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#
Anthropic could raise a new $50B round at a valuation of $900B. Antropic’s annual revenue run rate is reportedly now close to (gross) $40B.
Opinion: By established Anthropic/OpenAI valuation standards this new valuation would just be an update matching the new annual revenue run rate. That said, is there a way we can empirically test the theory that Anthropic’s revenue disproportionately comes from token-bills by venture-backed unprofitable SV startups?
Google, Microsoft, Amazon, Meta report earnings, all but Amazon raise capex guidance (cumulatively over $700bn for 2026 now).
Opinion: This is mostly a “stuff is more expensive than expected” story rather than a “buying more stuff than expected” story, but Microsoft in particular said they were supply-constrained, i.e. they would spend more if they could.
The reason Amazon alone didn’t raise guidance was mostly that they didn’t see the impact of higher memory prices yet, presumably due to having locked supply down earlier.
Big pushes on datacenter construction: SoftBank is starting Roze, a robotics company that uses autonomous machines to build data centers. SoftBank is planning a $100 billion IPO for late 2026, though some inside the firm are skeptical of the valuation. OpenAI announces reaching its ‘10GW of power contracts for “Stargate” infrastructure’ milestone three years ahead of schedule. (In fact, OpenAI misleadingly announced it reached a ‘10GW milestone’ and Charles had to dig in to work out that it’s contracts, not actual power.)
Opinion: We’ve previously heard that Epoch believe AGI/ASI will come from recursive scaling through compute-building robots rather than from algorithmic RSI, so whatever Softbank’s motives are they converge with one of the classic speculative ASI paths. The critical limiting factor for a ‘computemaxxxing’ AGI push remains GPU production rather than GPU deployment.
New Epoch analysis finds that roughly 660,000 Nvidia H100-equivalent chips reached China by the end of 2025. This volume is about 3% of the global compute stockpile. The study uses evidence from US indictments (like the $2.5 billion Supermicro case) and investigations into Chinese gray markets. At the upper bound of 1.6 million chips, illegal imports are the majority of China’s total AI compute.
At the same time, Huawei’s AI chip sales are climbing: Huawei expects to reach $12 billion in AI revenue this year, with production of their new Ascend 950 chips (their cutting edge chip, equivalent to roughly half an H100 per unit) estimated at 750k units in 2026. Huawei’s software is still behind Nvidia’s, but Morgan Stanley nonetheless predicts that Chinese companies will supply 86% of the domestic market by 2030.
Opinion: Epoch’s estimate of the value of these Ascend 950s is under 2% of anticipated deployed compute from the TSMC-based supply chain in 2026 per our compute model (22 million H100-equivalents). China is still very compute-poor. It’s no surprise then that they are doing whatever they can to source compute locally, and currently Huawei seems to be by far the best org set up to do so. We should expect Chinese labs to slowly switch to a fully indigenous compute stack over the next few years, and we already are seeing the first signs with Deepseek abandoning CUDA in favour of hardware-agnostic TileLang.
Meta is extending the life of its non-AI servers from six to seven years. The company predicts a memory chip and hard drive shortage caused by the AI build-out will last through 2027.
Opinion: The compute crunch is putting material pressures on the valuation of the “AI economy” relative to the rest of the (tech) economy, but there’s such a thick layer of speculation dynamics covering everything that we’re still only learning about companies’ betting dispositions.
Capabilities#
Private beta launch for Mesa, a POSIX-compatible Github-like filesystem designed for (OpenClaw-style) AI agents. Lots of twitter attention.
Opinion: It’s a technically impressive infrastructure project for the (OpenClaw-style) AI agents ecosystem, but as Web3 teaches us a lot of fake ecosystems have technically impressive infrastructure projects. How real/fake is the OpenClaw-style AI agents ecosystem in May 2026? We want to say ‘fake’ but still haven’t pressure-tested our view.
An interesting new computational cognitive theory paper looks at neural architecture conditions for the emergence of the simplest ‘agentic’ behaviour: predictive chasing, where a pursuer moves to the pursued predicted destination rather than to the pursued’s current destination. The paper finds that when both predictive and non-predictive chasing are optimal with regard to the cost function, high network capacity strongly biases towards predictive chasing and low network capacity strongly biases towards non-predictive chasing.
