a Norwegian-ruled former Arctic whaling station “possibly featuring more polar bears than people,” both face 10% duties.“ —Semafor
The world spent a while scratching their heads at the left hand column… it definitely *isn’t* “Tariffs Charged To The USA”. Turns out it’s a rounded formula calculating the minimum of 10% OR the goods trade deficit for the US, divided by the total goods imports divided by 2 … the White House finally responded to confirm this: they don’t teach this in Econ 101:
Aha…
Rohit wrote this one up… gold:
“This is ‘Vibe Governing’.
The idea that you could just ask for an answer to any question and take the output and run with it. And, as a side effect, wipe out trillions from the world economy.”
The MSM is full of indignant chatter and noise… some of the more insightful voices that I’ve been listening to:
Singapore PM Lawrence Wong:
Aotearoa New Zealand opposition MP David Parker: talking up bringing Europe into the CPTPP trade agreement alongside the other 12 parties already there (Australia, Brunei, Canada, Chile, Japan, Malaysia, Mexico, New Zealand, Peru, Singapore, United Kingdom, Vietnam). An alternative “rules based” trading system to de facto replace the WTO.
Newly minted Canadian PM Mark Carney went to Europe (not Washington) on his first overseas trip, to bolster trade with the UK and EU.
Semafor says it out loud: An American Brexit?
🇨🇳🔒🔋China's rare earth restrictions now threaten US tech and defence industries In response to the tariffs, China has implemented new export restrictions on seven rare earth elements, including scandium and dysprosium, requiring exporters to obtain licenses and specify how buyers will use these materials:
Clearly threatening critical US tech and defence supply chains.
Price increases of up to 5X expected for materials essential to semiconductors and telecom.
US vulnerability is highlighted with only one domestic mine versus China's 85% global production.
(China playing Go, Trump playing TicTacToe?)
⏱️The clock keeps on ticking…
US President Trump extended TikTok's deadline by another 75 days to comply with the sell-or-ban law amid ongoing sale negotiations:
Algorithm ownership remains the critical sticking point in any potential TikTok sale deal.
Appartently Trump said that his 34% tariff (and counting…) against China will give him “great” negotiating power with Beijing.
Public support for a TikTok ban has dropped significantly, influencing Trump's political calculations.
My bet 75 days from now: another 75 days…
💭Meme stream
A few items of memetic eclectica that caught in my net this week…
🌳Museum of Nature, in Nature
BIG (Bjarke Ingels Group) unveiled plans for the Hungarian Natural History Museum in Debrecen, with a design that seamlessly integrates with the surrounding Great Forest landscape.
🧙Harry Potter, 2025 edition
(I think this meme has been around for a while… first time I saw it)
🎨Studio Ghibli Strikes Back
The Ghibli-fication of everything has faded from last week’s ChatGPT image generation frenzy … be thankful for satire:
🦑Trivial alignment
🙏He always said thank yoo!
🙏🙏🙏 Thanks as always to everyone who takes the time to get in touch with links and feedback. Back in your inbox again this time next week!
Namaste
Ben