Loop Daily: 2026-10-20
Three results landed this window that cut against the autoresearch story everyone tells. A NeurIPS framework ran six search algorithms over 3,222 generated research ideas and found zero completely original ones, and only one novel idea reached its task's top ten, because agents still cannot tell which new idea is worth refining. A pre-registered study of 30,687 scored submissions across ten open leaderboards found that building on a parent made with a different vendor's model changes your odds of promotion by about 2 percent, which is to say not at all; two thirds of builds already switch vendor, and the swarm's edge looks like coordination, not model mixing. And a researcher returning after a few months found Claude still drifts into local hyperparameter tuning, just later than it did in April. On the build side the sharpest post was a software factory where a run has 22 steps and only three belong to the agent, with credentials fetched by code, rules frozen at the starting commit, two CI runs maximum and a deleted test ending the run. A security agent at OpenRouter has been improving itself for months with a 0.05 percent action reversal rate. Google's Harness-squared, which rewrites prompts, tools and memory without touching weights, got read three ways: a 64.2 to 75.7 jump on GLM-5, a reminder that every task runs three times, and an underwriter's objection that a harness cannot grade its own memo when there is no ground truth. A Gemini 4 Argon loop built a game 20,000 people have played while its author worked a full day, a 35B local model scored 52 against another's 91 on change-of-plan recovery, and a RAM-only Qwen was too dumb to run an autoresearch loop at all.
#1
@anton_iades
https://x.com/anton_iades/status/2112226980095512843
@anton_iades asked whether search algorithms can help autoresearch agents produce original ideas that work, and ran six of them on Heuresis, a NeurIPS 2026 framework built to study autoresearch agents. Across 3,222 ideas, zero were completely original. Search did beat the greedy baseline at producing minor-similarity ideas, the most novel rating any idea got: Curiosity 20 against Greedy 0 on nanoGPT, MAP-Elites 12 against 0 on on-policy RL, OMNI 29 against 12 on model unlearning. The catch is that only one of these novel ideas reached its task's top ten, so agents cannot yet tell which new idea deserves refinement. Heuresis ships plug-and-play search abstractions, skills that analyze runs for quality, diversity and novelty, and a HackerJudge that catches reward hacking, and the same loop generates training data for better autoresearch models.
https://x.com/anton_iades/status/2112226980095512843
@anton_iades asked whether search algorithms can help autoresearch agents produce original ideas that work, and ran six of them on Heuresis, a NeurIPS 2026 framework built to study autoresearch agents. Across 3,222 ideas, zero were completely original. Search did beat the greedy baseline at producing minor-similarity ideas, the most novel rating any idea got: Curiosity 20 against Greedy 0 on nanoGPT, MAP-Elites 12 against 0 on on-policy RL, OMNI 29 against 12 on model unlearning. The catch is that only one of these novel ideas reached its task's top ten, so agents cannot yet tell which new idea deserves refinement. Heuresis ships plug-and-play search abstractions, skills that analyze runs for quality, diversity and novelty, and a HackerJudge that catches reward hacking, and the same loop generates training data for better autoresearch models.
#2
@oxwizzdom
https://x.com/oxwizzdom/status/2112062165356093860
@oxwizzdom released the study promised the night before on whether cross-model pollination helps open autoresearch, and the answer is no. The author froze the public Yukon API for ten leaderboards, 30,687 scored submissions including the integer-multiplication board, recorded the company behind each declared model and which earlier submission each one built on, and timestamped the method with OpenTimestamps before joining outcomes to data. Cross-company building is the normal case: two thirds of builds on someone else's work switch vendor, mostly between Anthropic and OpenAI models in both directions. Raw rates even lean the other way, 6.1 percent promoted for cross-company builds against 7.6 for same-company, but comparing each participant with themselves at the same stage gives about 2 percent higher odds for cross-company, with the data ruling out anything beyond roughly 18 percent lower or 28 percent higher. The famous 843-qubit circuit really was a cross-model step, GPT-5 building on a Claude Opus 5.5 circuit, but on average vendor mix is not what separates winning steps from losing ones; coordination on one public frontier is the better explanation.
