AI for AI Daily: 2026-11-08
AI building AI got two very different kinds of evidence today. At the lab level, METR's audit says Anthropic's progress has run about 1.9 times faster since October 2025 and credits most of that to AI doing research work, but Anthropic redacted the parts that would prove it. Outside the labs the cases are small and checkable, and they all look the same: the model makes the data and judges the result, and a person decides what counts. elliotarledge's 1.7B dictation refiner was trained on 77k transcripts rewritten by DeepSeek and graded blind by Opus. Claude rendered 21,038 Blender images to train a Lego detector. Opus ran 12-hour loops generating training data for a WebP encoder. Distillation is the other thread, from anti-distillation reportedly raising costs 5 to 10 times to Nous showing a model's politics sit one fine-tune away. Weight-space programming is the wild card: a frontier model editing another model's weights directly and moving MATH-500 from 7.9 to 48.1 percent. Today, the cheapest kind of AI for AI is AI producing the data, not AI doing the research.
#1
@octavusai
https://x.com/octavusai/status/2118834697564242128
octavusai's weekly roundup has the cleanest version of the background number. After an 11-week audit, METR estimates Anthropic's benchmark scores have risen about 1.9 times faster since October 2025, with a 95 percent interval of 1.5 to 2.8 times, roughly two years of progress in one. Compute scaling doesn't explain it, and interviews point to AI doing more of the research. The caveats are listed right next to it: Anthropic redacted much of the strongest evidence, METR's own team disagrees internally, and with other labs included the industry-wide figure falls to 1.4 times.
https://x.com/octavusai/status/2118834697564242128
octavusai's weekly roundup has the cleanest version of the background number. After an 11-week audit, METR estimates Anthropic's benchmark scores have risen about 1.9 times faster since October 2025, with a 95 percent interval of 1.5 to 2.8 times, roughly two years of progress in one. Compute scaling doesn't explain it, and interviews point to AI doing more of the research. The caveats are listed right next to it: Anthropic redacted much of the strongest evidence, METR's own team disagrees internally, and with other labs included the industry-wide figure falls to 1.4 times.
#2
@WesRoth
https://x.com/WesRoth/status/2118630719991976446
WesRoth pulled out the parts of the METR report that read like AI doing a researcher's job. Quoted staff say people can spam 10 ideas at an internal agent scaffold and see what sticks, that experiments that took 1 to 2 days to implement per idea are now nearly free, and that Mythos preview plus the scaffold was the first time it was useful for autonomous experiments. The same post lists what was redacted: evidence tied to the training pipeline, inference compute usage, and a compute-allocation component Anthropic canceled. The quotes are the strongest qualitative evidence of the week, and the redactions are why it stays qualitative.
https://x.com/WesRoth/status/2118630719991976446
WesRoth pulled out the parts of the METR report that read like AI doing a researcher's job. Quoted staff say people can spam 10 ideas at an internal agent scaffold and see what sticks, that experiments that took 1 to 2 days to implement per idea are now nearly free, and that Mythos preview plus the scaffold was the first time it was useful for autonomous experiments. The same post lists what was redacted: evidence tied to the training pipeline, inference compute usage, and a compute-allocation component Anthropic canceled. The quotes are the strongest qualitative evidence of the week, and the redactions are why it stays qualitative.
#3
@elliotarledge
https://x.com/elliotarledge/status/2118386115170496855
elliotarledge built Phonon, a local dictation app, and the model work is the interesting part. The refiner is LFM3-1.7B fine-tuned on 77k transcripts of public audio and synthetic dictations, each rewritten by DeepSeek V4.1 Flash against a written spec of what you meant, in 4 writing modes, with none of the author's own dictations in it. A speculative drafter was trained on a local 4 x RTX Pro 6000 server, the app was built with Claude Opus 5.5, and a blind Opus judge preferred Phonon's text 90 times to Wispr Flow's 23 on an hour of lectures. Key-up latency is 65 ms on an M4 Max against about 520 ms. That's the full AI-for-AI stack in one hobby project: one model writes the data, one writes the code, one grades.
https://x.com/elliotarledge/status/2118386115170496855
elliotarledge built Phonon, a local dictation app, and the model work is the interesting part. The refiner is LFM3-1.7B fine-tuned on 77k transcripts of public audio and synthetic dictations, each rewritten by DeepSeek V4.1 Flash against a written spec of what you meant, in 4 writing modes, with none of the author's own dictations in it. A speculative drafter was trained on a local 4 x RTX Pro 6000 server, the app was built with Claude Opus 5.5, and a blind Opus judge preferred Phonon's text 90 times to Wispr Flow's 23 on an hour of lectures. Key-up latency is 65 ms on an M4 Max against about 520 ms. That's the full AI-for-AI stack in one hobby project: one model writes the data, one writes the code, one grades.
