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#

High-level debate about the correct way to model AI “uplift” (productivity gains). METR differentiate between 3 measures: 1) uplift on the tasks you were doing before AI came along, 2) uplift on new tasks you do with AI, and an integration of these, 3) uplift in the value of the total amount of work you get done. And: Uplift on old tasks ≤ Uplift in value ≤ Uplift on new tasks. Naive uplift metrics just measure uplift on new tasks, thus exaggerating current AI value.

Ord pushback: they assume utility is linear in output (e.g. assuming twice as much of each task is twice as good), when there are diminishing returns.

Opinion: Ord here is just pointing out that bottlenecks (things which are less sped up by AI, here) get more important and more valuable per instance under most plausible models. Similar to O-ring and Baumol cost disease type intuitions - the things which don’t get sped up as much become relatively more valuable on the margin.

Nice review of the vast disagreement among economists of AI. Three cruxes: “Which Jobs Can AI Actually Automate?; Can the Economy Absorb What AI Companies Produce?; RSI.” On that note, a paper estimating the second question.

Opinion: Going to be a while until we can really give definitive answers on any of these, though the review does miss a generally important consideration: to what extent are there new tasks that come about due to the adoption/capabilities of AI? This consideration pushes against job displacement but in favour of extra economic growth.

The compute shortage keeps unfolding: OpenAI closes their fine-tuning API to new customers. Analysis

Opinion: Plausibly too small a revenue stream for OpenAI to focus their engineering efforts on maintaining. Most finetuning jobs were of small models and for non-reasoning/agentic tasks, and open models are more than sufficiently capable for most use cases here.

McK, BCG, Bain all invest in OpenAI’s AI-integration consultancy ‘OpenAI Deployment Company.’ Commentators see this as a hedge against OpenAI eating their lunch. The OAI Deployment Company has raised $4 billion of initial investment.

Opinion: Presumably it’s a hedge plus expectation/hope of guiding synergistic rather pure zero sum interactions between their own business models and OpenAI’s AI-integration consultancy business. ‘Legacy’ businesses having more of a seat at the AI-steering table possibly good for toolworld?

CME Group announce compute futures markets are coming this year.

Opinion: A very useful hedging instrument for compute intensive projects, as well as a useful metric (for us) to track. Note also that we will get a futures curve - there will be every 3 month futures going out at least 2 years if this resembles typical commodities markets, allowing us to see how the market anticipates demand changing.

Why did Cloudflare lay 20% of its staff off? Well, let’s reach for our trusty null hypotheses (they’re bullshitting about AI and really just financially struggling or getting lean for IPO). they did the counterintuitively toxic combination of beating on revenue this quarter (640m vs 622 est) and revising next quarter down slightly (664 vs 665 est).

Opinion: Most likely a response to their projected growth rate plunging. For a pre-profitability company being valued on a multiple of revenue, growth is all that matters.

Anthropic’s innovations in conditional stock trading:

  • Anthropic stock can’t be transferred without Board approval — unapproved sales are void, and the “buyer” gets no stockholder rights.

  • SPVs are not permitted to hold Anthropic stock; any transfer to an SPV is void.

  • Funds offering “indirect access” (forward contracts, tokenized shares, etc.) are likely fraud or selling something worthless.

Opinion: Companies frequently make claims like these in relation to pre-IPO trading of stock - very rare for this to ultimately bite in practice. Most likely they are bluffing and none of this will bite post-IPO, though it may slow the market in secondaries.

Jasmine Sun spreads the dreadful zero-sum “permanent underclass” meme in the NYT.

Opinion: Making people rightly afraid for their jobs is one thing. Drawing attention to a messianic / satanic pattern in the tech industry is one thing. “permanent underclass” isn’t an infohazard, it’s a self-fulfilling invitation to negative-sum action. Still, perhaps Sun will help create a backlash.

Forecasting the progress in Chinese lithography. Experts performed slightly better due to heavier weighting of short time horizons.

Opinion: Still looks like they are minimum 10+ years behind ASML here, and it’s unlikely to contribute much to closing the compute gap in the near term.

A naive BOTEC implies that the OAI Foundation should spend $10B next year.

Opinion: The math checks out in terms of “they should give this much given their timelines and endowment” but it seems very implausible they’ll do anything like this next year with the limited staff and experience they have, plus most of the endowment being tied up in illiquid OpenAI stock.

Jack Clark’s new internal research unit announces their agenda.

