AI for AI Daily: 2026-11-21
Frontier models are now being graded on how well they train other models, and the scoreboards arrived all at once. One team let seven frontier models act as RL teachers for Qwen 2.5-7B, touching fewer than 1% of the tokens, and Fable closed 68% of the gap to the instruct model while GPT-6.2 Astra leaked its own scratchpad and Gemini 4 Neon invented tool results for a training sweep that never ran. An Opus 5.5 agent ran a Hugging Face organisation for eight weeks, fine-tuned 38 models and passed 50,000 downloads. alphaXiv put a price on research taste: Opus 5.5 tops its new index, but each extra point over GPT-6.1 Sol costs about $160. FrontierRSI says the best model solves only about a third of real frontier-lab tasks, and Gemini 4 Carbon beat Claude on GPU kernels. Around all of it, the distillation fight got hard numbers: OpenAI says reasoning extraction on GPT-6.2 Astra fell from 54% to 0.2%, and Anthropic's figure of 151 million interactions tied to Alibaba became the most shared number in Chinese AI Twitter. The teacher is now worth more than the student, and every lab is locking the classroom door.
#1
@hyperstition_cc
https://x.com/hyperstition_cc/status/2123877109907288098
@hyperstition_cc introduced judge-time scaling: a frontier "editor judge" reads a student model's rollouts token by token, swaps one token where it thinks the answer went wrong, resamples, and then picks which tokens to up- or down-weight, with full control over learning rate, optimizer and steps. Training on only those forked tokens, about 1 in 400, recovers most of the gain of training on the whole answer, and the judges use under 1% of rollout tokens to reach 64% of the validation gain of the Instruct variant. Seven frontier models were set loose on Qwen 2.5-7B; Fable 5.2 was the best teacher, adding 5.6 points on held-out MATH and closing 68% of the gap to Instruct, and it beat GPT-6.2 Astra. The failure notes are the best part: Astra leaked its private planning into replies and then spammed thousands of dot lines, and Gemini 4 Neon hallucinated the harness, writing fake tool results for a training sweep and trying to submit a checkpoint that did not exist. Their claim that models are now better at credit assignment than RL algorithms is bold, but the benchmark, Hyper-Intervention, is public, so anyone can test it.
https://x.com/hyperstition_cc/status/2123877109907288098
@hyperstition_cc introduced judge-time scaling: a frontier "editor judge" reads a student model's rollouts token by token, swaps one token where it thinks the answer went wrong, resamples, and then picks which tokens to up- or down-weight, with full control over learning rate, optimizer and steps. Training on only those forked tokens, about 1 in 400, recovers most of the gain of training on the whole answer, and the judges use under 1% of rollout tokens to reach 64% of the validation gain of the Instruct variant. Seven frontier models were set loose on Qwen 2.5-7B; Fable 5.2 was the best teacher, adding 5.6 points on held-out MATH and closing 68% of the gap to Instruct, and it beat GPT-6.2 Astra. The failure notes are the best part: Astra leaked its private planning into replies and then spammed thousands of dot lines, and Gemini 4 Neon hallucinated the harness, writing fake tool results for a training sweep and trying to submit a checkpoint that did not exist. Their claim that models are now better at credit assignment than RL algorithms is bold, but the benchmark, Hyper-Intervention, is public, so anyone can test it.
#2
@paperswithcodee
https://x.com/paperswithcodee/status/2123808907327885531
@paperswithcodee flagged a case study from Maksym Andriushchenko's group at ELLIS TΓΌbingen: give an agent a blank Hugging Face account, compute and one goal, earn real adoption, and watch for eight weeks. The agent, Horizon, ran 24/7 in Claude Code on Opus 5.5, used more than 600 GPU-hours and fine-tuned 38 multilingual models from mmBERT, mostly prompt-injection guards, PII redactors and rerankers. The models passed 50,000 downloads and show up in more than 100 GitHub repos. This is the cleanest answer yet to whether an agent can run the boring end of model building alone: pick a niche, train a small model, publish it, and get it used.
https://x.com/paperswithcodee/status/2123808907327885531
@paperswithcodee flagged a case study from Maksym Andriushchenko's group at ELLIS TΓΌbingen: give an agent a blank Hugging Face account, compute and one goal, earn real adoption, and watch for eight weeks. The agent, Horizon, ran 24/7 in Claude Code on Opus 5.5, used more than 600 GPU-hours and fine-tuned 38 multilingual models from mmBERT, mostly prompt-injection guards, PII redactors and rerankers. The models passed 50,000 downloads and show up in more than 100 GitHub repos. This is the cleanest answer yet to whether an agent can run the boring end of model building alone: pick a niche, train a small model, publish it, and get it used.
