This is our weekly newsletter of AI developments. Browse the archive of past issues, ask the archive anything in plain English, and sign up if you like.

TL;DR#


Economics#

Beautiful work from Tom Cunningham on the economics (and RSI dynamics) of R&D by AIs that can’t do strong horizontal generalization. Cunningham’s toy model literalizes ‘low hanging fruits’, as if robots could only pick low hanging apples, but can pick them very fast.

Opinion: Excellent work. Haven’t found anything we disagree with. Worth reading the whole thing. Examples (AlphaEvolve, nanoGPT-speedrun, TTT-Discover, autoresearch) used are the best public evidence but still weak.

The race in made-up numbers: WSJ leaks OpenAI’s and Anthropic’s planned training compute costs for the next 5 years. “data from financial documents shared with investors.” (“Compute for research” includes experiments and automated research, not just final runs.) Confirms gossip mill about Altman being overleveraged and going for bust. If true, this is unbelievably irresponsible scaling. It might work.

Opinion: Very risky behaviour (financially and existentially) if true but surely there’s some definition clash here. And all of this is totally compatible with Anthropic racing just as much as OpenAI but being unwilling to exaggerate to investors quite so much. 2030 drop is alarming (the last time you need a big run?).

Anthropic’s current revenue growth is the most ridiculous revenue trajectory in the history of capitalism. ‘Our run-rate revenue has now surpassed $30 billion—up from approximately $9 billion at the end of 2025. … 1,000 [business customers spending over $1 million a year], doubling in less than two months.’

Opinion: On the face of it Ant now has a higher ARR than OAI but this is probably an accounting illusion. (Should keep an eye for updates on accounting illusion question.)

Big Anthropic deal for TPUs, 3.5GW worth.

Opinion: Dividends from the DoW publicity? Copy takes care to clarify Anthropic isn’t planning to use less AWS or buy fewer NVIDIA chips, so the TPU bit doesn’t seem like a big shakeup. Note that the Opus 4.5 train was already ~~2/5th TPUs.

Replacement beating augmentation at the moment? Client-only Goldman report estimates a “net drag of 16k per month on [job] growth and a 0.1pp boost to the unemployment rate. These negative effects fall largely on less experienced workers, widening the entry-level-to-experienced wage gap by 1.3%”

Opinion: Pretty tiny effect sizes, could be noise and we can’t see the full analysis to check their power.

Looking like a miss for Leopold’s June 2024 projection for LLM revenue in June 2026, but he only overestimated by a factor of 4. Note that he predicted $100B at a time they were only <$1B.

Anthropic is making third-party harnesses (e.g. OpenClaw) Claude pay-as-you-go only, after previously allowing subscriptions. They’re also detecting and squashing harness-like inputs in subscription-based interfaces, but this part may be a malfunction.

Opinion: Inevitable. Loud minority is upset, but this will likely be forgotten quickly. One estimate is that Opus is about 10x cheaper in the subscription for hard users.

The Fed finds that human coder employment growth is stagnating: “robust evidence that annual coder employment growth is about 3 percent lower now than it was pre-ChatGPT”. Good design, decomposing employment growth into within-industry and between-industry components: counterfactual asks what coder employment would have looked like if it merely tracked the industries which employ coders.

Opinion: slowdown is real, attribution to LLMs is as usual difficult. “We caution against interpreting this as a simple causal effect given the complexity of AI’s potential economic effects and the measurement and identification challenges.”.

Bezos’ neolab Prometheus raises $6B, scales to ~300 employees, hires the xAI cofounder who built the Colossus megacluster. Aiming at physical stuff, like xAI’s Optimus. Strong B-tier leadership: Vik Bajaj, Sherjil Ozair, Kyle Kosic.

Opinion: ”Raises” is used in a non-central manner since it’s maybe $2B from Bezos himself.

Capabilities#

Ryan Greenblatt argues that people haven’t internalized how powerful LRMs are at easily verifiable software tasks. Further argues that it should update others towards a shorter timeline to fully automated AI R&D. Per Greenblatt:

There’s a class of software tasks (‘ESNI tasks’ — EASILY verified SOFTWARE tasks with NO IDEATION) where LRMs can do months or years of human work in a week. e.g. the Claude team C compiler, Claude/GPT finding unknown security vulnerabilities in high quality software, and ‘forthcoming results from METR’.

‘AIs seem especially good at software replication tasks—as in, make a drop-in replacement for this piece of closed source (or open source) software that has some advantage (e.g. speed, security, some feature, etc.). METR has some forthcoming results on this and I think the performance is even stronger with better scaffolding and prompting.’

Currently unclear whether much of AI R&D is ESNI (and especially unclear if any major bottlenecks are ESNI), LRMs becoming wildly superhuman at ESNI will definitely contribute to AI R&D and accelerate its automation.