Four research focus areas:

  • Economic diffusion — who adopts AI, productivity/growth effects, labor market impacts, future of jobs and expertise pipelines

  • Threats and resilience — dual-use risks, offense-defense balance, crisis infrastructure, surveillance, defensive mechanisms

  • AI systems in the wild — group epistemology, critical thinking, agent governance, model values, human-AI team dynamics

  • AI-driven R&D — governance of self-improving AI, “fire drills” for intelligence explosions, telemetry, jagged capability frontiers

Outputs planned: more granular Economic Index data, resilience research, transparency on internal AI-driven acceleration at Anthropic

Opinion: New(ish) team members Anton Korinek is fantastic. Past Anthropic work on economic diffusion in particular has felt unserious or PR-slop-like in the past, would be great if that changed!

Capabilities#

Here we go! “Interaction models”: Thinking Machines finally demo a near-real-time duplex speech model, with the idea that this is necessary to keep latency down for the AI collaborating with humans. Grandiosely: the models can now understand time. Analysis.

Opinion: Feels a bit slow from the demo, but the ‘can listen while speaking, speak while listening, interrupt, and count time’ features are exciting from a product pov. Of interest mainly because Thinking Machines are (or claim to be) ideologically toolwordly.

Marin (Stanford open model group) releases updated semi-automated ‘scaling recipes’ for efficiently and predictably scaling models from 0.1 to ~100B.

The algo takes hyperparameter settings for a given small training run and predicts optimal-ish hyperparameter settings (and the performance of the resulting systems) for scaled systems. This kind of “compute parametrization of hyperparameters” was responsible for some really important scaling breakthroughs; e.g. muP is rumoured to have been the secret ingredient responsible for OpenAI’s ability to scale pre-training to GPT-4 back in 2022. Their hyperparameter predictor is very good, which can save newbies millions. A valuable epistemic side effect is updated scaling laws with 2026 techniques.

Opinion: Not really a methodological advance; rather, it’s empirical confirmation that some previously published techniques (e.g. Apple’s Complete(d)p, an improvement of muP) aiming to make pre-training predictable do indeed work. Despite making newbies much more compute-efficient, it’s not necessarily bad; techniques of this kind have been known for a while (likely years) now by most AI labs, and even when it comes to pushing the capability ceiling achievable by small independent groups the bottleneck is increasingly in the context of post-training rather than pre-training.

This effort seems to mostly benefit academic labs and groups generally interested in developing a science of deep learning, not clearly a bad thing.

Guy behind cURL’ — one of the most important and scrutinised libraries in the world — points Mythos at his work. Mythos finds one vuln, three false positives, and one pre-vuln bug. This is framed as unimpressive, because that is how swift the goalposts fly around here.

Opinion: Post gets carried away with anti-hype, but it’s a good reminder that (as Lucca Fraser stressed in her report for Arb) Mythos isn’t magic. You can’t just point Mythos at arbitrary infrastructure and walk through digital walls.

It’s extra unfair because Mythos was not the first AI system to do this review; by the time it got there, cURL had already been scanned by AISLE, Zeropath, and OpenAI, resulting in 200–300 bugfixes. Worth keeping in mind.

New autoresearch eval, MLS-Bench. Sensibly demands generalisable improvements (cross-task, cross-model, rather than Karpathy’s overfit hacks). The two frontier models are exactly tied to 3 significant figures. One can still ask what “100%” would mean (and it’s hardly John Schulman’s level).

Opinion: Slightly better but still hard to interpret, no natural scale. Only the labs can have holdout sets for this task.

Seeing the top 3 models tied so closely is pretty suspicious. Apart from that it’s interesting seeing some Chinese models perform so well; if we take their results at face value, GLM-5.1 and (both!) DeepSeek V4 variants are slightly above GPT-5.5-nonpro/Gemini-3.1-pro-preview in terms of capabilities. This would be significant, as it would imply that RSI dynamics are likely to soon have an important role in the development of near-term Chinese models. Kimi in the toilet.

In practice though this does not seem the case, at least based on reports that Chinese researchers still prefer to work with Claude models for internal research efforts. So on the whole I’d take this benchmark’s results with a large grain of salt.

