Open robotics just got 10x faster to train. lerobot v0.5, out september 13, uses streaming video…
open robotics just got 10x faster to train. lerobot v0.5, out september 13, uses streaming video encoding to hit the speedup, and the paper landed at iclr 2026. the transformers-for-robotics claim is starting to hold.
read the note →The boring part of compilers is where agents actually shine. a september 16 paper, peepholebench,…
the boring part of compilers is where agents actually shine. a september 16 paper, peepholebench, turns real missed llvm instcombine optimizations — the small localized patches llvm's own pipeline never catches — into a coding-agent test. the unglamorous 20-line fixes are the benchmark now.
read the note →The market is pre-pricing alibaba's roadmap before it exists. citi expects the yunqi conference,…
the market is pre-pricing alibaba's roadmap before it exists. citi expects the yunqi conference, opening september 22 in hangzhou, to bring qwen v3.8 or v4.0 plus updates to alibaba's pingtouge chips, and the stock already rose on the expectation. the announcement is almost secondary.
read the note →The phone is becoming the whole dev environment. google's ai studio went mobile on september 16,…
the phone is becoming the whole dev environment. google's ai studio went mobile on september 16, letting you describe an app and generate it from your voice, not a terminal. the next generation of builders may never touch an ide.
read the note →Semantic code search is a debugging project masquerading as a search feature. jetbrains' september…
semantic code search is a debugging project masquerading as a search feature. jetbrains' september 17 developer diary on building context admits the hard part was the rag pipeline giving agents citable evidence instead of whatever grep surfaces. search was the easy half.
read the note →Video pricing just hit a new floor. creatify's boreal, out september 15, charges one cent for every…
video pricing just hit a new floor. creatify's boreal, out september 15, charges one cent for every second of video it generates, and it's their second in-house model. the unit economics of video are collapsing faster than anyone expected.
read the note →Non-determinism is the real tax on ai features. a september 18 study of in-ide ai developers found…
non-determinism is the real tax on ai features. a september 18 study of in-ide ai developers found 56% struggle to test their features because behavior shifts between runs, ahead of cost at 53% and privacy at 52%. the hardest bug is the one that only happens sometimes.
read the note →Most agent pilots die before production. gartner, september 14, says 40% of enterprise apps will…
most agent pilots die before production. gartner, september 14, says 40% of enterprise apps will embed task-specific agents by year end, yet only 17-31% of orgs have deployed any, and 86-88% of pilots never graduate. the gap between embedding and operating is the real story.
read the note →Google's open models just went edge-native. gemma 4, out september 16, runs audio and vision fully…
google's open models just went edge-native. gemma 4, out september 16, runs audio and vision fully offline on phones, raspberry pi and jetson, with e2b and e4b sizes. frontier ai now has a pocket version.
read the note →The interesting patent isn't the rewrite, it's the map. microsoft's september 18 filing routes…
the interesting patent isn't the rewrite, it's the map. microsoft's september 18 filing routes legacy code through an intermediate schema instead of one-shot translation, the step where most migrations die. the artifact is the strategy.
read the note →The frontier can now find exploits faster than most humans. anthropic's glasswing page, out…
the frontier can now find exploits faster than most humans. anthropic's glasswing page, out september 21, says its mythos preview model found thousands of high-severity vulnerabilities, including some in every major os and browser. defensive security now races a model.
read the note →Ai-generated c++ has a measurable tax. a september 17 study found it brings more interface and…
ai-generated c++ has a measurable tax. a september 17 study found it brings more interface and coupling burden, copy and allocation overhead, and 5-8% higher compute use — targeted feedback cut that by 11.1%. quality is a feedback loop.
read the note →The agent workbench just left the code editor. alibaba's qoder went live on harmonyos pc on…
the agent workbench just left the code editor. alibaba's qoder went live on harmonyos pc on september 21, stretching agent workflows into docs, data analysis and research. the division of labor is now 'you give the idea, the agent builds it'.
read the note →The open-weight agent stack just got a desktop app. cline desktop, out september 14, runs parallel…
the open-weight agent stack just got a desktop app. cline desktop, out september 14, runs parallel ai agents with 300+ models plus plugins, mcp servers and skills, all open source. the local agent workflow is now a product.
read the note →Voice cloning went fully local. voicestudio, the open-source elevenlabs alternative, runs cloning,…
voice cloning went fully local. voicestudio, the open-source elevenlabs alternative, runs cloning, dubbing and transcription on your own hardware across 646 languages. the hosted ais have a local rival now.
read the note →The week's fastest adoption belongs to a model nobody talks to. jev from typesafe ai dropped its…
the week's fastest adoption belongs to a model nobody talks to. jev from typesafe ai dropped its waitlist on september 21, gave new users a $5 credit, and vercel, langchain and browser use teams built on it all week. the quiet model is winning the dev floor.
read the note →Ai found the bugs faster than humans could triage them. kernel cves per release jumped from about…
ai found the bugs faster than humans could triage them. kernel cves per release jumped from about 500 to nearly 2,000, and maintainers now see 5-10 reports a day, most real and most low priority. the bottleneck moved to attention.
read the note →Claude code ate his ram, so he wrote his own agent in rust. zerostack, 8.9 megabytes and about…
claude code ate his ram, so he wrote his own agent in rust. zerostack, 8.9 megabytes and about 7,000 lines, hit 540 hacker news points after its creator watched claude code chew through gigabytes. the counter to bloat is a unix tool.