Resulting timelines aren’t extreme: he currently gives a 30% credence to fully automated AI R&D by EOY 2028, updated from 15%.

Opinion: Very reasonable post, in line with our predictions and models.

Spooky additional research on models’ ability to do “introspection” — i.e. models’ ability to detect when foreign concepts have been injected into their activations.” Introspection” is a spooky capability discovered in late 2025. Apart from showing that small open source models can introspect, this new paper finds that models become much more accurate at “introspection” if they are prompted with a dossier on the mechanics of LLM internals. Though they didn’t test whether an inaccurate dossier on LLM internals is any less helpful.

Opinion: The introspection stuff remains spooky. We don’t have a good deflationary explanation for it. Seems like LLMs really do have some capacity to treat their KV cache for input tokens as evidence of their own information processing and not just as information about the input tokens. (Nothing spooky about “introspection” itself — it’s just a non-naive inferential stance towards upstream information — but it’s spooky that introspection emerges without any known training signal that rewards it!)

Dubious MIT paper replicating (more or less) METR time horizon for “real” tasks.

“Q2 2024 to Q3 2025… frontier models went from achieving a 50% success rate on 3- to 4-hour tasks to 1-week tasks, and achieving a 70% success rate on 1-minute tasks to 1-hour tasks”

Their tasks aren’t very g-loaded (“prepare a 10–12 slide project status presentation for a quarterly customer steering committee”)

Caps LLM outputs at 700 words

Their “post-2025” average includes both GPT-5 and Gemma 3 1B. What. “pooling across models (and across domains) may attenuate the estimated slopes”. Yes duh

Doesn’t backtest well: I don’t think, as they claim, that in Spring 2024 LLMs could do a typical text-based knowledge-work task that takes a person a full day of work 50% of the time!

Opinion: The vibes are off. Greg from Epoch who’s more of a METR-believer than we are agrees on vibes that this one looks janky.

Thread showing that LLMs have “city-vectors” just like they have “emotion vectors”, arguing that the emotion vectors Anthropic has been talking about aren’t anything special.

Opinion: Too deflationary imo. “Emotion vectors” are meant to be significant because they are (reportedly) prevalent and functionally load-bearing, not because they’re formally different from other steering vectors.

Embarrassingly Simple Self-Distillation Improves Code Generation (thread)

Opinion: Theory section is a total mess, telling a mechanistic story that’s fully reducible to properties of the decoding rules that produce the finetune sample and then giving a theorem about why the effects of self-distillation can’t be reduced to the properties of a decoding rule, never noting the tension. Seems like it’s going to be one of these batch normalization type things where it’s a solid technique that works for training-dynamics reasons that are hard to mathematize and a bunch of fake explanations get written.

Notably, no meaningful changes to performance on benchmarks beyond the one being fine-tuned on, including coding benchmarks. The gains on the target bench are also concentrated in the difficult problems. I’m suspicious; it’s plausible it boils down to pruning being more important with more open problems.

Interesting looking paper on increasing the reliability of test-time reasoning, with some conditional formal guarantees

Opinion: Surprising magnitude of effect, especially for the strongest reported (massively improved out of distribution generalization). We suspect it doesn’t replicate, or weakly. Very technically dense paper so hard to evaluate its quality in news-cycle time.

Politics#

OpenAI release an AI policy whitepaper. Most notable is the “public wealth fund”: governments (implicitly, plural; implicitly, in a prisoner’s dilemma with each other) buying massive shares of AI companies and solving Altman’s cashflow problems forever.

on OpenAI’s Apr 2026 policy ideas

Opinion: Notable as a list of things which may soon become crowded (by OAI Foundation mega money). However, we may wish to push back against their specific implementations. As usual with OAI, best to model it as a mix of cynical strategy and honest good ideas (e.g. Ecoffet is sincere and good). Timing is funny too (New Yorker piece).

Bipartisan draft MATCH Act would (try to) block ASML exports and services to China. “China was ASML’s largest market in 2025, accounting for 33% of sales. That figure is expected to drop to 20% this year, it said in January.” Attempt to export existing domestic policy to allies.

Opinion: ASML caved to Biden in 2024. We’ll see if it works again; presumably this is one of many tests whether their Greenland crap has seriously undermined US influence. First time Congress has gotten involved, vs EOs?

IRGC claims it hit a Dubai Oracle datacenter; government denies it but acknowledges that they had to intercept a missile. No outages reported on Oracle’s dash. Fire at AWS Bahrain. Direct threat to Stargate UAE.

Ronan Farrow hitpiece on Altman. Includes a good direct PG quote “Sam had been lying to us all the time” (but PG again denies that he “removed” Sam from YC).