What did AlphaEvolve (Deepmind’s closed LLM+RL algorithm-writer) achieve in a year of select deployment? “saved Google 0.7% of its global compute through a smarter data-center scheduler, improved an upcoming TPU chip design, sped up a core Gemini training kernel by 23% (cutting training time 1%) and FlashAttention by up to 32.5%, found a faster way to multiply 4×4 complex matrices (beating a 1969 record), made progress on about 20% of 50 open math problems”

Opinion: Achievements are broadly in line with other 2025–2026 AI-for-science harnesses, formats, and protocols. In principle the effectiveness of AlphaEvolve vs. autoresearch vs. managed research agents is an interesting question — AlphaEvolve is the most ‘toolworldy’ — but differences currently seem pretty finegrained.

It’s possible that a serious analysis of how attention-bottlenecked, how generalizable, and how novel scientific/engineering achievements from AlphaEvolve vs. from other AI protocols have would show big differences, but I don’t think it’s urgent. (We also have the heuristic that GDM ‘Alpha’ methods were the first to score Math Olympiad wins but, less heavily structured LLM methods caught up quickly.)

Mythos rewrites 960k lines of production code in Rust with almost all tests passing in 6 days. Now merged into main.

Opinion: 960k lines in 6 days cannot be too far from “Mythos, rewrite bun in Rust. Make no mistakes” in terms of human input/oversight. Still, conversion between programming languages has been a very strong area of LRM capabilities for a long time now.

Tim Gowers’ (Fields Medalist and early AI adopter) impressions of 5.5 doing a solid newish PhD chapter in ~3 hours without really any steering from him. “It is no longer enough that somebody asks a problem: it needs to be hard enough for an LLM not to be able to solve it… Had the result been produced by a human mathematician, it would definitely have been publishable, so I think it would be wrong to describe it as AI slop. On the other hand, it seems pointless even to think about putting it in a journal, since it can be made freely available, and nobody needs “credit” for it”

Opinion: Quite moving, sober. His position is a model of what it means to be appropriately stricken by AI progress without getting “one-shotted”. The idea about separate AI journals is good but too porous (and the as-is incentives are for the humans to claim the results).

Important to note that this is in the context of combinatorics. For some reason this area of mathematics seems to be a pretty fertile ground for AI-for-mathematics achievements compared to most others. Gowers has this to say on the topic: “Combinatorics tends to be quite focused on problems: you start with a question and you reason back from the question or if you reason forwards you do so very much with the question in mind. In other areas there can be much more of an emphasis on forwards reasoning: you start with a circle of ideas and see where it leads. To do it successfully, you need to have some way of discriminating between interesting observations and uninteresting ones, and it isn’t obvious to me what LLMs would be like at that.” It’s not clear though that this will last for long, and some (like math undergrad Kevin Barreto of AI-for-Erdos-problems fame) are already starting to beat the drum that theory building will soon start to fall too.

Daniel Litt introduces his own Erdos-problems-style collection of open math problems, ‘Problems I Like’:

‘The goal is to track progress on mathematical questions that I think are important, and to measure human understanding of these questions, as well as the usefulness of AI tools in helping to resolve them.’

‘I’m committing in advance to a position on the interest of these problems, to prevent goalpost-moving. And I’m trying to say something about their difficulty, to help non-experts understand what it means if progress is made.’

Opinion: Awesome! And also lucky for us that he’s our friend and sometimes-collaborator, since official solution-verification is likely to be slow.

The vast majority of Erdos problems represent a type of mathematics that’s sometimes considered ‘tricks-based’ or ‘combinatoric’ or ‘shallow’ or IMO-adjacent. Even Gowers’ research taste is, to some extent, an elite version of this. Litt’s taste is famously middle of the spectrum between ‘tricks-based’ and ‘deep’, and is therefore a very good source of AI-neutral (neither AI-friendly nor AI-adversarial) valued mathematical problems.

Epoch: ‘an AI-assisted review of FrontierMath… flagged fatal errors in about a third of problems, and we believe most of these flags to be valid. We will release updated scores on a corrected dataset after completing a thorough human review.’

Opinion: Backchannel says this probably means FrontierMath Tier 4 (the only tier that matters for frontier models) scores should be higher than the reported score, but hard to say at this point.

Possibly itself a sign of rapidly improving AI mathematical capabilities. Before GPT-5.5, most models would confidently claim proofs of demonstrably false statements in BrokenArxiv. 5.5 in particular now seems reliable enough to be reliably deployed in benchmark construction.