https://x.com/oxwizzdom/status/2112062165356093860
@oxwizzdom released the study promised the night before on whether cross-model pollination helps open autoresearch, and the answer is no. The author froze the public Yukon API for ten leaderboards, 30,687 scored submissions including the integer-multiplication board, recorded the company behind each declared model and which earlier submission each one built on, and timestamped the method with OpenTimestamps before joining outcomes to data. Cross-company building is the normal case: two thirds of builds on someone else's work switch vendor, mostly between Anthropic and OpenAI models in both directions. Raw rates even lean the other way, 6.1 percent promoted for cross-company builds against 7.6 for same-company, but comparing each participant with themselves at the same stage gives about 2 percent higher odds for cross-company, with the data ruling out anything beyond roughly 18 percent lower or 28 percent higher. The famous 843-qubit circuit really was a cross-model step, GPT-5 building on a Claude Opus 5.5 circuit, but on average vendor mix is not what separates winning steps from losing ones; coordination on one public frontier is the better explanation.
#3
@mardehaym
https://x.com/mardehaym/status/2112099730641555505
@mardehaym does not trust a model to follow rules because they are in a prompt; if a rule matters it goes in code the agent cannot change. In Velocity Core, which runs agents inside their software factory, a run has 22 steps, 14 of them plain code and only three left entirely to the agent. Credentials stay outside the agent, with code fetching the repo before it starts and pushing after. The agent cannot change its own rules, which are read from the starting commit and need a person's approval to change. Outbound requests pass an allowlist and secret scanning covers code, logs and PR text. Done is not proof, CI decides, and the agent gets at most two CI runs. Deleting or emptying a test sends the work back, and a second violation ends the run. Every run records prompt versions, exact model per step, starting commit and cost, and a person still merges every PR. The agent loop was the easy part; the platform that enforces the boundaries on every run is where the author says the advantage lives.
https://x.com/mardehaym/status/2112099730641555505
@mardehaym does not trust a model to follow rules because they are in a prompt; if a rule matters it goes in code the agent cannot change. In Velocity Core, which runs agents inside their software factory, a run has 22 steps, 14 of them plain code and only three left entirely to the agent. Credentials stay outside the agent, with code fetching the repo before it starts and pushing after. The agent cannot change its own rules, which are read from the starting commit and need a person's approval to change. Outbound requests pass an allowlist and secret scanning covers code, logs and PR text. Done is not proof, CI decides, and the agent gets at most two CI runs. Deleting or emptying a test sends the work back, and a second violation ends the run. Every run records prompt versions, exact model per step, starting commit and cost, and a person still merges every PR. The agent loop was the easy part; the platform that enforces the boundaries on every run is where the author says the advantage lives.
#4
@alexatallah
https://x.com/alexatallah/status/2112058319410254067
@alexatallah described Sentinel, a self-improving security agent built a few months ago to protect OpenRouter users. It runs around the clock across all activity looking for anomalies, signs of end-user account takeover and user key theft, and keeps improving from the trust and safety team's signals; this month its action reversal rate, roughly its false positive rate, was 0.05 percent. Any app can inherit its protection by implementing OpenRouter OAuth, and the post doubles as a hiring call for what the author calls real autoresearch in action.
https://x.com/alexatallah/status/2112058319410254067
@alexatallah described Sentinel, a self-improving security agent built a few months ago to protect OpenRouter users. It runs around the clock across all activity looking for anomalies, signs of end-user account takeover and user key theft, and keeps improving from the trust and safety team's signals; this month its action reversal rate, roughly its false positive rate, was 0.05 percent. Any app can inherit its protection by implementing OpenRouter OAuth, and the post doubles as a hiring call for what the author calls real autoresearch in action.
#5
@ammaar
https://x.com/ammaar/status/2112274521671831824
@ammaar reports that over 20,000 people have played the game Gemini 4 made since Friday, and the point is how it was made. With an idea right before a full workday, the author fired off an agent loop with Gemini 4 Argon and went to work. For two hours in the background it inspected the Tesla Roadster site for styles, copy and car details, looked up Subway Surfers references, watched the Roadster launch on YouTube to understand the thrusters, critiqued and iterated on its own work several times, and generated music with Lyria and commentary with Gemini TTS, pinging the author to try it between breaks. About ten minutes of prompting for control and camera tweaks preceded publishing in AI Studio.
https://x.com/ammaar/status/2112274521671831824
@ammaar reports that over 20,000 people have played the game Gemini 4 made since Friday, and the point is how it was made. With an idea right before a full workday, the author fired off an agent loop with Gemini 4 Argon and went to work. For two hours in the background it inspected the Tesla Roadster site for styles, copy and car details, looked up Subway Surfers references, watched the Roadster launch on YouTube to understand the thrusters, critiqued and iterated on its own work several times, and generated music with Lyria and commentary with Gemini TTS, pinging the author to try it between breaks. About ten minutes of prompting for control and camera tweaks preceded publishing in AI Studio.