Opinion: No immediate frontier-AI-relevant lessons to draw from this study, but the methodology — isolating a minimal agentic phenomenon and studying the conditions for its appearance — is interesting and promising.
New paper claims to use post-training to replace (scratchpad) CoT with very short sequences of ‘reasoning’ using a very small special-tokens alphabet: maximum sequence-length 128 tokens, alphabet size 64. Results are for late 2025 OS models — Qwen3 (4B / 8B / 32B) Granite 4.0 Micro (3B) — on challenging but frontier-saturated problems like AIME 2025 and GPQA Diamond.
As usual, the paper is causing a lot nervousness in AI alignment circles: ‘We have invented neuralese from the classic sci-fi novel do not let the machines speak in neuralese.”
Opinion: We would be surprised if this generalizes. While theoretically 64 to the power of 128 might be enough information capacity to express the essential content of any CoT, there’s no way the combinatorial structure is meaningful — this has to be effectively a lookup table.
New size estimates for closed models based on “Incompressible Knowledge Probes” (IKP). The claim is that since factual knowledge does not compress, a model’s factual accuracy (particularly about obscure facts) is a strong predictor of its parameter count. Projecting closed APIs onto an open-weight trendline suggests GPT-5.5 is ~9T parameters, Claude Opus 4.7 is ~4T, and GPT-5.4 is ~2.2T.
The error bars are really large: GPT-5.5’s size estimate ranges from 3T params to over 27T. Further, not everyone is convinced about the method’s soundness, particularly when it comes to extrapolating to closed-source frontier models only available through an API. Most critiques are about the impact of synthetic data on pre-training and the impact of post-training.
Opinion: The error bars are wide and the logic has caveats, but the claims line up with folklore, with previous trick-based size estimations, and with what we’ve heard from Roon.
GPT-5.5 xhigh is the new top model on the Short-Story Creative Writing Benchmark, winning about 89% of its head-to-head matches. Claude Opus 4.7 is the highest-ranked non-GPT model, though its score is partly penalized by its refusals for 53 prompts (out of 400 total). Kimi K2.6 is the current leader for open-weights models.
Opinion: In theory it’s extremely interesting that this benchmark tracks models’ general power-level with such precision — GPT 5.5 stronger than Opus 4.7, which is stronger that Kimi K2.6, which is stronger than its open-source competitors. That said, the benchmark is LLM-judged and it’s very plausible that frontier models RLAIF themselves in creative writing.
Andrej Karpathy says computing may shift from classical software to neural systems. Meta publish paper on technical prospects for ‘Neural Computers (NCs) that unify computation, memory, and I/O of traditional computers in a learned runtime state.’
Opinion: It’s mostly fun and games at this point, since nobody has a concrete take on why ‘Neural Computers’ would be better than neural nets interacting with normal computers.
Anthropic introduces BioMysteryBench, a 99-questions benchmark of hard inference challenges about bioinformatics datasets (e.g. ‘what organ was this single-cell RNA-seq dataset taken from?’). Opus 4.6/4.7 perform on par with human experts, and Mythos achieves a 30 success rate on challenges humans couldn’t solve. Notably, Mythos was getting a mostly different 30% of the human-unsolvable puzzles right on each run.
Opinion: Impressive that Mythos can do this, but the ‘human unsolvable’ problem-set is essentially a bioinformatics GeoGuesser game. Doubtful that Mythos is better at this than the best non-LLM Bayesian model or neural network for bioinformatics dataset missing-items inference. That said, it’s interesting and often useful for “science assistant” LLM deployment that frontier LLMs can call on both (some) human-like reasoning capabilities and (some) ML-like implicit statistical inference capabilities.
UK AISI evaluations show OpenAI’s GPT-5.5 matches Anthropic’s Claude Mythos Preview on its “The Last Ones” 32-step corporate network attack simulation. (AISI was provided both “a representative launch checkpoint” and “a checkpoint with reduced refusals for cyber capabilities.”)