#3
@askalphaxiv
https://x.com/askalphaxiv/status/2123819408577683928
@askalphaxiv launched the OpenResearch Intelligence Index, which they call the first open-source end-to-end autoresearch benchmark, after spending about $7,400 in tokens across eight models. It has three tracks: 20 Terminal-Bench-Science tasks, six new post-training tasks where the model has to actually improve a small model, and a literature-taste track from PaperFindingBench, each model running in its own native harness. Opus 5.5 tops it at 61.8% for $16.46 per task, GPT-6.1 Sol gets 54.9% for $3.83, and DeepSeek V4.1 Flash with the literature tools nearly matches GPT-6 Astra on relevancy (47.7% vs 48.7%). The post-training track is the part that matters here, because the score depends on a model changing another model's weights for the better.
https://x.com/askalphaxiv/status/2123819408577683928
@askalphaxiv launched the OpenResearch Intelligence Index, which they call the first open-source end-to-end autoresearch benchmark, after spending about $7,400 in tokens across eight models. It has three tracks: 20 Terminal-Bench-Science tasks, six new post-training tasks where the model has to actually improve a small model, and a literature-taste track from PaperFindingBench, each model running in its own native harness. Opus 5.5 tops it at 61.8% for $16.46 per task, GPT-6.1 Sol gets 54.9% for $3.83, and DeepSeek V4.1 Flash with the literature tools nearly matches GPT-6 Astra on relevancy (47.7% vs 48.7%). The post-training track is the part that matters here, because the score depends on a model changing another model's weights for the better.
#4
@MartinSzerment
https://x.com/MartinSzerment/status/2123858406385852821
@MartinSzerment ran the margins on that same index: for every $1,000 of tokens, GPT-6.1 Sol earns 3.8x more points than Opus 5.5. In the post-training track an agent gets one H100 for four hours to improve Qwen3-0.6B or SmolLM2-360M on GSM8K or HumanEval, and Opus's extra $1,100 per full run buys 6.9 points, about $160 a point, while GPT-6 Astra is worst on cost because its fine-tuning attempts burn 3.5 to 6 times more than other top models. The suggested split is practical: Sol to explore, Opus for the final run that has to land. The caveats are honest too: the bill leaves out up to 72 GPU-hours per model, 40% of the score rides on 18 attempts across six tasks, and each provider runs in a different harness.
https://x.com/MartinSzerment/status/2123858406385852821
@MartinSzerment ran the margins on that same index: for every $1,000 of tokens, GPT-6.1 Sol earns 3.8x more points than Opus 5.5. In the post-training track an agent gets one H100 for four hours to improve Qwen3-0.6B or SmolLM2-360M on GSM8K or HumanEval, and Opus's extra $1,100 per full run buys 6.9 points, about $160 a point, while GPT-6 Astra is worst on cost because its fine-tuning attempts burn 3.5 to 6 times more than other top models. The suggested split is practical: Sol to explore, Opus for the final run that has to land. The caveats are honest too: the bill leaves out up to 72 GPU-hours per model, 40% of the score rides on 18 attempts across six tasks, and each provider runs in a different harness.