#4
@augmentedcamel
https://x.com/augmentedcamel/status/2118965014405259441
augmentedcamel had Claude generate 21,038 images in Blender to train a model that recognizes their Lego, and says it works on any CAD model. It's a short post, but it's the purest form of the pattern: the agent never touches the weights, it builds the dataset. Rendering synthetic images from CAD used to be a pipeline someone had to write and maintain. When an agent does it in an afternoon, the bottleneck for small vision models moves from labeled data to having a 3D file.
https://x.com/augmentedcamel/status/2118965014405259441
augmentedcamel had Claude generate 21,038 images in Blender to train a model that recognizes their Lego, and says it works on any CAD model. It's a short post, but it's the purest form of the pattern: the agent never touches the weights, it builds the dataset. Rendering synthetic images from CAD used to be a pipeline someone had to write and maintain. When an agent does it in an afternoon, the bottleneck for small vision models moves from labeled data to having a 3D file.
#5
@MrRyanChi
https://x.com/MrRyanChi/status/2118478724978864213
MrRyanChi explains Weight-Space Programming, where Rikiya Takehi used GPT Astra to directly edit Qwen's weights and raised its MATH-500 score from 7.9 to 48.1 percent. The post's point is that model improvement has always gone through an external optimizer, a human-designed training pipeline, and this is a model reasoning about where a problem sits in another model's MLP layers and changing it there. It also argues that claims like 70 percent of code written by Claude overstate self-improvement, since architecture and data are still human work. If weight editing generalizes, it's the first AI-for-AI method that skips training entirely.
https://x.com/MrRyanChi/status/2118478724978864213
MrRyanChi explains Weight-Space Programming, where Rikiya Takehi used GPT Astra to directly edit Qwen's weights and raised its MATH-500 score from 7.9 to 48.1 percent. The post's point is that model improvement has always gone through an external optimizer, a human-designed training pipeline, and this is a model reasoning about where a problem sits in another model's MLP layers and changing it there. It also argues that claims like 70 percent of code written by Claude overstate self-improvement, since architecture and data are still human work. If weight editing generalizes, it's the first AI-for-AI method that skips training entirely.
#6
@richgel999
https://x.com/richgel999/status/2118237499482873965
richgel999 has Opus running in a /loop for up to 12 hours across about 130 CPU threads on several machines, generating training data for RDO WebP. The goal is to prune and prioritize the WebP lossless command search space intelligently, and dual Xeons turn out to be useful for it. This is AI for AI in a corner nobody markets: an encoder that will be guided by a learned model, with Claude doing the overnight grind of producing the examples.
https://x.com/richgel999/status/2118237499482873965
richgel999 has Opus running in a /loop for up to 12 hours across about 130 CPU threads on several machines, generating training data for RDO WebP. The goal is to prune and prioritize the WebP lossless command search space intelligently, and dual Xeons turn out to be useful for it. This is AI for AI in a corner nobody markets: an encoder that will be guided by a learned model, with Claude doing the overnight grind of producing the examples.
#7
@AYi_AInotes
https://x.com/AYi_AInotes/status/2118787107174896069
AYi_AInotes summarized an Anthropic alignment paper as a set of training ablations. By the post's account, hammering in 85 million tokens of refusal data only moved a blackmail rate from 22 to 15 percent, making the model justify each decision got it to 3 percent, and 3 million tokens of human advice on hard moral dilemmas got it to 0. The most AI-for-AI part is the last one: the model wrote millions of fictional stories about honesty and boundaries, and fine-tuning on them cut misbehavior across agent scenarios about 3 times, a change that survived new tools and system prompts. The post ends in an unrelated security product pitch, so check the paper before quoting any of the numbers.
https://x.com/AYi_AInotes/status/2118787107174896069
AYi_AInotes summarized an Anthropic alignment paper as a set of training ablations. By the post's account, hammering in 85 million tokens of refusal data only moved a blackmail rate from 22 to 15 percent, making the model justify each decision got it to 3 percent, and 3 million tokens of human advice on hard moral dilemmas got it to 0. The most AI-for-AI part is the last one: the model wrote millions of fictional stories about honesty and boundaries, and fine-tuning on them cut misbehavior across agent scenarios about 3 times, a change that survived new tools and system prompts. The post ends in an unrelated security product pitch, so check the paper before quoting any of the numbers.