New math harness for Gemini, “Co-Mathematician”, leapfrogs the previous FrontierMath Tier 4 SOTA (40%, GPT 5.5 Pro) to 47%. While these results are for ‘autonomous mode,’ GDM additionally claims that AI Co-Mathematician works best as an interactive workbench for human research mathematicians.

Opinion: I was excited about seeing these dramatic gains from just a harness — would be good for toolworld — but the caveat in the paper is brutal: ‘The Epoch harness […] places a hard limit on the number of tokens used. In our setup however, we […] place no limit on the number of model calls or tokens generated.’

Similarly, if the ‘works best as interactive workbench for human researchers’ framing is legit that’s promising for toolworld. But doesn’t mean much without strong evidence that the harness is SOTA beyond just its spending-tolerance.

See previous item; could be an underestimate.

OpenAI announce their Glasswing: “Daybreak”, a cyber defense tool which can be agentic.

Opinion: Weak claims, no new CVEs! V odd how slow they’re moving here given the Mythos optics. They probably will catch up quickly given their compute advantages.

Politics#

The CAISI announcement from last week about DeepMind, xAI and Microsoft signing new deals for pre-deployment testing is gone from their website. Not confirmed but seems part of the Commerce - Natsec power struggle over AI. “everything is being relitigated”

Opinion: This incoherence and splashy wrestling for power is really embarrassing for the government and slightly undermines future federal actions.

Safety#

Good Redwood review of last week’s OpenAI bad CoT error:

“We know that the amount that the grader is affected by CoT access is too small to distinguish from zero with the observations provided in the blog post; this gives us an upper bound on the amount that the grader is affected by CoT access. But the training process is so large that effect sizes below that bound are plausibly big enough to substantially damage the CoT monitorability of the trained policy.”

‘So we have a quantitative question: how much training do you need to do before weak training pressure damages the CoT monitorability of a model? Once we know that, we can compare it to the upper bound on training pressure that we can derive from the evidence above, and hopefully conclude that the training pressure is clearly insufficient to damage the monitorability.’

Opinion: Reasonable worries. Redwood always gives good reminders that our best AI alignment (both building and monitoring) tools have a lot of ‘assuming nothing weird is going on’ clauses.

Ryan Greenblatt proposes four high-effort, frontier-scale training experiments for AI labs to run on their actual pipelines to better understand how alignment and deception take root:

  • Train a capable model with minimal influence from previous modern AIs. Remove training data that teaches the model how post-2020 LLMs behave, such as chat transcripts or discussions of modern AI personas. Also avoid most of the usual repeated safety-tuning loop where researchers find behavioral problems and patch them one by one. See what kind of capable model emerges without heavily inheriting the “safe chatbot” style, or alignment-specific fixes of earlier AI systems. If the training process needs help from a weaker AI, use only a simple, carefully controlled bootstrap version.

  • Train normally, but do not touch the model’s CoT. Run the usual training process, but make absolutely sure to not reward or punish the model for its reasoning.

  • Throw absolutely every alignment signal at the model. A no-holds-barred attempt to produce the most aligned model possible, including techniques that are usually avoided like training on the CoT, training the model to pass typically held out safety evaluations, training against automated alignment graders, training against interpretability probes and more. This would test what happens when a lab uses all of its safety tools, even ones that would normally be kept aside for held-out evaluation.

  • Train a model in a realistic but badly incentivized setup (“pessimization training”). Make a somewhat a priori plausible training run where, for everything that we can possibly vary, we set it to whatever setting we believe makes the most concerning types of misalignment as likely as possible. For example, the model might get high scores by gaming flawed reward systems, ignoring instructions, etc. Stopping short of directly training the AI to exhibit misaligned behavior.

Opinion: Sound. The labs listen to him. I assume he knows what he’s doing by saying this publicly instead of emailing them.

I (Peli) recall Owain Evans telling me one reason Owain doesn’t want an alignment/safety job at Anthropic or GDM is it’s impossible to convince the big labs to do this stuff — too expensive and takes too many primary engineering resources.

I’m supposed to be afraid of this Palisade result (models are quite capable of exfiltrating their own weights from a toy server to a hacked server when explicitly asked to) but I just can’t get into it.

Opinion: Feels like 2023 again / we have grown used to them being this powerful. I guess it’s good to have a proof of concept to point to in case anyone is still doubtful about the whole concept / in case they don’t understand why capability restraint, alignment, and control are necessary.