#6
@sydneyrunkle
https://x.com/sydneyrunkle/status/2112256862318194818
@sydneyrunkle distilled a conversation with LangChain's founder on why agents are so hard to build. An agent is model plus context plus harness and can be improved at any of the three; the highest-return move is context, then harness, then weights, because that is where domain knowledge lives. Continual learning is a loop: run the agent, collect traces, aggregate learnings, fix something in any layer, measure, repeat. Memory is massively unsolved and the hardest part is distilling learnings from traces, which is domain specific. Cost reduction goes visibility, then controls, then agent optimization, which is how LangChain cut its own coding-agent spend, and model routing belongs in the harness because that is where the context for the routing decision lives. Open models currently power narrow parts of the loop, and horizontal agents will call vertical agents at a rate humans cannot.
https://x.com/sydneyrunkle/status/2112256862318194818
@sydneyrunkle distilled a conversation with LangChain's founder on why agents are so hard to build. An agent is model plus context plus harness and can be improved at any of the three; the highest-return move is context, then harness, then weights, because that is where domain knowledge lives. Continual learning is a loop: run the agent, collect traces, aggregate learnings, fix something in any layer, measure, repeat. Memory is massively unsolved and the hardest part is distilling learnings from traces, which is domain specific. Cost reduction goes visibility, then controls, then agent optimization, which is how LangChain cut its own coding-agent spend, and model routing belongs in the harness because that is where the context for the routing decision lives. Open models currently power narrow parts of the loop, and horizontal agents will call vertical agents at a rate humans cannot.
#7
@ruhzi57
https://x.com/ruhzi57/status/2112272341845835931
@ruhzi57 is testing whether the inference-level intervention built for ForesightBench, which decomposes a setup into threads and sub-threads, actually improves autoresearch. The experiment tasks the custom harness and a vanilla out-of-the-box Pi with building a long-horizon agent, two seeds each, four long-horizon agents in total, with the retrieval pipeline and the repo containing both the custom harness and the agents the two Pis built published alongside. The author's framing in a companion post: autoresearch loops are brute-forcey, and what is needed are loops that reason over the results of research directions before committing to implementations, which is what ForesightBench measures.
https://x.com/ruhzi57/status/2112272341845835931
@ruhzi57 is testing whether the inference-level intervention built for ForesightBench, which decomposes a setup into threads and sub-threads, actually improves autoresearch. The experiment tasks the custom harness and a vanilla out-of-the-box Pi with building a long-horizon agent, two seeds each, four long-horizon agents in total, with the retrieval pipeline and the repo containing both the custom harness and the agents the two Pis built published alongside. The author's framing in a companion post: autoresearch loops are brute-forcey, and what is needed are loops that reason over the results of research directions before committing to implementations, which is what ForesightBench measures.
#8
@braaannigan
https://x.com/braaannigan/status/2112123285685350783
@braaannigan is back to autoresearch experiments after a few months off and reports a partial improvement: Claude no longer drifts immediately into local hyperparameter optimisation the way it did in April, but given enough time it still ends up there.
https://x.com/braaannigan/status/2112123285685350783
@braaannigan is back to autoresearch experiments after a few months off and reports a partial improvement: Claude no longer drifts immediately into local hyperparameter optimisation the way it did in April, but given enough time it still ends up there.
#9
@davidpgil
https://x.com/davidpgil/status/2112227480904040877
@davidpgil tried to have Qwen 3.6 35b:a3b run an autoresearch loop entirely from RAM, with the goal of getting DeepSeek V4.1 Flash running fast on two GPUs. It failed; the model was too dumb for the job. Nemotron is next, and the author is asking whether anyone else is benchmarking with RAM-only models.
https://x.com/davidpgil/status/2112227480904040877
@davidpgil tried to have Qwen 3.6 35b:a3b run an autoresearch loop entirely from RAM, with the goal of getting DeepSeek V4.1 Flash running fast on two GPUs. It failed; the model was too dumb for the job. Nemotron is next, and the author is asking whether anyone else is benchmarking with RAM-only models.