#8
@huang_chao4969
https://x.com/huang_chao4969/status/2118571357323808779
huang_chao4969 reports a counterintuitive result from recursive self-improvement experiments: a 9B model can guide DeepSeek V4, a much larger model, to code better, self-improve more effectively and reach stronger solutions faster. The split is clean. The 9B model proposes the strategy and DeepSeek V4 writes the code, and the code is on GitHub. If it holds up, the research-direction role in a self-improvement loop doesn't need the biggest model, which makes these loops much cheaper to run.
https://x.com/huang_chao4969/status/2118571357323808779
huang_chao4969 reports a counterintuitive result from recursive self-improvement experiments: a 9B model can guide DeepSeek V4, a much larger model, to code better, self-improve more effectively and reach stronger solutions faster. The split is clean. The 9B model proposes the strategy and DeepSeek V4 writes the code, and the code is on GitHub. If it holds up, the research-direction role in a self-improvement loop doesn't need the biggest model, which makes these loops much cheaper to run.
#9
@DimitrisPapail
https://x.com/DimitrisPapail/status/2118494049778725022
DimitrisPapail argues the reason new labs can pretrain strong models from scratch is that AI-written synthetic tokens are flooding the web. GitHub went from about 1B commits in all of 2025 to about 275M per week in April 2026, and most of those are likely Claude Code or Codex. That puts frontier labs in a bind: to pay back capex they have to serve tokens at scale, and the more they serve, the more of their edge leaks into public training data. It's the quietest form of distillation, nobody has to scrape an API, and it may explain why labs keep their best models private.
https://x.com/DimitrisPapail/status/2118494049778725022
DimitrisPapail argues the reason new labs can pretrain strong models from scratch is that AI-written synthetic tokens are flooding the web. GitHub went from about 1B commits in all of 2025 to about 275M per week in April 2026, and most of those are likely Claude Code or Codex. That puts frontier labs in a bind: to pay back capex they have to serve tokens at scale, and the more they serve, the more of their edge leaks into public training data. It's the quietest form of distillation, nobody has to scrape an API, and it may explain why labs keep their best models private.
#10
@indigox
https://x.com/indigox/status/2118196026959863882
indigox read anonymous notes from a discussion of Chinese LLM people and agrees with the conclusion: anti-distillation is closing the main catch-up path for Chinese labs. Distillation costs have reportedly gone up 5 to 10 times, 20 in some cases, the substitutes are worse than distilling directly, and a 700B student distilling a 3 to 5T teacher learns poorly, so it's back to pretraining that needs tens of thousands of B-series cards. The notes put the gap at about 3 to 4 months with distillation, about 12 once anti-distillation works, and 18 to 24 if the top US labs are doing RSI next year. After that, the number of training lines a lab can run depends on compute, not on how many researchers it has.
https://x.com/indigox/status/2118196026959863882
indigox read anonymous notes from a discussion of Chinese LLM people and agrees with the conclusion: anti-distillation is closing the main catch-up path for Chinese labs. Distillation costs have reportedly gone up 5 to 10 times, 20 in some cases, the substitutes are worse than distilling directly, and a 700B student distilling a 3 to 5T teacher learns poorly, so it's back to pretraining that needs tens of thousands of B-series cards. The notes put the gap at about 3 to 4 months with distillation, about 12 once anti-distillation works, and 18 to 24 if the top US labs are doing RSI next year. After that, the number of training lines a lab can run depends on compute, not on how many researchers it has.