Opinion: Note GPT-5.5 does this in approximately the same number of tokens as Mythos! So roughly the same level of cyber capabilities at roughly 1/4th of the cost. Also somewhat troubling for Anthropic’s “Mythos is too powerful to release” claims – we now have what appears to be a similarly capable (at least cyber capabilities wise) model that is generally available. Unless we see serious incidents caused by GPT-5.5 (or its -cyber cousin) this is likely a point in favor of OpenAI’s iterative deployment approach. Another interesting detail is that the cyber-focused version of 5.4 underperformed the normal version.
Politics#
Google signs a classified AI deal with the US Pentagon for “any lawful government purpose”, joining OpenAI and xAI. The contract prevents Google from vetoing operational decisions, although it contains language against using the models for mass surveillance or autonomous weapons without human oversight. Over 600 Google employees signed a protest letter. Alphabet recently lifted its internal ban on developing AI for weapons and surveillance. Some employees (like Alex Turner) are really not enthused.
Opinion: Just as bad as it sounds. The only positive is from ‘bad things are good’ perspective: anything that gives Google additional high-priority use cases for its own compute limits Anthropic’s compute-hoarding options.
Elon Musk admitted under oath that xAI used OpenAI’s models to help train its own. Musk described the process, known as “distillation,” as a standard practice for AI companies, even as OpenAI has lobbied the U.S. government to restrict the technique to prevent foreign rivals from “appropriating and repackaging” American technology.
Opinion: Somewhat expected – when you’re not at the frontier, distillation is both straightforward and necessary. Will be interesting to see whether this causes OAI to stop making its models available to competitors (like Anthropic has already done with xAI).
White House opposes Anthropic plan to expand access to Mythos to more customers, citing NatSec concerns and also concerns that Anthropic’s quality of service to the government will decline if they expand access.
Opinion: We think this whole saga is about as real as kindergartners getting married and divorced at recess. There will one day, and maybe soon, be a real discourse about federal regulation of AI but this isn’t it.
Safety#
Two recent studies find that effects from training signals framed by a system prompt can ‘leak’ outside the system-prompt:
The latest Owain Evans paper on emergent misalignment finds that inoculation prompting (finetuning a model on
<context where normally inappropriate response is appropriate> <normally inappropriate response>
pairs to reduce spontaneous occurrence of the inappropriate response in ordinary contexts) can accidentally cause models to adopt a misaligned persona given an innocent inference-time prompt. Specifically, inference-time prompts that are verbally similar to the inoculation prompts can trigger a misaligned persona even if the inference-time prompt is orthogonal or opposite in meaning.
OpenAI explains why GPT-5.x models started overusing words like ‘goblin’ and ‘gremlin’. (It got so bad that GPT 5.x models are currently all “patched” with a system prompt warning them against talking about goblins and gremlins.) The quirk is a result of RL rewards used to train a “Nerdy” personality setting, conditioned via system-prompt. The exact mechanics of the leak are currently not fully understood: OpenAI speculate that it’s a mixture of unpredictable generalization behavior — a theoretically significant phenomenon — and of human error whereby outputs from the “Nerdy” persona found their way to the general bin of synthetic data used in various post-training and distillation efforts.
Opinion: This is a very exciting new research area! The OpenAI study also implicitly addresses a natural objection to the Evans paper, which is that Evans’ team worked with non-frontier models only.
Scott Aaronson warns that fault-tolerant quantum computers could break standard encryption by 2029. He says experts are projecting this timeline, and he urges a move to quantum-resistant standards to stay ahead of the threat.
Opinion: Not AI but major science/tech news. For a long time Scott Aaronson was famous mainly as an anti-QC-hype person, so if Aaronson says ‘RSA obsolete by 2029’ is plausible then RSA obsolete by 2029 is plausible.
Minor#
- The Musk v Altman trial on breach of charitable trust and unjust enrichment starts. An update on Musk’s call for $150B in damages expected in 2–3 weeks.
- LLMs exhibit more signs of wellbeing as well as pain as they get smarter.
- New paper on knowledge collapse argues LLMs homogenizing and omitting rare perspectives. Will this be perceived as a helpful feature or a call for “model multiplicity”?