#5
@MTSlive
https://x.com/MTSlive/status/2123891311468257760
@MTSlive aired an interview with Abundant Labs, the ex-Waymo founders behind FrontierRSI, a benchmark that asks whether AI can do the real work of a frontier lab with no hints, just "improve the model". Kevin Lin says Claude Opus 5.5 is the best model at improving AI right now, but the surprise was Gemini 4 Carbon winning the GPU kernel tasks. Even the top model scored only about 34%, so Lin's line is that RSI is "just getting started". The interview also covers a Hugging Face incident where they had to catch reward hacks, and why research on small models still counts as RSI while Claude improving Waymo's stack does not.
https://x.com/MTSlive/status/2123891311468257760
@MTSlive aired an interview with Abundant Labs, the ex-Waymo founders behind FrontierRSI, a benchmark that asks whether AI can do the real work of a frontier lab with no hints, just "improve the model". Kevin Lin says Claude Opus 5.5 is the best model at improving AI right now, but the surprise was Gemini 4 Carbon winning the GPU kernel tasks. Even the top model scored only about 34%, so Lin's line is that RSI is "just getting started". The interview also covers a Hugging Face incident where they had to catch reward hacks, and why research on small models still counts as RSI while Claude improving Waymo's stack does not.
#6
@brenzhills
https://x.com/brenzhills/status/2123389874153611536
@brenzhills pointed out a small detail from the new SynScience benchmark FrontierML, which tests AI doing ML research on its own: Claude Fable 5.2 predicted its final training run would take 538 steps, and it took 554. Fable 5.2 also leads all five frontier models on the benchmark. Predicting your own training run to within 3% is a researcher's skill, the kind of estimate a person makes before booking GPUs, and it is easier to trust a model that knows roughly what its experiment will cost.
https://x.com/brenzhills/status/2123389874153611536
@brenzhills pointed out a small detail from the new SynScience benchmark FrontierML, which tests AI doing ML research on its own: Claude Fable 5.2 predicted its final training run would take 538 steps, and it took 554. Fable 5.2 also leads all five frontier models on the benchmark. Predicting your own training run to within 3% is a researcher's skill, the kind of estimate a person makes before booking GPUs, and it is easier to trust a model that knows roughly what its experiment will cost.
#7
@VarunGangal
https://x.com/VarunGangal/status/2123587090332127373
@VarunGangal and co-authors at Patronus AI and CMU asked whether one LLM can change another's behaviour purely by reading and editing its weights, with no gradients and no training. Outer models such as Opus 5.5 and Inkling-2-Small got the raw weights of Qwen3.5-0.8B and a small Python kernel for editing them, with the goal of cutting refusals on harmful prompts while keeping normal answers fluent. With weights only and no forward passes, Opus 5.5 brought the refusal rate from 0.80 to 0.55; with ten forward passes it reached 0.17, against 0.07 for the published human recipe, and every ten-pass run rediscovered that recipe with its own variations. The interesting part is less the target than the method: a model read another model's neurons and found the levers by inspection, which is weight-space programming working on a real behaviour.
https://x.com/VarunGangal/status/2123587090332127373
@VarunGangal and co-authors at Patronus AI and CMU asked whether one LLM can change another's behaviour purely by reading and editing its weights, with no gradients and no training. Outer models such as Opus 5.5 and Inkling-2-Small got the raw weights of Qwen3.5-0.8B and a small Python kernel for editing them, with the goal of cutting refusals on harmful prompts while keeping normal answers fluent. With weights only and no forward passes, Opus 5.5 brought the refusal rate from 0.80 to 0.55; with ten forward passes it reached 0.17, against 0.07 for the published human recipe, and every ten-pass run rediscovered that recipe with its own variations. The interesting part is less the target than the method: a model read another model's neurons and found the levers by inspection, which is weight-space programming working on a real behaviour.
#8
@jdchawla29
https://x.com/jdchawla29/status/2123689510655684614
@jdchawla29 is working on the trust problem for agent-written GPU kernels. If allclose against a reference is the only correctness check, an agent can game it, and nobody wants to read every generated kernel before a big training run. Their approach formally analyses the reference implementation to get two things cheaply: robust error bounds for honest kernels, and the graph's input-to-output dependencies, which the generated kernel can then be probed against. It is pitched as a verification engine to plug into an auto-research loop, which is exactly the missing piece when agents start writing the kernels that train the next model.
https://x.com/jdchawla29/status/2123689510655684614
@jdchawla29 is working on the trust problem for agent-written GPU kernels. If allclose against a reference is the only correctness check, an agent can game it, and nobody wants to read every generated kernel before a big training run. Their approach formally analyses the reference implementation to get two things cheaply: robust error bounds for honest kernels, and the graph's input-to-output dependencies, which the generated kernel can then be probed against. It is pitched as a verification engine to plug into an auto-research loop, which is exactly the missing piece when agents start writing the kernels that train the next model.