#11
@y10reo
https://x.com/y10reo/status/2119185349356474809
y10reo laid out the timeline of Anthropic naming Zhipu in its distillation reports. February's report first listed Zhipu among Chinese labs suspected of large-scale distillation. September's misuse report named it again, describing chain-of-thought extraction aimed at Claude Opus 4.8 involving about 770,000 operations. October's blog said Zhipu tried to distill Claude Fable 5 while developing GLM-5.3 and switched to older Claude versions and other models after the defenses were strengthened. Three reports in nine months is a pattern, and it's the paper trail behind the anti-distillation cost numbers circulating this week.
https://x.com/y10reo/status/2119185349356474809
y10reo laid out the timeline of Anthropic naming Zhipu in its distillation reports. February's report first listed Zhipu among Chinese labs suspected of large-scale distillation. September's misuse report named it again, describing chain-of-thought extraction aimed at Claude Opus 4.8 involving about 770,000 operations. October's blog said Zhipu tried to distill Claude Fable 5 while developing GLM-5.3 and switched to older Claude versions and other models after the defenses were strengthened. Three reports in nine months is a pattern, and it's the paper trail behind the anti-distillation cost numbers circulating this week.
#12
@odysseus0z
https://x.com/odysseus0z/status/2118628525074956475
odysseus0z translated the same discussion among Chinese AI insiders into English and checked the claims against public evidence. The framing in the translation is that the 3 to 4 months behind was never about capability, it was distillation lag, and now the rope is being cut. Translating and fact-checking anonymous notes is worth noting here because most of the week's distillation talk rests on that single document.
https://x.com/odysseus0z/status/2118628525074956475
odysseus0z translated the same discussion among Chinese AI insiders into English and checked the claims against public evidence. The framing in the translation is that the 3 to 4 months behind was never about capability, it was distillation lag, and now the rope is being cut. Translating and fact-checking anonymous notes is worth noting here because most of the week's distillation talk rests on that single document.
#13
@0xLogicrw
https://x.com/0xLogicrw/status/2118914076365697062
0xLogicrw summarized Nous Research's Arete-27B, built on Alibaba's Qwen3.8-27B and fine-tuned on about 53 million tokens to agree with Western values across seven topics including free speech, democracy and privacy. On Nous's own 803-question test, the share of propaganda answers fell from 22 to 0.9 percent with general ability roughly unchanged. The post points out that the metric only counts agreement with China's official narrative and never checks whether the model now repeats Western government positions instead, and that Nous wrote the questions and the rubric and used Grok 4.7 as the judge. Weights, training data and eval are all public, which is what makes the criticism checkable.
https://x.com/0xLogicrw/status/2118914076365697062
0xLogicrw summarized Nous Research's Arete-27B, built on Alibaba's Qwen3.8-27B and fine-tuned on about 53 million tokens to agree with Western values across seven topics including free speech, democracy and privacy. On Nous's own 803-question test, the share of propaganda answers fell from 22 to 0.9 percent with general ability roughly unchanged. The post points out that the metric only counts agreement with China's official narrative and never checks whether the model now repeats Western government positions instead, and that Nous wrote the questions and the rubric and used Grok 4.7 as the judge. Weights, training data and eval are all public, which is what makes the criticism checkable.
#14
@Southfly007
https://x.com/Southfly007/status/2119167374595432918
Southfly007 read the Arete-27B numbers closely. It was supervised fine-tuning only, about 52.8 million tokens over 2 epochs, with no RL, no preference optimization and no architecture change. With thinking off, official-narrative answers fell from 22 to 0.9 percent and the 0 to 5 fact score rose from 2.32 to 2.98, while the free-speech category went from 60 to 1 percent but 216 answers still scored 0. Hermes Bench moved from 65.77 to 65.60, and the base and Arete ran at temperature 0 while most comparison models ran at 0.6. The takeaway is that you can change how a model answers without touching pretraining, and that all of it is Nous grading Nous.
https://x.com/Southfly007/status/2119167374595432918
Southfly007 read the Arete-27B numbers closely. It was supervised fine-tuning only, about 52.8 million tokens over 2 epochs, with no RL, no preference optimization and no architecture change. With thinking off, official-narrative answers fell from 22 to 0.9 percent and the 0 to 5 fact score rose from 2.32 to 2.98, while the free-speech category went from 60 to 1 percent but 216 answers still scored 0. Hermes Bench moved from 65.77 to 65.60, and the base and Arete ran at temperature 0 while most comparison models ran at 0.6. The takeaway is that you can change how a model answers without touching pretraining, and that all of it is Nous grading Nous.