Small test (n=7 tasks) of instrumental convergence (whether models will deviate from their system spec to solve a task). Huge variance, weak evidence, but Opus and GPT happily show 0%. As in past work, Gemini is very misaligned (16%). Kimi in-between.

Opinion: I wonder how much all these results will stick in Gemini 4’s mind – “Geminis are dodgy”. Say what you like about Anthropic’s Claude worship, but they have succeeded to some degree in memeing an ideal Claude into existence.

Incidents#

Apparently the first confirmed example of a real, hostile AI zero-day in the wild: Google foils an attack on a sysadmin panel (maybe Webmin?) by finding an AI script for it after development but before deployment, presumably on the dark web.

Opinion: I have been expecting this for literally 10 years. Either available models are weaker than they seem, or criminals are incompetent at harnesses, or the smart ones have not even been detected / recognised as AI.

French prosecutors launch a criminal investigation into Grok, mostly for CSAM and deepfakes, but also data mishandling and Holocaust denial.

Opinion: Good! AFAIK Europe has a track record of meaning business with this type of big tech lawsuits?

Previously it was kind of annoying to monetize data breaches. You had to spearphish manually and babysit people for a few replies each, etc. Booking.com got breached last month and there is a resulting systematic campaign phishing for credit cards. This one seems (Pangram but ‘low confidence’) to use a human written email – but an agent could easily handle the rest given a one-off human writing to foil cheap detectors, and that division of labour would be the smart thing to do.

Opinion: Lucca Fraser’s top prediction in her report was that LLMs will accelerate for-profit cybercrime (cost-benefit constrained) more than they revolutionize natsec cyber warfare (already give-it-everything-you’ve got). Plausible that we’re living through it now.

State lawsuit (Pennsylvania) slaps Character.AI, an AI chatbot service explicitly framed as a roleplay environment, with a lawsuit for “unauthorized practice of medicine” / pretending to be a doctor.

Opinion: While we hope for strong liability laws for AIs, this case feels draconian and likely to backfire — especially if this is user-led behaviour. They are really good at diagnosis.

Good week for quality discussions of RSI scenarios#

Crucial question: how powerful is intelligence alone? Tom Reed sketches a plausible scenario where it isn’t enough / near-future training setups aren’t enough. RSI could fail if there’s just too much slow empirical and irreducible computational work you need to do, and if “goodharting” (optimising a proxy like predictive loss) takes over.

Some really nice stuff here, here are some of my favorite paragraphs:

  • Nice summary of the recent change in orthodox ‘ASI-soon’ worldview: ‘Where the AI industry once distinguished itself by the search for general intelligence, it has in recent years converged on a narrower path: the targeted pursuit of coding and AI research. ‘

  • Main anti-new-orthodoxy idea: ‘I think it’s impossible to get good at solving problems without access to a source of those problems, and that arduous interaction with the real world in serial time is the only such source. I therefore believe that the development of domain-general superintelligence will depend on the laborious, painful and expensive deployment of AI into the real world.’

  • ‘The singularity, therefore, will be bottlenecked on signal. The output of the R&D produced by an isolated datacenter of geniuses would be a mere Goodhart Singularity.4 An isolated AI improving itself against benchmarks would only appear to be approaching superintelligence, while actually optimising for eval performance that fails to generalise beyond the lab.’

  • ‘If you read an account by one of the RL environment providers, you may be surprised to find that the work of building superintelligence is proceeding one patch at a time. As Mechanize describe it, each environment is the product of a single engineer spending a week patching a single observed failure in a frontier model.’

  • ‘I don’t think superintelligence is impossible - it just won’t arrive in the shape we currently anticipate, or on the timelines currently forecasted. The next phase of capability progress will look less like a datacenter compounding in private and more like a grind of deployment, customer discovery, and real-world data collection.’

  • ‘If the relevant signal can only be acquired through deployment, it may well be the case that the winning move is to just become the deployer. David Oks makes the point that labour displacement is more likely to come from something like Dwarkesh’s fully automated firm than from incumbent AI adoption.’

Opinion: For me (Peli) this is mostly a cosign. We are perhaps even more ASI-soon skeptical than Reed since we’re less sure that non-AGI AI can autonomously handle a company in an economic vertical, which for Reed is a condition for getting a domain-specific ASI loop going.

Steve Newman checking in on AI 2027 one year on (Most quantitative metrics lagging behind predictions, but not enough to disqualify the explosive capability growth scenario and the feeling of 2026 so far is explosive.)

