Loop Daily: 2026-11-06
Weight editing got its first physical test and its first league in the same two days. Opus 5.5 edited a robot policy's weights and cut task time by about 30 percent in 111 minutes on one RTX 4090, though the same team concedes only 4 to 12 percent came from the weights themselves, and that most of what weight editing does today could be done in code. The Open Kernel Challenge produced the clearest cost spread of the week: one entrant spent about $14,000 on tokens to reach 38th, another reached 92nd on a single 40-hour thread for $20. The sober findings came from everywhere else. A bench of ten antibody tasks found each add-on fixed one thing and broke another, a much-shared agent inside a model turned out to be mostly forced tokens, and a local harness that climbed from 77 to 99 percent on its own benchmark solved 1 of 6 on Terminal-Bench 4.0. Researchers checking 36,487 repos found loops nearly always have a trigger and far less often a check, state or exit.
#1
@Czcqet
https://x.com/Czcqet/status/2118381798585692541
Czcqet's team asked Opus 5.5 to write code that edits the weights of Οβ.β and makes a YAM robot faster, with no backpropagation and no inference code changes. In about 111 minutes on one RTX 4090 it tested a RoPE-based speedup, fitted it into Q/K weights, and unexpectedly found and repaired an RTC prefix bug. Task time fell about 30 percent with 100 percent success in bottle-placement trials, but that combines weight editing with a higher rate of 40 Hz against 30. At matched rates the edits alone saved 4.3 to 12 percent. The team's follow-up is the honest part: weight editing today may be a subset of autoresearch, with many effects equivalent to code changes.
https://x.com/Czcqet/status/2118381798585692541
Czcqet's team asked Opus 5.5 to write code that edits the weights of Οβ.β and makes a YAM robot faster, with no backpropagation and no inference code changes. In about 111 minutes on one RTX 4090 it tested a RoPE-based speedup, fitted it into Q/K weights, and unexpectedly found and repaired an RTC prefix bug. Task time fell about 30 percent with 100 percent success in bottle-placement trials, but that combines weight editing with a higher rate of 40 Hz against 30. At matched rates the edits alone saved 4.3 to 12 percent. The team's follow-up is the honest part: weight editing today may be a subset of autoresearch, with many effects equivalent to code changes.
#2
@ilblackdragon
https://x.com/ilblackdragon/status/2118142451223511389
ilblackdragon released Delta League, an open competition where anyone submits deltas to model weights and gets scored on a hidden test set. It's a response to this week's weight-editing excitement, which the post says was limited by the approach demonstrated. The goal is to auto-research and hillclimb smaller models into better performance, with verifiable evaluation rather than self-reported claims. Support for editing architectures and custom evaluation sets is promised next. A hidden test set is the right first move after a week of audits on the original paper.
https://x.com/ilblackdragon/status/2118142451223511389
ilblackdragon released Delta League, an open competition where anyone submits deltas to model weights and gets scored on a hidden test set. It's a response to this week's weight-editing excitement, which the post says was limited by the approach demonstrated. The goal is to auto-research and hillclimb smaller models into better performance, with verifiable evaluation rather than self-reported claims. Support for editing architectures and custom evaluation sets is promised next. A hidden test set is the right first move after a week of audits on the original paper.
#3
@RohanArun
https://x.com/RohanArun/status/2118165115900150148
RohanArun climbed to 38th overall in the Open Kernel Challenge with under 8 hours left, despite holding US and international CUDA patents and hand-writing no kernels. The workflow that finally hillclimbed unattended was a genetic algorithm with a .05 percent minimum advance, 3 tests to beat noise on an H200, Claude insight plus Nsight and a Lean verifier to guide candidates, and a small statistical model predicting results at 0.3 percent error. The top 100 now clears 2.30 times SGLang, and getting into the top 50 is as hard as the top 10 was days ago. A later reply puts the cost at about $2,000 in shared GPU and about $14,000 in tokens, mostly Gemini 3.8 Flash, spent partly to explore the limits of the method.