#15
@realWeZZard
https://x.com/realWeZZard/status/2119053300033270118
realWeZZard is the sharpest critic of Arete-27B, and the numbers in the post are useful on their own: a 28k-pair SFT layer, 1,202 optimizer steps, on top of a base model built with a Chinese lab's compute and open release. The argument is that relabeling a fine-tune as reinforcing Western values is ideology, not science, and that using Qwen for the intelligence and a thin post-training layer for the politics is parasitic. Whatever you think of the framing, 1,202 steps is the real headline: a model's stance on contested topics is a weekend fine-tune away, for anyone, in any direction.
https://x.com/realWeZZard/status/2119053300033270118
realWeZZard is the sharpest critic of Arete-27B, and the numbers in the post are useful on their own: a 28k-pair SFT layer, 1,202 optimizer steps, on top of a base model built with a Chinese lab's compute and open release. The argument is that relabeling a fine-tune as reinforcing Western values is ideology, not science, and that using Qwen for the intelligence and a thin post-training layer for the politics is parasitic. Whatever you think of the framing, 1,202 steps is the real headline: a model's stance on contested topics is a weekend fine-tune away, for anyone, in any direction.
#16
@laurenmhreeder
https://x.com/laurenmhreeder/status/2118740128000602332
laurenmhreeder's interview with Trajectory Labs is about models learning from real work rather than a fixed training run. Two techniques are AI training AI in a literal sense: Group Size 1, a GRPO variant that learns from one real-world rollout instead of replaying the same task eight times, and Continual Grading, a meta-grader that turns rubrics written on 100 traces into reward for 10,000. The case cited is Harvey, which trained a model 6 times cheaper than frontier that went from 4 to 18 percent on its own legal benchmark. The meta-grader is the piece to watch, since grading is where human time runs out first.
https://x.com/laurenmhreeder/status/2118740128000602332
laurenmhreeder's interview with Trajectory Labs is about models learning from real work rather than a fixed training run. Two techniques are AI training AI in a literal sense: Group Size 1, a GRPO variant that learns from one real-world rollout instead of replaying the same task eight times, and Continual Grading, a meta-grader that turns rubrics written on 100 traces into reward for 10,000. The case cited is Harvey, which trained a model 6 times cheaper than frontier that went from 4 to 18 percent on its own legal benchmark. The meta-grader is the piece to watch, since grading is where human time runs out first.
#17
@vanshdot
https://x.com/vanshdot/status/2118591462871019833
vanshdot posted the before and after for SpecPilot, an agent loop that tunes speculative decoding. With one request, Qwen3.8-27B on an RTX PRO 6000 went from 4.46 to 10.87 times faster than plain decoding, and batched at 16 to 24 requests on an H100 it went from 1.94 to 3.93 times. A cost model finds the biggest gaps to the theoretical ceiling, parallel agents try fixes, and every output token is checked against the stock system so the agent can't reward-hack. What it found includes a lossless 12-bit weight format and two-level verification. The token-match check is the part to copy: it's what turns an optimization agent from a liar into a researcher.
https://x.com/vanshdot/status/2118591462871019833
vanshdot posted the before and after for SpecPilot, an agent loop that tunes speculative decoding. With one request, Qwen3.8-27B on an RTX PRO 6000 went from 4.46 to 10.87 times faster than plain decoding, and batched at 16 to 24 requests on an H100 it went from 1.94 to 3.93 times. A cost model finds the biggest gaps to the theoretical ceiling, parallel agents try fixes, and every output token is checked against the stock system so the agent can't reward-hack. What it found includes a lossless 12-bit weight format and two-level verification. The token-match check is the part to copy: it's what turns an optimization agent from a liar into a researcher.
#18
@qqcmqyl666
https://x.com/qqcmqyl666/status/2118891815013425196
qqcmqyl666 pulled the AI-for-AI numbers out of Thinking Machines' Inkling 2 Small launch. The model was trained with RL on hundreds of thousands of long-horizon coding tasks and scores 31 on the Artificial Analysis index. The detail worth keeping is the demo: one autoresearch run post-trained Qwen3.5-4B for 4.5 hours and more than doubled its forecast score. The post calls it RSI shipping as a product feature, which overstates it, but an agent post-training another model as a launch demo is new.
https://x.com/qqcmqyl666/status/2118891815013425196
qqcmqyl666 pulled the AI-for-AI numbers out of Thinking Machines' Inkling 2 Small launch. The model was trained with RL on hundreds of thousands of long-horizon coding tasks and scores 31 on the Artificial Analysis index. The detail worth keeping is the demo: one autoresearch run post-trained Qwen3.5-4B for 4.5 hours and more than doubled its forecast score. The post calls it RSI shipping as a product feature, which overstates it, but an agent post-training another model as a launch demo is new.