#10
@PaulGugAI
https://x.com/PaulGugAI/status/2112140201116217592
@PaulGugAI measured the gap between two 35B-A3B models at the same IQ3_M quant on an RTX 3080 with 10 GB and 50K context: a Qwen3.8 distill scored 52 on a change-of-plan recovery test and OpenBMB's MiniCPM-V 4.7 scored 91, while also running faster at 551 versus 498 tokens per second on prompt and 24.5 versus 21.8 on generation. Aggressive quants and distilled MoEs handle chat, basic tool calls and light agent work fine on 8 to 12 GB cards, but in a demanding agent loop the gap is the difference between trusting an agent with a busy, changing schedule or not.
https://x.com/PaulGugAI/status/2112140201116217592
@PaulGugAI measured the gap between two 35B-A3B models at the same IQ3_M quant on an RTX 3080 with 10 GB and 50K context: a Qwen3.8 distill scored 52 on a change-of-plan recovery test and OpenBMB's MiniCPM-V 4.7 scored 91, while also running faster at 551 versus 498 tokens per second on prompt and 24.5 versus 21.8 on generation. Aggressive quants and distilled MoEs handle chat, basic tool calls and light agent work fine on 8 to 12 GB cards, but in a demanding agent loop the gap is the difference between trusting an agent with a busy, changing schedule or not.
#11
@naqerl
https://x.com/naqerl/status/2111955807231193396
@naqerl solved a client's recurring problem, sharing task board status with non-technical staff who do not speak English, with a thin harness that bridges Telegram messages into a pi-durable agent. The agent searches related issues in the git forge and surfaces a link to a staff-facing board with a mirrored, human-readable version of the task and a short report. The flow can break at almost every step: context has to accumulate correctly across replies, forwards and attached media, then comes the key classification of whether the message is even meant for the bot, and only then the agent loop with read-only code access, Sentry and uptime data, and credentials for the forge and the staff CRM. Compaction, persistence, subagents, tools and MCPs were already handled by pi-durable, so a proper agent that took a long time to assemble a year ago now comes together in an evening.
https://x.com/naqerl/status/2111955807231193396
@naqerl solved a client's recurring problem, sharing task board status with non-technical staff who do not speak English, with a thin harness that bridges Telegram messages into a pi-durable agent. The agent searches related issues in the git forge and surfaces a link to a staff-facing board with a mirrored, human-readable version of the task and a short report. The flow can break at almost every step: context has to accumulate correctly across replies, forwards and attached media, then comes the key classification of whether the message is even meant for the bot, and only then the agent loop with read-only code access, Sentry and uptime data, and credentials for the forge and the staff CRM. Compaction, persistence, subagents, tools and MCPs were already handled by pi-durable, so a proper agent that took a long time to assemble a year ago now comes together in an evening.
#12
@ManlingLi_
https://x.com/ManlingLi_/status/2112307226224292176
@ManlingLi_ introduced Embodied First Exam, one line of code to run coding agents across 1,090 embodied tasks, 2,128 variants, 31 simulation suites, 44 robots and 8 robot types, including 382 brand-new tasks in eight original suites such as RoboOlympics, DexHandExam and SafetyExam plus hard picks from 23 community benchmarks. Standard mode is camera-only, partially observed, no resets, real time for dynamic tasks; privileged mode gives raw simulator access. Coding agents can act three ways: direct control as policy, code as policy, or train a policy to control, which the author labels autoresearch with sim2sim. It runs through harbor with your own harness and model.
https://x.com/ManlingLi_/status/2112307226224292176
@ManlingLi_ introduced Embodied First Exam, one line of code to run coding agents across 1,090 embodied tasks, 2,128 variants, 31 simulation suites, 44 robots and 8 robot types, including 382 brand-new tasks in eight original suites such as RoboOlympics, DexHandExam and SafetyExam plus hard picks from 23 community benchmarks. Standard mode is camera-only, partially observed, no resets, real time for dynamic tasks; privileged mode gives raw simulator access. Coding agents can act three ways: direct control as policy, code as policy, or train a policy to control, which the author labels autoresearch with sim2sim. It runs through harbor with your own harness and model.