https://x.com/RohanArun/status/2118165115900150148
RohanArun climbed to 38th overall in the Open Kernel Challenge with under 8 hours left, despite holding US and international CUDA patents and hand-writing no kernels. The workflow that finally hillclimbed unattended was a genetic algorithm with a .05 percent minimum advance, 3 tests to beat noise on an H200, Claude insight plus Nsight and a Lean verifier to guide candidates, and a small statistical model predicting results at 0.3 percent error. The top 100 now clears 2.30 times SGLang, and getting into the top 50 is as hard as the top 10 was days ago. A later reply puts the cost at about $2,000 in shared GPU and about $14,000 in tokens, mostly Gemini 3.8 Flash, spent partly to explore the limits of the method.
#4
@kaanaricioglu
https://x.com/kaanaricioglu/status/2118454524881039769
kaanaricioglu reached 92nd on the same challenge's MoE track, 70th at one point, with GPT-6.1 Sol at max and xhigh in one thread over 40 hours. There was no autoresearch harness, just Codex with personal skills and playbooks, and almost no prompting beyond telling it to research further with a swarm of Terra agents. The task cost $20 of free Modal credits. Set next to the $14,000 run above, it's a useful reminder that leaderboard position and spend don't move together, and that a good playbook does a lot of what a harness promises.
https://x.com/kaanaricioglu/status/2118454524881039769
kaanaricioglu reached 92nd on the same challenge's MoE track, 70th at one point, with GPT-6.1 Sol at max and xhigh in one thread over 40 hours. There was no autoresearch harness, just Codex with personal skills and playbooks, and almost no prompting beyond telling it to research further with a swarm of Terra agents. The task cost $20 of free Modal credits. Set next to the $14,000 run above, it's a useful reminder that leaderboard position and spend don't move together, and that a good playbook does a lot of what a harness promises.
#5
@uctive
https://x.com/uctive/status/2118284347358744689
uctive has been running an experimental self-improving harness for Qwen3.8-Flash-Next locally for 5 days, and asked it for a progress report. Solved tasks on a 178-task benchmark rose from 77 to 99 percent without retraining, across 15 releases and 37 million tokens, while over half the elapsed time was idle waiting on hardware swaps and human decisions. The same tokens would cost about $3.25 on Qwen Flash and $143 on GPT-5.4 by API, against an estimated β¬7 to β¬10 of electricity. The post states plainly that this isn't a claim about model intelligence, and that the split and the power use are estimates.
https://x.com/uctive/status/2118284347358744689
uctive has been running an experimental self-improving harness for Qwen3.8-Flash-Next locally for 5 days, and asked it for a progress report. Solved tasks on a 178-task benchmark rose from 77 to 99 percent without retraining, across 15 releases and 37 million tokens, while over half the elapsed time was idle waiting on hardware swaps and human decisions. The same tokens would cost about $3.25 on Qwen Flash and $143 on GPT-5.4 by API, against an estimated β¬7 to β¬10 of electricity. The post states plainly that this isn't a claim about model intelligence, and that the split and the power use are estimates.
#6
@uctive
https://x.com/uctive/status/2118287404486312337
uctive then pointed the same harness at Terminal-Bench 4.0, with 66 complex tasks, and the first attempt was humbling: 1 solved, 5 failed, and an estimated full run of about 105 hours, so they stopped it. Stage two adds real-world tasks, multi-agent coordination, independent verification, regression checks and outside models reviewing through Hermes. The stated goal has moved from making the model look good on its own benchmark to solving independently verified tasks. A 99 percent on your own test followed by 1 of 6 on someone else's is the most instructive pair of numbers in today's feed.
https://x.com/uctive/status/2118287404486312337
uctive then pointed the same harness at Terminal-Bench 4.0, with 66 complex tasks, and the first attempt was humbling: 1 solved, 5 failed, and an estimated full run of about 105 hours, so they stopped it. Stage two adds real-world tasks, multi-agent coordination, independent verification, regression checks and outside models reviewing through Hermes. The stated goal has moved from making the model look good on its own benchmark to solving independently verified tasks. A 99 percent on your own test followed by 1 of 6 on someone else's is the most instructive pair of numbers in today's feed.