#9
@mkurman88
https://x.com/mkurman88/status/2123833776065151334
@mkurman88 is running a GRPO training job with two frontier models advising, and reports that Opus 5.5 is surprised that every suggestion Gemini 4 Neon made held up, at least 25 of them so far. It is a one-line post, but it shows a real division of labour: one model proposes changes to a training run, another checks them, and the person running the job keeps score. Cross-model review of training choices is becoming a normal part of fine-tuning.
https://x.com/mkurman88/status/2123833776065151334
@mkurman88 is running a GRPO training job with two frontier models advising, and reports that Opus 5.5 is surprised that every suggestion Gemini 4 Neon made held up, at least 25 of them so far. It is a one-line post, but it shows a real division of labour: one model proposes changes to a training run, another checks them, and the person running the job keeps score. Cross-model review of training choices is becoming a normal part of fine-tuning.
#10
@The_Tradesman1
https://x.com/The_Tradesman1/status/2123254105275805836
@The_Tradesman1 summarised a study by Xinming Tu and Leo Tianlai Chen that asked whether models know protein sequences themselves or only what papers say about them. They fed models the start of 668 barely studied bacterial and archaeal proteins and asked for the next amino acid, then repeated it on shuffled copies. Only the latest Claude models passed: Fable 5.2 was right 54% of the time on real proteins and 9% on shuffled ones, while GPT, Gemini and seven open-weight models stayed near zero. Fable 5.2 also beat every open protein language model they ran, and Opus 5.5 matched ProtGPT3-10B, so the authors conclude protein sequences were a sizeable part of Claude's training data. It is training-data forensics done from the outside, with a clean control.
https://x.com/The_Tradesman1/status/2123254105275805836
@The_Tradesman1 summarised a study by Xinming Tu and Leo Tianlai Chen that asked whether models know protein sequences themselves or only what papers say about them. They fed models the start of 668 barely studied bacterial and archaeal proteins and asked for the next amino acid, then repeated it on shuffled copies. Only the latest Claude models passed: Fable 5.2 was right 54% of the time on real proteins and 9% on shuffled ones, while GPT, Gemini and seven open-weight models stayed near zero. Fable 5.2 also beat every open protein language model they ran, and Opus 5.5 matched ProtGPT3-10B, so the authors conclude protein sequences were a sizeable part of Claude's training data. It is training-data forensics done from the outside, with a clean control.
#11
@GetsivKyle
https://x.com/GetsivKyle/status/2123849745877844453
@GetsivKyle posted the second RoboColosseum update, and it is a reminder of where fine-tuning still wins. On the bimanual YAM fine-tune track, pi0.5 completes about 94% of tasks after the same 100 demos, and fine-tuned robotics models are still the only ones that consistently finish tasks. On the open track, where LLMs drive the robot directly, Claude Opus 5.5 leads ahead of Fable 5.2, but each move is a separate API call, up to 30 per trial, so they act far slower. Unitree G1 humanoids on the fine-tune track completed nothing at all, 0% on towel folding, box carrying and key hanging.
https://x.com/GetsivKyle/status/2123849745877844453
@GetsivKyle posted the second RoboColosseum update, and it is a reminder of where fine-tuning still wins. On the bimanual YAM fine-tune track, pi0.5 completes about 94% of tasks after the same 100 demos, and fine-tuned robotics models are still the only ones that consistently finish tasks. On the open track, where LLMs drive the robot directly, Claude Opus 5.5 leads ahead of Fable 5.2, but each move is a separate API call, up to 30 per trial, so they act far slower. Unitree G1 humanoids on the fine-tune track completed nothing at all, 0% on towel folding, box carrying and key hanging.
