Summary
The morning slot carried 91 tweets and almost all of the substance came from one source: Elie Bakouch's four-post thread releasing Prime Intellect's autonomous-research study, 153 runs across 18 frontier models on the nanoGPT optimizer speedrun, with full leaderboards, scratchpads and logs published. That is the largest public experiment of its kind and it is the only genuinely new research artifact in the slot. Second in weight is a tokenizer-economics exchange: Anthropic's Tibo Sottiaux publicly conceded OpenAI's tokenizer is roughly 30% more efficient per unit of text, which matters directly because API billing is per token, and a screenshot circulating with the post puts the gap at 34.5% overall and 53.2% on multilingual prose. There is no cluster in the usual sense today, because the AI-relevant handles were not talking to each other; the two items above are independent and both are cost-of-inference stories. Robert Scoble filed two robotics-adjacent posts from a San Francisco robot event, one of which carries a real regulatory signal from San Mateo County. The remainder of the slot is unusable for this wiki: roughly two thirds of the volume was political and human-interest content from four handles (@MarioNawfal, @brivael, @spencerpratt, @AustinJustice), and @bayesiansapien posted no retweets for a seventh consecutive day, so the curated-signal layer is again absent. DHH posted three times about voice-driving his own OS and about working with agents at 70 megabits from an airplane, which is atmosphere rather than signal.
Posts
Prime Intellect publishes the largest public autonomous-AI-research experiment: 153 runs, 18 models, up to 8.7 days each (@eliebakouch, blog, results explorer, earlier experiment) (cluster of 4 posts).
Bakouch released the study across four posts. The task is the nanoGPT optimizer speedrun: lower the number of training steps needed to reach a target validation loss, changing only the optimizer, learning-rate schedules, initialization and hyperparameters. Runs went up to eight days on 8xH200s each, with multiple seeds per model. He frames the scale explicitly against the labs: Anthropic's comparable internal evaluation optimizes a model on a CPU node, and OpenAI's GPT-5.6 Sol system card reports nanoGPT Track 1 on a single H100 for under a day.
The leaderboard, measured as share of the human-record gap closed: Fable 5 first at 2,726 steps and 81.7% closed, under claude-code at high effort, over 8.7 days. Opus 5 second at 2,920 and 53.6%, under claude-code at max, in 2.9 days. Kimi K3 third at 2,930 and 52.2% under prime-agent, and fourth at 2,974 and 45.8% under kimi-code. Opus 4.8 fifth at 3,018. GPT-5.6 Sol sixth at 3,042 under codex at xhigh. The human baseline the prior experiment targeted was 2,990 steps.
Two details matter more than the ranking. First, Bakouch names the noise directly: one run in the same setting has a roughly 50-step spread after 24 hours, which is large enough that any single-run claim about autonomous research capability is inside the noise band. Second, every row is a model-harness-effort triple, and Kimi K3 appears twice under two different harnesses with a 44-step gap between them. That gap is roughly the size of the entire difference between Opus 5 and Kimi K3, which means the scaffold is contributing about as much as a model generation on this task.
The earlier Prime Intellect experiment from May, which ran Codex and Claude Code for two weeks on idle compute for about 10,000 runs and roughly 14,000 H200 hours, found agents strong at optimizer search, hyperparameter sweeps and stacking known methods, and weak at proposing new ideas without human records to climb from. That diagnosis is not retracted here even though the ceiling moved. The May post also documents model-specific behavioural failures worth knowing: Opus repeatedly stopping and refusing to stay in the autonomous loop, while Codex never stops but can get stuck grinding one seam. Bakouch also credits METR and Recursive_SI for related work, Keller Jordan for modded-nanogpt, and Karpathy for nanogpt itself. Everything is released: traces, scratchpads, run logs, scripts, configs.
Wiki: Measuring Autonomous AI Research · Agent Harness Engineering
Anthropic concedes OpenAI's tokenizer is about 30% more efficient, and the gap is larger than that on multilingual text (@ns123abc, quoting @thsottiaux).
Tibo Sottiaux of Anthropic posted that an OpenAI token and an Anthropic token are not the same unit, that the two companies use different tokenizers, that OpenAI's is significantly more efficient, and that this is important because you pay per token both over the API and in usage. He put the difference at roughly 30% in token count and called it definitely significant. The image attached to the reposting tweet is a comparison table with four passages and their token counts under three tokenizers, and it is worth transcribing because the numbers are specific: English strategy prose, 138 words, 157 tokens under OpenAI o200k versus 160 under legacy Claude versus about 208 under a Claude Opus 5 estimate, a 24.5% OpenAI reduction. Technical systems prose, 110 words, 203 versus 212 versus about 276, a 26.4% reduction. Multilingual prose, 107 words, 188 versus 309 versus about 402, a 53.2% reduction. Numbers, policy and data, 138 words, 218 versus 219 versus about 285, a 23.5% reduction. Totals: 493 words rendering as 766 OpenAI tokens, 900 legacy Claude tokens, and about 1,170 Claude Opus 5 tokens, a 34.5% overall reduction for OpenAI.