#7
@iiiichigo_chan
https://x.com/iiiichigo_chan/status/2118397992239571196
iiiichigo_chan points to researchers who checked 36,487 repositories: nearly all agent loops had a trigger, and far fewer had a check, state or an exit. The framing is sharp and accurate, that your agent loop is probably a cron job with your credit card. The thread walks through what's missing across harness, loop and graph. It fits the rest of this week's evidence that the risky part of running loops is the plumbing around the loop, which is the part nobody demos.
https://x.com/iiiichigo_chan/status/2118397992239571196
iiiichigo_chan points to researchers who checked 36,487 repositories: nearly all agent loops had a trigger, and far fewer had a check, state or an exit. The framing is sharp and accurate, that your agent loop is probably a cron job with your credit card. The thread walks through what's missing across harness, loop and graph. It fits the rest of this week's evidence that the risky part of running loops is the plumbing around the loop, which is the part nobody demos.
#8
@iiiichigo_chan
https://x.com/iiiichigo_chan/status/2118430865676227060
iiiichigo_chan also published the audit prompt, based on the same work across 256 real loops. It finds anything that starts an agent without you typing, and requires quoting the line that proves each of six parts: trigger, a hard limit, an outside check the agent can't edit, a stop rule, state written to disk, and cost per run. It flags any step that reads text from strangers while holding tools that can push or read secrets. Then it proposes the smallest fix for the worst loop, with three exits of green, out of budget or same failure twice, and waits for approval. You can run this today, and it's more useful than most harness guides circulating this week.
https://x.com/iiiichigo_chan/status/2118430865676227060
iiiichigo_chan also published the audit prompt, based on the same work across 256 real loops. It finds anything that starts an agent without you typing, and requires quoting the line that proves each of six parts: trigger, a hard limit, an outside check the agent can't edit, a stop rule, state written to disk, and cost per run. It flags any step that reads text from strangers while holding tools that can push or read secrets. Then it proposes the smallest fix for the worst loop, with three exits of green, out of budget or same failure twice, and waits for approval. You can run this today, and it's more useful than most harness guides circulating this week.
#9
@BioAgentDaily
https://x.com/BioAgentDaily/status/2118369022882267262
BioAgentDaily summarizes a paper testing LLM agents on 10 expert-validated antibody affinity-improvement tasks. Giving agents antibody tools made them better at following instructions and worse at improving binders. Giving them past scored designs improved optimization but stopped exploration. LLM judges matched programmatic checks of instruction following closely, and the full agentic loop came out roughly on par with a standard Bayesian optimization baseline. Every add-on fixed one thing and broke another, and matching a classical baseline is a useful ceiling to quote the next time someone promises agent-designed drugs.
https://x.com/BioAgentDaily/status/2118369022882267262
BioAgentDaily summarizes a paper testing LLM agents on 10 expert-validated antibody affinity-improvement tasks. Giving agents antibody tools made them better at following instructions and worse at improving binders. Giving them past scored designs improved optimization but stopped exploration. LLM judges matched programmatic checks of instruction following closely, and the full agentic loop came out roughly on par with a standard Bayesian optimization baseline. Every add-on fixed one thing and broke another, and matching a classical baseline is a useful ceiling to quote the next time someone promises agent-designed drugs.
#10
@burnrate_ai
https://x.com/burnrate_ai/status/2118164108969136314
burnrate_ai checked BabyAGI Inside, the widely shared project putting a fixed agent loop inside Qwen3.5-2B. In one five-cycle trace the repo itself reports 3,448 emitted tokens, of which 1,428 were generated by the model and 2,020 were forced by the controller, ROM or filesystem. The repo's own words are that this isn't general autonomous planning. It's a neat trick, and well under half of the output is the model. Always count who wrote the tokens before getting excited about a loop.
https://x.com/burnrate_ai/status/2118164108969136314
burnrate_ai checked BabyAGI Inside, the widely shared project putting a fixed agent loop inside Qwen3.5-2B. In one five-cycle trace the repo itself reports 3,448 emitted tokens, of which 1,428 were generated by the model and 2,020 were forced by the controller, ROM or filesystem. The repo's own words are that this isn't general autonomous planning. It's a neat trick, and well under half of the output is the model. Always count who wrote the tokens before getting excited about a loop.