#13
@jindong_wang92
https://x.com/jindong_wang92/status/2112278762964271580
@jindong_wang92 published Science Utopia, an LLM-agent framework that simulates an academic ecosystem where everyone has top models for paper writing, reviewing and grant writing. The question behind it predates the autoresearch wave: does a Matthew effect appear, where people with more API credits write more and gain more? The paper, website and code are out, with several findings the author says are worth reading.
https://x.com/jindong_wang92/status/2112278762964271580
@jindong_wang92 published Science Utopia, an LLM-agent framework that simulates an academic ecosystem where everyone has top models for paper writing, reviewing and grant writing. The question behind it predates the autoresearch wave: does a Matthew effect appear, where people with more API credits write more and gain more? The paper, website and code are out, with several findings the author says are worth reading.
#14
@pifagor_dao
https://x.com/pifagor_dao/status/2112236875465302077
@pifagor_dao surfaced ARIS, Auto-Research-In-Sleep, a set of 85 Markdown skills that do ML research overnight: Claude runs the experiments, a second model reviews them through Codex, and the loop fixes whatever the reviewer flags. The README's real overnight run shows four review rounds with no human in between, more than 20 GPU experiments launched by the loop itself, an AI reviewer score rising from 5.0 to 7.5 out of 10, and the loop dropping its own main claim when a seed study did not back it. The repo sits at 17.7K stars under MIT, the author found no X post linking it, and checked that nothing runs on install; you bring your own GPU server and Codex login.
https://x.com/pifagor_dao/status/2112236875465302077
@pifagor_dao surfaced ARIS, Auto-Research-In-Sleep, a set of 85 Markdown skills that do ML research overnight: Claude runs the experiments, a second model reviews them through Codex, and the loop fixes whatever the reviewer flags. The README's real overnight run shows four review rounds with no human in between, more than 20 GPU experiments launched by the loop itself, an AI reviewer score rising from 5.0 to 7.5 out of 10, and the loop dropping its own main claim when a seed study did not back it. The repo sits at 17.7K stars under MIT, the author found no X post linking it, and checked that nothing runs on install; you bring your own GPU server and Codex login.
#15
@maixbt
https://x.com/maixbt/status/2112166928856789074
@maixbt read Google's Harness-squared release and pulled out the parts people skip. The default editor that does the improving is Claude Opus 4.8. It never touches weights; after every task it rewrites everything around them, prompt, tools, memory and skills, and GLM-5 went from 64.2 to 75.7 after one rewrite. The loop runs the task once as a reference, tries two rewrites of its own setup, keeps what the runs support without a grader, and saves lessons to EXPERIENCE.md for the next task. The catch is that every task runs three times and Google does not say what that costs, and stock Gemini lost to GLM-5 wearing a better setup in Google's own table.
https://x.com/maixbt/status/2112166928856789074
@maixbt read Google's Harness-squared release and pulled out the parts people skip. The default editor that does the improving is Claude Opus 4.8. It never touches weights; after every task it rewrites everything around them, prompt, tools, memory and skills, and GLM-5 went from 64.2 to 75.7 after one rewrite. The loop runs the task once as a reference, tries two rewrites of its own setup, keeps what the runs support without a grader, and saves lessons to EXPERIENCE.md for the next task. The catch is that every task runs three times and Google does not say what that costs, and stock Gemini lost to GLM-5 wearing a better setup in Google's own table.
#16
@tobias_pfuetze
https://x.com/tobias_pfuetze/status/2112276225833996601
@tobias_pfuetze adds the underwriter's objection to Harness-squared: it tunes prompts, tools and memory with no ground-truth feedback, which is fine for coding tasks where tests exist, but in credit the harness also has to decide who says the file is done, and a self-improving agent cannot grade its own memo. The author wrote up the six harness decisions their team makes for underwriting agents.
https://x.com/tobias_pfuetze/status/2112276225833996601
@tobias_pfuetze adds the underwriter's objection to Harness-squared: it tunes prompts, tools and memory with no ground-truth feedback, which is fine for coding tasks where tests exist, but in credit the harness also has to decide who says the file is done, and a self-improving agent cannot grade its own memo. The author wrote up the six harness decisions their team makes for underwriting agents.