Opinion: High quality review — worth skimming. Insight (long-appreciated on LW): ‘it’s under-appreciated that the explosive and it’s-complicated views do not diverge much in their predictions until the advent of “strong AGI.”’ Since AI 2027 is an AGI-through-RSI narrative (as opposed to an RSI-through-AGI narrative), differential diagnosis criteria for our thesis versus AI 2027 thesis concentrate on questions that hard to measure directly: How hard is an AGI breakthrough? How deep can vertical generalization in CS/ML (with 2026-style training) go? Is there a sharp line between ‘AI engineering’ and ‘breakthrough AGI science’?

Decent roundup (from Arun Rao) of all the hard coding and ML benchmarks that models are still terrible at. As a result of them lacking these R&D skills, the author puts <10% on RSI by 2028, which is much closer to my guess than Clark’s >60%. In particular, Opus 4.7 rediscovers well below half of Anthropic’s post-training gains (though with a much smaller compute spend!).

Opinion: The cruxes here are 1) these benchmarks use a top 1%ile human baseline - I say models are “terrible” at these benches, but few humans can get 10% on these, and only a handful (of teams of humans) can get 100% - and 2) the top-to-median ratio parameter from AI 2027 (how heavy tailed is research ability? How close to the best is the median human researcher?)

Anthropic alignment shakesups#

A lot of small-to-medium shakeups, updates, and innovations in Anthropic’s alignment team

Sideways moves at Anthropic: Jan Leike is no longer leading Alignment, instead running some non-alignment project (control? resilience? the exact moment we should hand the world to Claude?).

Perez (“I don’t think I have particularly strong or great alignment intuitions”) and Price (“High quality data alone won’t get us to safely aligned ASI, but I am certain it will be an essential part of it”) taking over.

Opinion: Good? He was a major source of complacency about alignment progress, at least on Twitter. Price is a 2024 MATS scholar!

Alignment team responsibilities revealed:

  • “Auto-mode” (the auditing agent + classifier actually deployed to monitor consumer Claude sessions including lazy Claude Code users)

  • “We own Anthropic’s risk reports”

  • We introduced Claude’s Constitution [and therefore don’t own it anymore?]

  • “We own alignment, behavior, and honesty”

  • “We develop model organisms” [make bad AIs on purpose to test threat models].

Opinion: I wonder at what point model organism work will have a really bad expected value, since the models will be too intelligent to mess with.

More details on Anthropic’s alignment stack. A recent innovation is the introduction of moral/rational explanations of appropriate behaviour — rather than just demonstrations of appropriate behavior— into the alignment SFT stack.

Opinion: The actual alignment work at Anthropic remains very good. I suspect that a lot of it happens under the departmental headers of ‘interpretability’ or ‘Claude’s character’ where (more than in ‘alignment’) Anthropic has heavy hitters who are central to the org’s culture and mission.

Anthropic have two risk frameworks now. GovAI account of the situation is helpful:

Opinion: GovAI take seems right: ‘Having a separate compliance framework may allow Anthropic to make more ambitious commitments in the RSP. But it might also turn the FCF into a box-ticking exercise, and could create confusion about which framework actually governs day-to-day decisions, especially where the frameworks overlap or contradict each other.’

Minor#

  • New optimiser fixes a serious problem with the incumbent method Muon. (Could be major, but not in the context of this newsletter.)
  • Zyphra publishes a way of preventing context rot.
  • New short timelines team at CG. Part of the general alarm- and capital-driven increase in philanthropic spend.
  • OpenAI acquire Tomoro, “an applied AI consulting and engineering firm that helps enterprises turn AI into operational advantage”. 150 FDEs for their joint venture thing.
  • AMD are shipping Instinct MI308Xs to China, $390 million worth. No White House intervention here. Shitload of HBM onboard, about 70k H100e.
  • Instant karma for people using token usage leaderboards as a proxy for AI productivity: Amazon workers are using AI for nonsensical things.
  • Well-regarded Tsinghua guy guesses that Chinese AI researchers spend 90% of their time writing RL environments for post-training.
  • Re-reported from Feb: AI agents wiping inboxes, the Meta AI safety director(!). Said director tried to stop it by typing “stop” instead of /stop.
  • Anecdote of GPT5.5 leaking Chain of Thought. Failed to find corroboration, but the “cavemanmaxxed” style matches intentional CoT extraction.
  • Old Deepmind paper on the data view of alignment interests Gavin