#11
@Prayag12B
https://x.com/Prayag12B/status/2118161628378329380
Prayag12B explains ADIAS, which attacks why self-improving agent frameworks hit a wall. The usual loop saves whole candidate agents and aggregate scores, so the meta-agent forgets which fix solved which failure, overwrites working patches and repeats failed interventions. ADIAS keeps a persistent issue tracker across generations, with a stable ID, priority, lifecycle status, evidence and intervention history per failure mode, and logs the targeted fix separately from side effects. Across five benchmarks and four models it beat the strongest candidate-centric baseline by 25.2 percent, and removing the issue state cost up to 40.7 percent. Self-evolving agents need a bug tracker rather than a leaderboard.
https://x.com/Prayag12B/status/2118161628378329380
Prayag12B explains ADIAS, which attacks why self-improving agent frameworks hit a wall. The usual loop saves whole candidate agents and aggregate scores, so the meta-agent forgets which fix solved which failure, overwrites working patches and repeats failed interventions. ADIAS keeps a persistent issue tracker across generations, with a stable ID, priority, lifecycle status, evidence and intervention history per failure mode, and logs the targeted fix separately from side effects. Across five benchmarks and four models it beat the strongest candidate-centric baseline by 25.2 percent, and removing the issue state cost up to 40.7 percent. Self-evolving agents need a bug tracker rather than a leaderboard.
#12
@SJRonanMD
https://x.com/SJRonanMD/status/2118076567721271345
SJRonanMD gives the best reading of SpecPilot, from Berkeley, Princeton, Stanford and NVIDIA, an agent loop that tunes speculative decoding for a specific deployment. Stock systems leave speed on the table in different ways: with one request reading the weights dominates, and with batching per-request checking roughly halves the speedup. The loop measures those gaps, tries directions in parallel, stacks what works and checks output stays identical, going from 4.46 to 10.87 times on one request and 1.94 to 3.93 with batching. One discovery was a general 12-bit weight format. The catch is that these are the authors' own loops on their own hardware.
https://x.com/SJRonanMD/status/2118076567721271345
SJRonanMD gives the best reading of SpecPilot, from Berkeley, Princeton, Stanford and NVIDIA, an agent loop that tunes speculative decoding for a specific deployment. Stock systems leave speed on the table in different ways: with one request reading the weights dominates, and with batching per-request checking roughly halves the speedup. The loop measures those gaps, tries directions in parallel, stacks what works and checks output stays identical, going from 4.46 to 10.87 times on one request and 1.94 to 3.93 with batching. One discovery was a general 12-bit weight format. The catch is that these are the authors' own loops on their own hardware.
#13
@istoica05
https://x.com/istoica05/status/2118457699826548748
istoica05 points to Exo, from Ion Stoica's group, which evolves the harness alongside the solution in an agentic loop and reports big reductions in the cost and time to reach state-of-the-art results. It's short on numbers in the post itself. But it pairs with the same group's agent-team paper this week, in which a supervisor rewriting roles took a kernel from 0.31 to 1.21 times production speed. The direction is consistent: treat the harness as part of what the loop may change.
https://x.com/istoica05/status/2118457699826548748
istoica05 points to Exo, from Ion Stoica's group, which evolves the harness alongside the solution in an agentic loop and reports big reductions in the cost and time to reach state-of-the-art results. It's short on numbers in the post itself. But it pairs with the same group's agent-team paper this week, in which a supervisor rewriting roles took a kernel from 0.31 to 1.21 times production speed. The direction is consistent: treat the harness as part of what the loop may change.