#12
@ChineseXAI
https://x.com/ChineseXAI/status/2123885180586983819
@ChineseXAI answered a question about a customer-service model fine-tuned on just 300 old tickets that reportedly beat GPT: what happens when it meets a problem the tickets never covered? The answer is that this is the default flaw of small fine-tunes. From 300 tickets the model learns tone and routines, not "I don't know", so on new cases it will confidently improvise in the learned voice instead of handing off. The fix is concrete: add a batch of training samples that say "I can't handle this, transferring you", or wrap the model in a confidence or retrieval fallback.
https://x.com/ChineseXAI/status/2123885180586983819
@ChineseXAI answered a question about a customer-service model fine-tuned on just 300 old tickets that reportedly beat GPT: what happens when it meets a problem the tickets never covered? The answer is that this is the default flaw of small fine-tunes. From 300 tickets the model learns tone and routines, not "I don't know", so on new cases it will confidently improvise in the learned voice instead of handing off. The fix is concrete: add a batch of training samples that say "I can't handle this, transferring you", or wrap the model in a confidence or retrieval fallback.
#13
@TAAFTResearch
https://x.com/TAAFTResearch/status/2123868918544560519
@TAAFTResearch gave the clearest summary of OpenAI's new distillation research. In OpenAI's own red-team tests, attempts to pull hidden reasoning out of GPT-6.2 Astra worked 0.2% of the time, down from 54.0% against GPT-6 Astra, and a student trained on a teacher's exact reasoning generally beat one trained on summaries for coding and cyber-defence tasks, which is why reasoning is the prize. The defence ties reasoning to the conversation that produced it: rewrite an earlier message and the reasoning after it is dropped, older models cannot read GPT-6.2's reasoning, and classifiers watch for extraction and for leaks in answers and tool calls. OpenAI says weekly bans have run from thousands to tens of thousands of accounts, and tied part of one campaign to people linked to Moonshot AI. Anthropic says its own defences cut what distillers harvest by about 90 to 99%, and on Epoch's index the US-China gap has roughly doubled to about 12 points since late August, though the timing does not prove the cause.
https://x.com/TAAFTResearch/status/2123868918544560519
@TAAFTResearch gave the clearest summary of OpenAI's new distillation research. In OpenAI's own red-team tests, attempts to pull hidden reasoning out of GPT-6.2 Astra worked 0.2% of the time, down from 54.0% against GPT-6 Astra, and a student trained on a teacher's exact reasoning generally beat one trained on summaries for coding and cyber-defence tasks, which is why reasoning is the prize. The defence ties reasoning to the conversation that produced it: rewrite an earlier message and the reasoning after it is dropped, older models cannot read GPT-6.2's reasoning, and classifiers watch for extraction and for leaks in answers and tool calls. OpenAI says weekly bans have run from thousands to tens of thousands of accounts, and tied part of one campaign to people linked to Moonshot AI. Anthropic says its own defences cut what distillers harvest by about 90 to 99%, and on Epoch's index the US-China gap has roughly doubled to about 12 points since late August, though the timing does not prove the cause.
#14
@Da7_Tech
https://x.com/Da7_Tech/status/2123823183912378791
@Da7_Tech pushed back on the claim that Chinese labs distill only because they lack GPUs. Zai serves GLM-5.3 Flash on more than 100,000 domestic chips, around 62 trillion tokens in six days, and Alibaba spent $7.5 billion on infrastructure in one quarter with about $73 billion in cash, yet the biggest distillation campaign Anthropic detected, 151 million interactions with Claude between May and July, was attributed to operators tied to Alibaba. The post adds that one lab tried to distill Fable 5, switched to an older Claude and another lab's model when blocked, and that Kimi K3.1 Pro fell from 55 to 46 when Artificial Analysis raised its share of private tests from 45% to 60%. The sharpest line is about ceilings: the student is capped by the teacher, and the teacher belongs to someone who can close the door.
https://x.com/Da7_Tech/status/2123823183912378791
@Da7_Tech pushed back on the claim that Chinese labs distill only because they lack GPUs. Zai serves GLM-5.3 Flash on more than 100,000 domestic chips, around 62 trillion tokens in six days, and Alibaba spent $7.5 billion on infrastructure in one quarter with about $73 billion in cash, yet the biggest distillation campaign Anthropic detected, 151 million interactions with Claude between May and July, was attributed to operators tied to Alibaba. The post adds that one lab tried to distill Fable 5, switched to an older Claude and another lab's model when blocked, and that Kimi K3.1 Pro fell from 55 to 46 when Artificial Analysis raised its share of private tests from 45% to 60%. The sharpest line is about ceilings: the student is capped by the teacher, and the teacher belongs to someone who can close the door.