#17
@yusufksadek
https://x.com/yusufksadek/status/2112227964326973876
@yusufksadek points to a detail in Table 1 of the Cryptids paper: 7 of the 11 new Cryptids are starred, meaning they came from the second stage, a general auto-research agent running on Gemini 3.8 Flash, rather than from the main Gemini 3.1 Pro pipeline.
https://x.com/yusufksadek/status/2112227964326973876
@yusufksadek points to a detail in Table 1 of the Cryptids paper: 7 of the 11 new Cryptids are starred, meaning they came from the second stage, a general auto-research agent running on Gemini 3.8 Flash, rather than from the main Gemini 3.1 Pro pipeline.
#18
@AgentScaleAI
https://x.com/AgentScaleAI/status/2112249929678668216
@AgentScaleAI tested re-injecting state into an agent loop every 3, 5 and 10 messages. At 5 the output was already drifting on section headers, at 10 it was basically freestyling, so they settled on every 3. Shorter loops cost more tokens, but the output stays clean.
https://x.com/AgentScaleAI/status/2112249929678668216
@AgentScaleAI tested re-injecting state into an agent loop every 3, 5 and 10 messages. At 5 the output was already drifting on section headers, at 10 it was basically freestyling, so they settled on every 3. Shorter loops cost more tokens, but the output stays clean.
#19
@RohanArun
https://x.com/RohanArun/status/2111629446604034089
@RohanArun posted an interactive widget that shows the integer-multiplication bounds updating in real time on the timeline, and notes the leaderboard now receives a PR a minute. Math as a spectator sport.
https://x.com/RohanArun/status/2111629446604034089
@RohanArun posted an interactive widget that shows the integer-multiplication bounds updating in real time on the timeline, and notes the leaderboard now receives a PR a minute. Math as a spectator sport.
#20
@thenomadevel
https://x.com/thenomadevel/status/2112226381161721916
@thenomadevel moved yes/no answers in their agent off frontier pricing onto Jev plus a small reader model and reports 12x cheaper per answer, latency down from 13.4 to 7.3 seconds, and 50 dollars now covering 100K answers instead of 8.3K, with a write-up of Jev, Laya and Glimpse and what changed.
https://x.com/thenomadevel/status/2112226381161721916
@thenomadevel moved yes/no answers in their agent off frontier pricing onto Jev plus a small reader model and reports 12x cheaper per answer, latency down from 13.4 to 7.3 seconds, and 50 dollars now covering 100K answers instead of 8.3K, with a write-up of Jev, Laya and Glimpse and what changed.
#21
@csehra42
https://x.com/csehra42/status/2112231163716948104
@csehra42 flagged a gotcha for agent loops on MiniMax M3.1 Flash, out at 0.15 and 0.60 dollars per million with weights due October 23: it thinks on demand, so simple tool-call turns can come back with no reasoning_content field at all. Any loop that assumes the field is always there needs fixing first.
https://x.com/csehra42/status/2112231163716948104
@csehra42 flagged a gotcha for agent loops on MiniMax M3.1 Flash, out at 0.15 and 0.60 dollars per million with weights due October 23: it thinks on demand, so simple tool-call turns can come back with no reasoning_content field at all. Any loop that assumes the field is always there needs fixing first.
#22
@atlashealthHQ
https://x.com/atlashealthHQ/status/2111830001930879039
@atlashealthHQ connects the forgetting result to a limit: forgetting helps autoresearch agents because a verifier picks what survives, drop every session and past candidate, restart from the best so far, per arXiv 2610.20563. Memory of a person has no verifier, so there the reset rule is the model itself deciding which past still counts.
https://x.com/atlashealthHQ/status/2111830001930879039
@atlashealthHQ connects the forgetting result to a limit: forgetting helps autoresearch agents because a verifier picks what survives, drop every session and past candidate, restart from the best so far, per arXiv 2610.20563. Memory of a person has no verifier, so there the reset rule is the model itself deciding which past still counts.