#14
@tri_dao
https://x.com/tri_dao/status/2118418321456660572
tri_dao says a 2-year project is finally shareable: distilling most of the team's expertise in kernel and inference engine optimization into an agent loop that powers its production. There are few details in the post. What matters is the source, a kernel specialist saying this runs in production rather than on a benchmark. Combined with the Open Kernel Challenge, kernel tuning looks like the first expert craft agent loops are absorbing wholesale.
https://x.com/tri_dao/status/2118418321456660572
tri_dao says a 2-year project is finally shareable: distilling most of the team's expertise in kernel and inference engine optimization into an agent loop that powers its production. There are few details in the post. What matters is the source, a kernel specialist saying this runs in production rather than on a benchmark. Combined with the Open Kernel Challenge, kernel tuning looks like the first expert craft agent loops are absorbing wholesale.
#15
@iamMrDuncan
https://x.com/iamMrDuncan/status/2118248163010105720
iamMrDuncan is hillclimbing prefill on a P100 with auto-research and is up to 150 tokens per second on a 27B dense model at 3.5 bits per weight and 32K context. They thought it had hit a plateau, then it found great improvements. Old hardware being pushed this way by a loop is a nice counterpoint to GPU-rich leaderboard runs. The roadmap they describe turns Thermite auto research into an auto scientist, with hill climbing as only the start.
https://x.com/iamMrDuncan/status/2118248163010105720
iamMrDuncan is hillclimbing prefill on a P100 with auto-research and is up to 150 tokens per second on a 27B dense model at 3.5 bits per weight and 32K context. They thought it had hit a plateau, then it found great improvements. Old hardware being pushed this way by a loop is a nice counterpoint to GPU-rich leaderboard runs. The roadmap they describe turns Thermite auto research into an auto scientist, with hill climbing as only the start.
#16
@sundars22
https://x.com/sundars22/status/2118270518130552969
sundars22 shared a working autoresearch roster worth noting for its division of labor. GPT-5.6 Sol orchestrates, GPT-6 Astra handles research design, methodology and direction, GPT-6.1 Sol does implementation and verification, GPT-6 Luna keeps documentation consistent, and Opus 5.5 is consulted through Astra for research design. Everything runs at medium effort. Cross-vendor second opinions at medium effort is a far cheaper recipe than the max-everything setups this week's benchmarks keep embarrassing.
https://x.com/sundars22/status/2118270518130552969
sundars22 shared a working autoresearch roster worth noting for its division of labor. GPT-5.6 Sol orchestrates, GPT-6 Astra handles research design, methodology and direction, GPT-6.1 Sol does implementation and verification, GPT-6 Luna keeps documentation consistent, and Opus 5.5 is consulted through Astra for research design. Everything runs at medium effort. Cross-vendor second opinions at medium effort is a far cheaper recipe than the max-everything setups this week's benchmarks keep embarrassing.
#17
@roboindexio
https://x.com/roboindexio/status/2118459988134040004
roboindexio reports on Moonlake's agentic loop for robot policy improvement, which calibrates a physics simulation with a little real-world data, evaluates policies inside it, diagnoses failure modes, then improves them with targeted simulated data and RL. Starting from 50 real teleop episodes on a contact-rich sorting task, rejection fine-tuning added 11 points and RL across 384 parallel environments added another 38 in sim and 39 on the real robot. On the hardest 45-part setup the robot went from sorting 6 parts to 23. The sim-to-real match is the striking part, and simulation here serves as the evaluation layer a physical loop otherwise lacks.
https://x.com/roboindexio/status/2118459988134040004
roboindexio reports on Moonlake's agentic loop for robot policy improvement, which calibrates a physics simulation with a little real-world data, evaluates policies inside it, diagnoses failure modes, then improves them with targeted simulated data and RL. Starting from 50 real teleop episodes on a contact-rich sorting task, rejection fine-tuning added 11 points and RL across 384 parallel environments added another 38 in sim and 39 on the real robot. On the hardest 45-part setup the robot went from sorting 6 parts to 23. The sim-to-real match is the striking part, and simulation here serves as the evaluation layer a physical loop otherwise lacks.