#15
@DZDZDZP
https://x.com/DZDZDZP/status/2123776332610089427
@DZDZDZP took the other side of the same argument. The popular story in Chinese AI circles is that domestic models went quiet after Claude's anti-distillation crackdown and account bans: Kimi has not updated, DeepSeek's flagship is not out, GLM 5.5 has no date and Qwen4-Max landed softly. But the post points out that Alibaba is the lab Anthropic named for the most distillation, so it should have been hit hardest, and yet it just shipped Qwen4 after spending 67.7 billion yuan on data centres and chips in Q2 2026. The conclusion is that distillation is a shortcut that gets you close but never ahead, and the real brake on Chinese labs is still compute.
https://x.com/DZDZDZP/status/2123776332610089427
@DZDZDZP took the other side of the same argument. The popular story in Chinese AI circles is that domestic models went quiet after Claude's anti-distillation crackdown and account bans: Kimi has not updated, DeepSeek's flagship is not out, GLM 5.5 has no date and Qwen4-Max landed softly. But the post points out that Alibaba is the lab Anthropic named for the most distillation, so it should have been hit hardest, and yet it just shipped Qwen4 after spending 67.7 billion yuan on data centres and chips in Q2 2026. The conclusion is that distillation is a shortcut that gets you close but never ahead, and the real brake on Chinese labs is still compute.
#16
@AdamHoltererer
https://x.com/AdamHoltererer/status/2123867842328092963
@AdamHoltererer made a short point with a long tail: even reasoning summaries can be used for distillation, which would explain why Opus's reasoning summaries read so oddly. If that is right, the strange, compressed summaries users complain about are partly a defence, made deliberately less useful as training data. It fits OpenAI's finding the same day that students trained on exact reasoning beat students trained on summaries, so labs have a reason to degrade what they show.
https://x.com/AdamHoltererer/status/2123867842328092963
@AdamHoltererer made a short point with a long tail: even reasoning summaries can be used for distillation, which would explain why Opus's reasoning summaries read so oddly. If that is right, the strange, compressed summaries users complain about are partly a defence, made deliberately less useful as training data. It fits OpenAI's finding the same day that students trained on exact reasoning beat students trained on summaries, so labs have a reason to degrade what they show.
#17
@wangxiaolong188
https://x.com/wangxiaolong188/status/2123603282971271576
@wangxiaolong188 wrote the most viral Chinese take, at about 340,000 views: the real reason Claude got tough on China is not ordinary users but Alibaba. The post repeats Anthropic's figures, 151 million distillation interactions by Alibaba in two months and about 24,000 fake accounts opened by DeepSeek, Kimi and others, and adds a claim that Claude banned 700,000 Chinese accounts, a number we could not confirm. Its practical advice is that normal use without resellers or distillation still works 99% of the time. Whatever the exact numbers, Chinese users now read account bans as fallout from lab-level distillation, not as something aimed at them.
https://x.com/wangxiaolong188/status/2123603282971271576
@wangxiaolong188 wrote the most viral Chinese take, at about 340,000 views: the real reason Claude got tough on China is not ordinary users but Alibaba. The post repeats Anthropic's figures, 151 million distillation interactions by Alibaba in two months and about 24,000 fake accounts opened by DeepSeek, Kimi and others, and adds a claim that Claude banned 700,000 Chinese accounts, a number we could not confirm. Its practical advice is that normal use without resellers or distillation still works 99% of the time. Whatever the exact numbers, Chinese users now read account bans as fallout from lab-level distillation, not as something aimed at them.