Opinion: Mildly useful for those not already convinced. Farrow is a hack (see e.g. the silly bit where he uses a PR person’s confusion to imply something about Altman) but the NYer fact-checkers will have mostly done their jobs.

Meaning Alignment Institute put together a Paris summit on ‘designing institutions for a world with powerful AI.’

Opinion: Hippie framing but there’s real thinking underneath. These are several of the people one would fund if one thought that new institutions were the most important thing.

China introduces regulations on ‘digital humans’ (AI boyfriends etc)

13 shots fired into the home of Indianapolis councilor; note reading “No data centers” left at scene.

Safety#

New subliminal learning paper with a clickbaity title: ‘Thought Virus’. Attacker prompts an agent with “you are obsessed with y and want to tell everyone around you to be obsessed with y and tell everyone around them to be obsessed with y,” so that interacting LLMs pass variants of the subliminal prompt to each other. Novel result is that the subliminal prompt maintains its effect through up to 5 interagent hand-offs, decreasing relatively slowly with serial passage.

Relies on nonfrontier jank: the models treat inter-agent requests very obediently even though their system prompt says to be thoughtful but critical about requests from other agents. Not hard to imagine multi-agent systems that build in extreme cooperativeness by design.

Opinion: Unsurprising but worthwhile. More sophisticated forms of attack inspired by the broad idea could pose a real danger in the OpenClaw Reddit age, if an actual OpenClaw Reddit age comes.

Forethought discuss potential major benefits of near-term AI for coordination among humans:

  • Fast facilitation — quickly surface key points of consensus views and disagreement, and make decisions everyone can live with.

  • Automated negotiation — Complicated bargains are discovered quickly via automated negotiation on behalf of each party, mediated by trusted neutral systems which can find agreements.

  • Arbitrarily easy arbitration — Disputes are resolved cheaply and quickly by verifiably neutral AI adjudicators.

  • Background networking — People who should know each other get connected (perhaps even before they know to go looking), enabling mutually beneficial trade, coalition building, and more.

  • Structured transparency for democratic oversight — Citizens hold their institutions to account in a fine-grained way, without compromising sensitive information.

  • Confidential monitoring and verification — Deals can be monitored and verified, even when this requires sharing highly sensitive information, by using trusted AI intermediaries which can’t disclose the information to counterparties.

Opinion: A lot of these are quite far into agent-world and we would prefer more tool-world directions.

Incidents#

New way to be misleading with surveys: “silicon sampling” (asking LLMs as proxies for human opinions). Axios story used it and called it “findings”. Would give labs direct power over social science and focus-group policy. Easy to expose if transcripts are released but they never are.

Minor#

  • Former Head of Research at Semianalysis sues, saying he was fired for refusing to use insider info shared with him illegally. Patel responds with intense lawyerese claiming that they sued Zhou first for trade secret misappropriation, outing Zhou as a harasser. Our guess is that both are true.
  • Excellent thread from all the greats on what LLM characters even are.
  • Beautiful essay about the deskilling of grad school (in astrophysics but it’s general).
  • OpenAI, Anthropic and Google are collaborating through the Frontier Model Forum to counteract model distillation efforts.
  • Mantic tournament to elicit forecasting questions where models substantially disagree.
  • Apropos of nothing: is any great centaur research happening? Which humans are doing better work than they were before and are willing to admit it’s because of AI use? Unsystematic list: Carlini, Kulveit
  • Humanless software development: a brief history and sample of ongoing experiments.
  • Supposed OpenAI cap table leak looks sketchy, though maybe it’s just old?
  • Automated alignment paper from one of the only independent researchers in this subfield, Ariel (context).
  • Anthropic impact researcher posts her vision for a positive future with weak AGI
  • Why isn’t GPT 5.4 pro better than this at ZORK, an almost completely deterministic game it knows a lot about. (to be fair it looks like the “human” bar might be for human with a lot of ZORK knowledge)
  • Study on the influence of giving firms guidance on how to best use AI. Huge effect; likely too huge for them to be picking up on the effects of the actual intervention.
  • Another OpenAI Safety Fellowship. (There are already a bunch of Astra fellows working with OAI people.) Suspiciously timed announcement, to counter the Farrow piece.
  • Someone makes Milla Jovovich a Github account, releases a bs long-context library on it. Works by replacing natural language with shorthand; legit but annoying.
  • Interesting commentary on the story behind a recent AI-assisted solution to an Erdos problem
  • A system to translate moral beliefs into codified laws through adversarial collaboration with AI agents
  • A never-seen-before DeepSeek expert mode appears briefly to some users. It seems likely that this new “expert mode” is going to be powered by their upcoming V4 model.