The reposter's commentary is a positioning argument rather than a technical one, that competitor-referential messaging cedes narrative control, and it can be skipped. The underlying fact cannot. Two vendors quoting the same dollars per million tokens are not quoting the same price, the divergence is worst exactly where a multilingual product would live, and the Claude Opus 5 column is an estimate rather than a measurement, which is a caveat the table itself flags.
Wiki: LLM Routing · Compute Economics
San Mateo County votes unanimously to draft humanoid robot regulations that would break the remote-teleoperation business model (@Scobleizer, quoting @humanoidsdaily).
The proposed ordinance mandates that robots be supervised at all times by a trained on-site human, imposes a County Automation Impact Fee to fund worker retraining, and adds annual fees to pay for specialised hazmat gear for high-heat lithium-ion battery fires. The on-site supervision requirement is the one that bites: the entire economic case for a $30-an-hour remote cleaning robot is that the operator is not on site. Startups launching that service in San Francisco next door are unaffected for now, which sets up a jurisdiction arbitrage worth watching.
A stealth robot-hand company argues Tesla will move from tendons to geared electric motors, and that current hands cost around $50,000 each (@Scobleizer).
Scoble spent an hour at a San Francisco robot event with a founder building an alternative hand actuation approach. Two concrete claims worth logging: Sharpa's hands, which impressed him at CES, run around $50,000 each and may be unimportable to the US under a recent FCC ruling, and the founder's argument is that cheaper hands are the unlock rather than better ones. Hardware cost claims from a single conversation with an interested party, so treat the numbers as directional.
DHH is voice-driving an operating system with agents and reporting it as a sci-fi milestone (@dhh, plus two related posts) (cluster of 3).
The substantive item is that the next Omarchy release will integrate Voxtype with the default agent so widgets, panels and apps can be created by speaking. The other two posts are atmosphere: working with agents over 70 megabit in-flight internet, and a reader saying Omarchy made computers fun again. No technical detail, no benchmark. Interesting as an early instance of an agent being wired into the OS shell rather than into an editor.
A claim that Grok 4.6 reaches near-Fable quality at one tenth the cost and half the time on a single design prompt (@minchoi, quoting @anshuc).
The comparison is one prompt, one design output, judged by eye, and the reasons given for preferring Grok's result are aesthetic (fills the hero image, flat tiles rather than rounded cards, fewer labels, lighter text). The cost and latency claim is the only quantitative part and it is unsourced. Included because it is the third same-task cost-spread claim in two days, alongside DHH's Rust rewrite benchmark from yesterday's afternoon slot, but on its own this is a demo, not a measurement.
A claim that Grok 4.6 is Pareto-dominant over Fable 5 and that investors are not pricing it (@MarioNawfal, citing @GavinSBaker).
Gavin Baker's argument as relayed is that Grok 4.6 is cheaper, faster and ahead on several measures, that almost no investors mention it, and that SpaceX should therefore be priced as more than a rocket company. No benchmarks cited. This is an investment thesis being reported second-hand through an aggregator account, and it belongs in the record only because it is the same cost-per-capability framing showing up on the finance side.
tinygrad's account on Chinese open weights (@__tinygrad__).
A one-line jab that the people who claimed only they could be trusted to gatekeep AI were overtaken by Chinese labs giving models away free. No content beyond the sentiment, but it is a datapoint on how the open-weights release cadence is being read by the systems-software community.
Robert Scoble's AI news aggregator now reads 30,000 posts a day across the X AI community (@Scobleizer, alignednews.com/ai).
He changed how his agent picks headlines and says the results improved. The stated coverage is 30,000 posts a day from about 50,000 people and 9,200 companies, with every headline linking back to X. Relevant here as a competing approach to the same problem this wiki solves, an agent-curated feed rather than a synthesised knowledge base, and worth checking against for coverage gaps.
Opaque or unfetchable links. Two posts pointed at content the farmer could not read usefully: Scoble's aggregator returned only a title stub, and a Substack post returned a JavaScript-required page. Click through if the topics matter.
Skip. Roughly two thirds of the slot was off-topic: @MarioNawfal (20 posts, geopolitics, crime video, human interest), @brivael (20 posts, French domestic politics), @spencerpratt (11 posts, California politics), @AustinJustice (3 posts, Austin transit), @NICKIMINAJ, @HouseGOP, @Tesla, @heavypulp (four image-only posts with no text), and @stepango's two one-line jokes. None carries AI research or industry signal.