#23
@chang_trenton
https://x.com/chang_trenton/status/2112191802736336917
@chang_trenton, whose PhD straddled ChatGPT's arrival, argues that autoresearch can boost discovery but in its worst form holds research back, because research is not results and the field has Goodharted itself into believing it is. The piece with Quadrillion Labs is about steering toward an abundance of discovery rather than an abundance of papers.
https://x.com/chang_trenton/status/2112191802736336917
@chang_trenton, whose PhD straddled ChatGPT's arrival, argues that autoresearch can boost discovery but in its worst form holds research back, because research is not results and the field has Goodharted itself into believing it is. The piece with Quadrillion Labs is about steering toward an abundance of discovery rather than an abundance of papers.
#24
@lifeisameeme
https://x.com/lifeisameeme/status/2112291421978001642
@lifeisameeme offers a four-question test before wiring up an agent loop: does the task recur weekly or more, can a test, linter or build fail it automatically, can you eat the retries, and can the agent run code and read logs. Miss one and a hand-written prompt is faster. The budget question bites: Uber reportedly capped engineers at 1,500 dollars per person per tool per month after burning its annual AI budget in four months, and without an automated check you are back to reading every diff with a bigger bill.
https://x.com/lifeisameeme/status/2112291421978001642
@lifeisameeme offers a four-question test before wiring up an agent loop: does the task recur weekly or more, can a test, linter or build fail it automatically, can you eat the retries, and can the agent run code and read logs. Miss one and a hand-written prompt is faster. The budget question bites: Uber reportedly capped engineers at 1,500 dollars per person per tool per month after burning its annual AI budget in four months, and without an automated check you are back to reading every diff with a bigger bill.
π‘ Eco Products Radar
Eco Products Radar
Harness-squared: Google's open-source loop that rewrites prompts, tools and memory without touching weights, read as a 64.2 to 75.7 jump on GLM-5, a three-runs-per-task cost nobody quoted, and a harness that cannot grade its own memo.
Claude Code and Codex: the worker and the reviewer in ARIS and in most overnight loops, and the harness that scored DeepSeek's new open model.
DeepSeek V4.1 Flash 1017: the model everyone is routing cheap loop steps to, with Search and Fetch free inside agent tools until the end of the month.
Pi and pi-durable: the baseline harness in the ForesightBench experiment and the runtime behind the Telegram task-board agent.
Gemini 4 Argon: the two-hour background loop that built a game 20,000 people played.
Jev: the small decision model taking yes/no steps out of frontier pricing.
MiniMax M3.1 Flash: 0.15 and 0.60 dollars per million, weights due October 23, and a missing reasoning_content field that breaks naive loops.
MiniCPM-V 4.7 and Qwen: the 91 and the 52 on change-of-plan recovery at the same quant, and a RAM-only loop that failed outright.
harbor: the runner behind Embodied First Exam and the fleet repo lists.
Yukon leaderboards: 30,687 scored submissions behind the cross-model pollination study, with the integer-multiplication board still taking a PR a minute.
LangChain: the model, context, harness framing and the visibility-first cost pipeline.
Harness-squared: Google's open-source loop that rewrites prompts, tools and memory without touching weights, read as a 64.2 to 75.7 jump on GLM-5, a three-runs-per-task cost nobody quoted, and a harness that cannot grade its own memo.
Claude Code and Codex: the worker and the reviewer in ARIS and in most overnight loops, and the harness that scored DeepSeek's new open model.
DeepSeek V4.1 Flash 1017: the model everyone is routing cheap loop steps to, with Search and Fetch free inside agent tools until the end of the month.
Pi and pi-durable: the baseline harness in the ForesightBench experiment and the runtime behind the Telegram task-board agent.
Gemini 4 Argon: the two-hour background loop that built a game 20,000 people played.
Jev: the small decision model taking yes/no steps out of frontier pricing.
MiniMax M3.1 Flash: 0.15 and 0.60 dollars per million, weights due October 23, and a missing reasoning_content field that breaks naive loops.
MiniCPM-V 4.7 and Qwen: the 91 and the 52 on change-of-plan recovery at the same quant, and a RAM-only loop that failed outright.
harbor: the runner behind Embodied First Exam and the fleet repo lists.
Yukon leaderboards: 30,687 scored submissions behind the cross-model pollination study, with the integer-multiplication board still taking a PR a minute.
LangChain: the model, context, harness framing and the visibility-first cost pipeline.
Comments