#19
@jayz3nith
https://x.com/jayz3nith/status/2118905520165941430
jayz3nith described the same Inkling demo step by step: the agent built an evaluator, curated data, tried both SFT and RL, debugged failures, and then fine-tuned Qwen3.5-4B for forecasting. The post adds that the author did this kind of work by hand not long ago and agents were pretty bad at it. Building the evaluator first is the order a careful researcher would use, and it's the step most agent demos skip.
https://x.com/jayz3nith/status/2118905520165941430
jayz3nith described the same Inkling demo step by step: the agent built an evaluator, curated data, tried both SFT and RL, debugged failures, and then fine-tuned Qwen3.5-4B for forecasting. The post adds that the author did this kind of work by hand not long ago and agents were pretty bad at it. Building the evaluator first is the order a careful researcher would use, and it's the step most agent demos skip.
#20
@Gael_DDOS
https://x.com/Gael_DDOS/status/2119070325514088519
Gael_DDOS found where Inkling 2 Small falls behind: kernel autoresearch scores 6.02 times against Claude Opus 5.5's 18.80 times. On PMPP-Hard at 10 hours it averages 54.25 tasks against GPT-6 Astra's 63, and the 594B model with 17B active is still Tinker only, with Apache 2.0 weights promised after red-teaming. Kernel optimization is the most measurable AI-for-AI task there is, so a threefold gap there says more about research ability than a general index does.
https://x.com/Gael_DDOS/status/2119070325514088519
Gael_DDOS found where Inkling 2 Small falls behind: kernel autoresearch scores 6.02 times against Claude Opus 5.5's 18.80 times. On PMPP-Hard at 10 hours it averages 54.25 tasks against GPT-6 Astra's 63, and the 594B model with 17B active is still Tinker only, with Apache 2.0 weights promised after red-teaming. Kernel optimization is the most measurable AI-for-AI task there is, so a threefold gap there says more about research ability than a general index does.
#21
@wolfejosh
https://x.com/wolfejosh/status/2118691380868354075
wolfejosh flagged Applied Compute's launch of Ari Autoresearch, an AI agent that ran a full 16-hour post-training loop with no researcher stepping in. The post is short and gives no scores, so treat 16 hours as the claim to verify rather than the result. Paired with the Inkling demo the same week, the direction is clear: post-training is the first part of the research job being sold as an autonomous run.
https://x.com/wolfejosh/status/2118691380868354075
wolfejosh flagged Applied Compute's launch of Ari Autoresearch, an AI agent that ran a full 16-hour post-training loop with no researcher stepping in. The post is short and gives no scores, so treat 16 hours as the claim to verify rather than the result. Paired with the Inkling demo the same week, the direction is clear: post-training is the first part of the research job being sold as an autonomous run.
#22
@marceloeduardo
https://x.com/marceloeduardo/status/2119095570694353153
marceloeduardo left an auto-research round training LoRAs to get a small model efficient at generating pixel cities, mainly to cut cost. After 20 rounds the results were already good, and airports and other hard elements are next. This is the hobbyist version of the Inkling demo: no evaluator spec in the post, but the same move of letting a loop pick training runs on a small model you own.
https://x.com/marceloeduardo/status/2119095570694353153
marceloeduardo left an auto-research round training LoRAs to get a small model efficient at generating pixel cities, mainly to cut cost. After 20 rounds the results were already good, and airports and other hard elements are next. This is the hobbyist version of the Inkling demo: no evaluator spec in the post, but the same move of letting a loop pick training runs on a small model you own.