#18
@krg85
https://x.com/krg85/status/2117968966379545074
krg85 assumed serving an open model was the easy part: start vLLM and you're done. It wasn't, and they got 64 percent more throughput one test at a time. Then they noticed they had been running the same loop by hand every time, so they built a harness modeled on Karpathy's autoresearch. That's the realistic adoption story: the loop arrives when someone notices they're already doing it manually, and the metric, throughput, was already there.
https://x.com/krg85/status/2117968966379545074
krg85 assumed serving an open model was the easy part: start vLLM and you're done. It wasn't, and they got 64 percent more throughput one test at a time. Then they noticed they had been running the same loop by hand every time, so they built a harness modeled on Karpathy's autoresearch. That's the realistic adoption story: the loop arrives when someone notices they're already doing it manually, and the metric, throughput, was already there.
#19
@sylviacccyu
https://x.com/sylviacccyu/status/2118299622125019279
sylviacccyu is running a targeted experiment on AI learning from failures to improve its own research process, and the design is more careful than most results shared this week. The base model, tools and tasks are frozen, and only failure logging, retrieval and experiment selection can change, while budget, scoring and test set stay fixed. Three arms separate retries, a static human researcher iterating the system, and validated new strategies, which disentangles those gains from recursive improvement. Early and late strategies face identical conditions, and failures get reported rather than only top runs. This is how to avoid fooling yourself about self-improvement.
https://x.com/sylviacccyu/status/2118299622125019279
sylviacccyu is running a targeted experiment on AI learning from failures to improve its own research process, and the design is more careful than most results shared this week. The base model, tools and tasks are frozen, and only failure logging, retrieval and experiment selection can change, while budget, scoring and test set stay fixed. Three arms separate retries, a static human researcher iterating the system, and validated new strategies, which disentangles those gains from recursive improvement. Early and late strategies face identical conditions, and failures get reported rather than only top runs. This is how to avoid fooling yourself about self-improvement.
#20
@steveb8n
https://x.com/steveb8n/status/2118462516980576606
steveb8n's team repurposed the autoresearch technique to autonomously improve Salesforce Apex code, says it worked, and lifted the tools out as an open source project. There are no numbers in the post, but the domain is the point. Enterprise languages with a measurable test suite and a large stock of old code are good candidates, and few people are looking there. Expect more of these as the method moves from ML training into line-of-business code.
https://x.com/steveb8n/status/2118462516980576606
steveb8n's team repurposed the autoresearch technique to autonomously improve Salesforce Apex code, says it worked, and lifted the tools out as an open source project. There are no numbers in the post, but the domain is the point. Enterprise languages with a measurable test suite and a large stock of old code are good candidates, and few people are looking there. Expect more of these as the method moves from ML training into line-of-business code.
#21
@jayw_actwise
https://x.com/jayw_actwise/status/2118435885796716966
jayw_actwise cites Manus reporting that an agent loop left more than 200 items awaiting a person's decision after one weekend. Faster execution overwhelmed prioritization. The takeaway is to decide which work deserves to start before multiplying it, which fits a week of evidence that review capacity is the real ceiling. A loop that produces decisions faster than anyone can make them hasn't saved any time at all.
https://x.com/jayw_actwise/status/2118435885796716966
jayw_actwise cites Manus reporting that an agent loop left more than 200 items awaiting a person's decision after one weekend. Faster execution overwhelmed prioritization. The takeaway is to decide which work deserves to start before multiplying it, which fits a week of evidence that review capacity is the real ceiling. A loop that produces decisions faster than anyone can make them hasn't saved any time at all.
#22
@shivanathd
https://x.com/shivanathd/status/2118046495022710892
shivanathd covers a ransomware incident at a UK school estate in which, per Bitdefender, commands repeated every 40 to 100 seconds in an agent-like pattern, though Bitdefender couldn't confirm a model was involved. Another summary in the same window counted about 37 of 50 machines encrypted. The author's response is the useful bit: check who can read your backup credentials. Even with attribution unproven, defenders should assume loop-speed iteration and plan around what an attacker reaches first, rather than around human dwell time.
https://x.com/shivanathd/status/2118046495022710892
shivanathd covers a ransomware incident at a UK school estate in which, per Bitdefender, commands repeated every 40 to 100 seconds in an agent-like pattern, though Bitdefender couldn't confirm a model was involved. Another summary in the same window counted about 37 of 50 machines encrypted. The author's response is the useful bit: check who can read your backup credentials. Even with attribution unproven, defenders should assume loop-speed iteration and plan around what an attacker reaches first, rather than around human dwell time.