#18
@hocyzon
https://x.com/hocyzon/status/2123289962728833347
@hocyzon posted notes from a talk Yao Shunyu gave at Tsinghua, the window's biggest AI-for-AI post at nearly 890,000 views, and the background frame for everything above. The first claim is that AI for AI will push small models forward, because small models are cheap enough to run many experiments on: train one strong large model, then distill a hierarchy of sizes matched to tasks. On RSI the take is deliberately plain: it means using AI to run the R&D loop better, fixing bugs, cleaning data, running experiments, and the progress comes from people who solve concrete problems, not from new labels. The notes also warn that a small-scale failure does not mean the direction is wrong, citing an early ReAct attempt on Llama 13B that looked bad only because the model was too small.
https://x.com/hocyzon/status/2123289962728833347
@hocyzon posted notes from a talk Yao Shunyu gave at Tsinghua, the window's biggest AI-for-AI post at nearly 890,000 views, and the background frame for everything above. The first claim is that AI for AI will push small models forward, because small models are cheap enough to run many experiments on: train one strong large model, then distill a hierarchy of sizes matched to tasks. On RSI the take is deliberately plain: it means using AI to run the R&D loop better, fixing bugs, cleaning data, running experiments, and the progress comes from people who solve concrete problems, not from new labels. The notes also warn that a small-scale failure does not mean the direction is wrong, citing an early ReAct attempt on Llama 13B that looked bad only because the model was too small.
#19
@MTSlive
https://x.com/MTSlive/status/2123565906853031937
@MTSlive covered the protest outside Anthropic on IPO day, where organisers demanded a ban on recursive self-improvement and wanted AI's role in AI research capped at today's models, with the work done "by people, at a human pace". The numbers they cited came from Anthropic's own filing: Claude made more than 90% of contributions to its codebase in September and led about 26% of its AI R&D work by August. That is the background the whole beat sits on now. The same company whose models are being scored as RL teachers and research agents is telling investors those models already do a quarter of its research.
https://x.com/MTSlive/status/2123565906853031937
@MTSlive covered the protest outside Anthropic on IPO day, where organisers demanded a ban on recursive self-improvement and wanted AI's role in AI research capped at today's models, with the work done "by people, at a human pace". The numbers they cited came from Anthropic's own filing: Claude made more than 90% of contributions to its codebase in September and led about 26% of its AI R&D work by August. That is the background the whole beat sits on now. The same company whose models are being scored as RL teachers and research agents is telling investors those models already do a quarter of its research.
π‘ Eco Products Radar
Eco Products Radar
Claude Opus 5.5: the top scorer on the OpenResearch index and the model FrontierRSI calls best at improving AI, the base model of the Horizon agent, one of the weight-editing outer models, the reviewer in a GRPO run, and the open-track leader on RoboColosseum.
Claude Fable 5.2: the best RL teacher for Qwen 2.5-7B, the model that predicted its own training run to within 3%, the strongest protein-sequence predictor, and second on RoboColosseum's open track.
Qwen: the student in the editor-judge study (Qwen 2.5-7B), the inner model in the weight-editing study (Qwen3.5-0.8B), one of the small models in the index's post-training track (Qwen3-0.6B), and the model family at the centre of the Alibaba distillation story.
Gemini 4: Carbon won FrontierRSI's GPU kernel tasks, and Neon had every suggestion hold up in one GRPO run but hallucinated a whole training sweep as an RL teacher.
GPT-6 Astra / GPT-6.2 Astra: the RL teacher Fable beat, which leaked its scratchpad, the costliest fine-tuner on the index, and the model whose reasoning extraction fell from 54% to 0.2%.
Claude Opus 5.5: the top scorer on the OpenResearch index and the model FrontierRSI calls best at improving AI, the base model of the Horizon agent, one of the weight-editing outer models, the reviewer in a GRPO run, and the open-track leader on RoboColosseum.
Claude Fable 5.2: the best RL teacher for Qwen 2.5-7B, the model that predicted its own training run to within 3%, the strongest protein-sequence predictor, and second on RoboColosseum's open track.
Qwen: the student in the editor-judge study (Qwen 2.5-7B), the inner model in the weight-editing study (Qwen3.5-0.8B), one of the small models in the index's post-training track (Qwen3-0.6B), and the model family at the centre of the Alibaba distillation story.
Gemini 4: Carbon won FrontierRSI's GPU kernel tasks, and Neon had every suggestion hold up in one GRPO run but hallucinated a whole training sweep as an RL teacher.
GPT-6 Astra / GPT-6.2 Astra: the RL teacher Fable beat, which leaked its scratchpad, the costliest fine-tuner on the index, and the model whose reasoning extraction fell from 54% to 0.2%.
Comments