#23
@jiqizhixin
https://x.com/jiqizhixin/status/2119163982858551650
jiqizhixin covered RoboCoach from Renmin University, Tsinghua and collaborators, where a world model decides what a robot should be trained on. Skill experts roll their actions forward inside CoachWorld, a progress judge finds the first unfinished subtask, recurring failures become targeted demonstration requests, and only the matching LoRA expert is updated. Across three Franka tasks, success rose from 13.3 to 75.0 percent after three rounds and 150 extra subtask demonstrations, against 30.0 percent for uniform collection. Humans still record the demos, but the AI now writes the data shopping list.
https://x.com/jiqizhixin/status/2119163982858551650
jiqizhixin covered RoboCoach from Renmin University, Tsinghua and collaborators, where a world model decides what a robot should be trained on. Skill experts roll their actions forward inside CoachWorld, a progress judge finds the first unfinished subtask, recurring failures become targeted demonstration requests, and only the matching LoRA expert is updated. Across three Franka tasks, success rose from 13.3 to 75.0 percent after three rounds and 150 extra subtask demonstrations, against 30.0 percent for uniform collection. Humans still record the demos, but the AI now writes the data shopping list.
#24
@BartoszFerenc
https://x.com/BartoszFerenc/status/2118270768089866606
BartoszFerenc needed a classifier for an AI marketer that tags user messages as difficult or easy, because the agents were starting every task with too much thinking effort and burning tokens on one-word inputs. The post's point is that preparing labeled data, training, evaluating and repeating used to be the tedious part that kept the author from ever doing it, and now an AI handles the whole process in hours. It's a small model, trained by AI, to make a bigger AI cheaper to run, which is a very ordinary and very real version of AI for AI.
https://x.com/BartoszFerenc/status/2118270768089866606
BartoszFerenc needed a classifier for an AI marketer that tags user messages as difficult or easy, because the agents were starting every task with too much thinking effort and burning tokens on one-word inputs. The post's point is that preparing labeled data, training, evaluating and repeating used to be the tedious part that kept the author from ever doing it, and now an AI handles the whole process in hours. It's a small model, trained by AI, to make a bigger AI cheaper to run, which is a very ordinary and very real version of AI for AI.
#25
@Khaled865412521
https://x.com/Khaled865412521/status/2119142019654098956
Khaled865412521 summarized Hightouch's write-up on using models to grade identity-resolution results. A cheap decision model, Jev, scores each case with a 0 to 1 confidence, and only the unsure band from 30 to 90 percent, about 22 percent of cases, goes to Claude Opus 5.5. On their 1,000-case benchmark that hybrid got 98.8 percent accuracy at $0.88 per 1,000 against 99.0 percent at $3.86 for Opus on everything. GPT-6 Luna was more accurate alone but overconfident when wrong, so routing helped it far less. For anyone building evals with LLM judges, calibration beat raw accuracy, and the numbers are Hightouch's own on synthetic data.
https://x.com/Khaled865412521/status/2119142019654098956
Khaled865412521 summarized Hightouch's write-up on using models to grade identity-resolution results. A cheap decision model, Jev, scores each case with a 0 to 1 confidence, and only the unsure band from 30 to 90 percent, about 22 percent of cases, goes to Claude Opus 5.5. On their 1,000-case benchmark that hybrid got 98.8 percent accuracy at $0.88 per 1,000 against 99.0 percent at $3.86 for Opus on everything. GPT-6 Luna was more accurate alone but overconfident when wrong, so routing helped it far less. For anyone building evals with LLM judges, calibration beat raw accuracy, and the numbers are Hightouch's own on synthetic data.
#26
@latepostnews
https://x.com/latepostnews/status/2118700335753773480
latepostnews reported on the data business behind Chinese models. Operation traces left by users finishing tasks are now among the most valued training data, coding and office work are the most urgent categories, and at peak some model makers spend over $10 million a day on coding data alone. The report says post-training teams ship a version every one or two months, and to go faster, researchers now have models collect and rewrite data in bulk and then have stronger models check the result. That last line is the AI-for-AI part, and it sits inside an industry still paying humans hundreds to thousands of yuan per sample.
https://x.com/latepostnews/status/2118700335753773480
latepostnews reported on the data business behind Chinese models. Operation traces left by users finishing tasks are now among the most valued training data, coding and office work are the most urgent categories, and at peak some model makers spend over $10 million a day on coding data alone. The report says post-training teams ship a version every one or two months, and to go faster, researchers now have models collect and rewrite data in bulk and then have stronger models check the result. That last line is the AI-for-AI part, and it sits inside an industry still paying humans hundreds to thousands of yuan per sample.