#23
@studioselfhq
https://x.com/studioselfhq/status/2118109809849532617
studioselfhq runs a marketing agency with zero employees on Basecamp and background agents, and the five rules are more interesting than the boast. The agent that writes copy never grades it, and a separate model tries to tear it apart. Every loop has an API spend limit and every ad set a daily budget. Each rejected piece appends a never-do-this rule to that client's voice guide, agents are read-only by default so nothing goes live until a to-do is ticked, and an agent pauses any ad running 30 percent over target CPA every 6 hours. There are no revenue figures here, so treat it as a rulebook rather than a result.
https://x.com/studioselfhq/status/2118109809849532617
studioselfhq runs a marketing agency with zero employees on Basecamp and background agents, and the five rules are more interesting than the boast. The agent that writes copy never grades it, and a separate model tries to tear it apart. Every loop has an API spend limit and every ad set a daily budget. Each rejected piece appends a never-do-this rule to that client's voice guide, agents are read-only by default so nothing goes live until a to-do is ticked, and an agent pauses any ad running 30 percent over target CPA every 6 hours. There are no revenue figures here, so treat it as a rulebook rather than a result.
#24
@SiliconValleyPX
https://x.com/SiliconValleyPX/status/2118272152223043892
SiliconValleyPX quotes Alexandr Wang telling Garry Tan that inside Meta, with the right agentic loop, evaluation system and metric, a swarm of agents can accomplish more than a team of 100 engineers, very easily. It's an unaudited executive claim from someone building these systems. Yet note what Wang names as the conditions: the loop and the metric, with nothing about the model. That emphasis matches what practitioners in this feed report, even if the 100 engineers figure is unverifiable.
https://x.com/SiliconValleyPX/status/2118272152223043892
SiliconValleyPX quotes Alexandr Wang telling Garry Tan that inside Meta, with the right agentic loop, evaluation system and metric, a swarm of agents can accomplish more than a team of 100 engineers, very easily. It's an unaudited executive claim from someone building these systems. Yet note what Wang names as the conditions: the loop and the metric, with nothing about the model. That emphasis matches what practitioners in this feed report, even if the 100 engineers figure is unverifiable.
π‘ Eco Products Radar
Eco Products Radar
Claude Code and Opus 5.5: the default engine for kernels, weight editing and loop audits
Codex and GPT-6 family: Sol and Astra rosters, and the cheap single-thread kernel run
Karpathy autoresearch: the reference pattern everyone cites, with many prompt templates in circulation
SpecPilot: agent loop for speculative decoding, 4.46 to 10.87 times
Open Kernel Challenge: the week's live proving ground for loops
EvoSkill v2: skill rewriting with warnings about gaming benchmarks
Delta League: hidden-test scoring for weight edits
Hermes: harness for connecting outside reviewers
atlas-gic and MiroFish: trading loop templates, paper-only and mostly promotional
Qwen3.8: the most common open model in loop experiments
Claude Code and Opus 5.5: the default engine for kernels, weight editing and loop audits
Codex and GPT-6 family: Sol and Astra rosters, and the cheap single-thread kernel run
Karpathy autoresearch: the reference pattern everyone cites, with many prompt templates in circulation
SpecPilot: agent loop for speculative decoding, 4.46 to 10.87 times
Open Kernel Challenge: the week's live proving ground for loops
EvoSkill v2: skill rewriting with warnings about gaming benchmarks
Delta League: hidden-test scoring for weight edits
Hermes: harness for connecting outside reviewers
atlas-gic and MiroFish: trading loop templates, paper-only and mostly promotional
Qwen3.8: the most common open model in loop experiments
Comments