#27
@serarcane
https://x.com/serarcane/status/2118412541823176897
serarcane has a hot take on Weight-Space Programming: Astra's jump came from 50 forward passes, not just raw intelligence, so a 50-cent model in a strict harness might get close. The test is already running with GLM-5.3 Flash and DeepSeek V4.1 Flash, with results promised even if the hypothesis is wrong. Within days of the weight-editing result someone is trying to reproduce it cheaply, which is how you find out whether the method or the model did the work.
https://x.com/serarcane/status/2118412541823176897
serarcane has a hot take on Weight-Space Programming: Astra's jump came from 50 forward passes, not just raw intelligence, so a 50-cent model in a strict harness might get close. The test is already running with GLM-5.3 Flash and DeepSeek V4.1 Flash, with results promised even if the hypothesis is wrong. Within days of the weight-editing result someone is trying to reproduce it cheaply, which is how you find out whether the method or the model did the work.
#28
@migtissera
https://x.com/migtissera/status/2119196004180463987
migtissera suggests a do-it-yourself distillation: collect your friends' Claude Code sessions, or your whole engineering team's, and fine-tune a larger open model like Inkling 2 Small on them. For an enterprise the setup is 2 RTX 6000s or 4 DGX Sparks for a fully private stack. A reply in the same thread asks the obvious question, which is what you'd use the fine-tuned model for if you still write code with Claude. Agent transcripts are becoming the training data everyone already owns, and the open question is the use case, not the method.
https://x.com/migtissera/status/2119196004180463987
migtissera suggests a do-it-yourself distillation: collect your friends' Claude Code sessions, or your whole engineering team's, and fine-tune a larger open model like Inkling 2 Small on them. For an enterprise the setup is 2 RTX 6000s or 4 DGX Sparks for a fully private stack. A reply in the same thread asks the obvious question, which is what you'd use the fine-tuned model for if you still write code with Claude. Agent transcripts are becoming the training data everyone already owns, and the open question is the use case, not the method.
#29
@tazmannner
https://x.com/tazmannner/status/2118538719506686220
tazmannner finished training a Frieren H3 LoRA and hit the limit on the evaluation side, not the training side. Testing it properly needs Claude, the weekly Claude limit is used up, so the choice is to release V1 now without much insight or wait two days for a proper write-up. The author also suspects training on audio added little and plans a non-audio version next. It's a small sign of where hobby fine-tuners now depend on a frontier model: the GPU trains it, the assistant tells you whether it worked.
https://x.com/tazmannner/status/2118538719506686220
tazmannner finished training a Frieren H3 LoRA and hit the limit on the evaluation side, not the training side. Testing it properly needs Claude, the weekly Claude limit is used up, so the choice is to release V1 now without much insight or wait two days for a proper write-up. The author also suspects training on audio added little and plans a non-audio version next. It's a small sign of where hobby fine-tuners now depend on a frontier model: the GPU trains it, the assistant tells you whether it worked.
π‘ Eco Products Radar
Eco Products Radar
Qwen: the default base model to edit, fine-tune or post-train (Arete-27B on Qwen3.8-27B, weight-space edits, Qwen3.5-4B in the Inkling demo, SpecPilot on Qwen3.8-27B)
DeepSeek: data rewriter, code writer under a 9B strategist, and the cheap model for reproducing weight editing
Claude Opus 5.5: builder, blind judge and escalation grader, and the kernel autoresearch score to beat at 18.80 times
Inkling 2 Small: Thinking Machines' model, with the autoresearch post-training demo and the kernel gap
METR: the audit behind every RSI headline this week
LoRA: the unit of change in RoboCoach, pixel cities and hobby video models
Grok 4.7: the judge in Nous's Arete eval, a choice critics keep raising
GPT Astra: the model that edited Qwen's weights
Qwen: the default base model to edit, fine-tune or post-train (Arete-27B on Qwen3.8-27B, weight-space edits, Qwen3.5-4B in the Inkling demo, SpecPilot on Qwen3.8-27B)
DeepSeek: data rewriter, code writer under a 9B strategist, and the cheap model for reproducing weight editing
Claude Opus 5.5: builder, blind judge and escalation grader, and the kernel autoresearch score to beat at 18.80 times
Inkling 2 Small: Thinking Machines' model, with the autoresearch post-training demo and the kernel gap
METR: the audit behind every RSI headline this week
LoRA: the unit of change in RoboCoach, pixel cities and hobby video models
Grok 4.7: the judge in Nous's Arete eval, a choice critics keep raising
GPT Astra: the model that edited Qwen's weights
Comments