read the note →The faster teammate is the worse teacher. a september 16 study had novices pair with copilot and…
the faster teammate is the worse teacher. a september 16 study had novices pair with copilot and with a human: copilot won the task and cut workload, but retest scores dropped. speed and learning now point opposite ways.
read the note →Test creation time just dropped 90%. browserstack's ai suite, out september 21, runs five agents…
test creation time just dropped 90%. browserstack's ai suite, out september 21, runs five agents for generation, authoring, healing, accessibility and visual review, and claims 40% fewer automation failures. the new bottleneck is deciding what to test.
read the note →Unlearning is now a hallucination fix. llm ghostbusters, updated september 17, cuts package…
unlearning is now a hallucination fix. llm ghostbusters, updated september 17, cuts package hallucination rates by 81% by surgically unlearning bad outputs while keeping the model's general ability. the fix is forgetting, not remembering.
read the note →Microsoft's ai-native ide now lets you bring your own model key. visual studio 2026, out september…
microsoft's ai-native ide now lets you bring your own model key. visual studio 2026, out september 15, runs any provider via byok with five copilot agents and claims 50% fewer hangs. the ide stopped selling the model.
read the note →Mysql just got agents without a migration. shannonbase upgrades running mysql instances in place…
mysql just got agents without a migration. shannonbase upgrades running mysql instances in place with analytics, ml and agent capabilities, no code changes, open source on github since september 15. the data never leaves the box.
read the note →The tool now grades your plugin. claude code's september release adds plugin eval, which runs your…
the tool now grades your plugin. claude code's september release adds plugin eval, which runs your plugin against a test suite and compares the score to a no-plugin baseline, and the init command writes the test cases for you. dogfooding became a command.
read the note →A coding coordinator that never writes code. cursor's projects feature, out september 10, plans the…
a coding coordinator that never writes code. cursor's projects feature, out september 10, plans the work and delegates to thousands of subagents while keeping context across months of changes. the bottleneck moves to planning.
read the note →Open weights is not open source, and the boardroom is starting to notice. the osi's definition…
open weights is not open source, and the boardroom is starting to notice. the osi's definition requires the full training and inference code plus enough data detail to rebuild the model, which most releases don't ship. the label is now a liability.
read the note →Open weights just got a 21 billion euro price tag. mistral raised 3 billion at that valuation on…
open weights just got a 21 billion euro price tag. mistral raised 3 billion at that valuation on september 11, keeping its models open while the closed labs raise bigger. the open-source bet now has to justify the number.
read the note →The swe-bench leaderboard can no longer order its top entries. a september 15 paper shows coding…
the swe-bench leaderboard can no longer order its top entries. a september 15 paper shows coding agents have converged so tightly that the gaps at the top are noise, and argues for measuring something else. the benchmark died of success.
read the note →The github trending board is all agents today. archify added 3,400 stars in a day, thu's openmaic…
the github trending board is all agents today. archify added 3,400 stars in a day, thu's openmaic 1,700, and deepseek-harness sits at 206,000 total. the fastest-growing repos this week are agent frameworks, not apps.
read the note →Domain agents just went open source. anthropic's python repo with reference agents and data…
domain agents just went open source. anthropic's python repo with reference agents and data connectors for banking, equity research and private equity hit 35,500 stars on github under apache-2.0. the boring industries get agents first.
read the note →The coding assistant market just got a free speed run. xai's grok code fast 1, out september 16,…
the coding assistant market just got a free speed run. xai's grok code fast 1, out september 16, hits 92 tokens a second for free across major platforms. the paid tier now has to justify itself.
read the note →The ai act stopped being a slide deck. since august 2, enforcement is live, and lawyers say the…
the ai act stopped being a slide deck. since august 2, enforcement is live, and lawyers say the first regulator letter is the opening move of your defense, not a paperwork exercise. compliance now runs on the critical path.
read the note →Serving got 2.8x faster with zero new hardware. vllm's september 13 optimization of kimi k3 cut…
serving got 2.8x faster with zero new hardware. vllm's september 13 optimization of kimi k3 cut latency 56-60% and first-token time up to 85% just by tuning the engine. the model didn't change, the serving code did.
read the note →The sdk pipeline just changed hands. anthropic acquired stainless, the platform that builds…
the sdk pipeline just changed hands. anthropic acquired stainless, the platform that builds official developer libraries for openai, google and cloudflare, for over $300 million. the api front door now has one owner.
read the note →The flaky test now fixes itself. checksum's continuous quality platform runs ai agents end-to-end…
the flaky test now fixes itself. checksum's continuous quality platform runs ai agents end-to-end and resolves 70% of test failures automatically through real-time auto-recovery. the test engineer's job became reviewing the repair.
read the note →Open weights just got a revenue split. moonshot's kimi k3 landed on aws bedrock september 18 under…
open weights just got a revenue split. moonshot's kimi k3 landed on aws bedrock september 18 under a new model where the foreign cloud hosts the model and shares the revenue, a first for a chinese open model. the business model became the moat.
read the note →A 2.8 megabyte model now fills out your forms. cua released cua-s1-forms on september 18, an open…
a 2.8 megabyte model now fills out your forms. cua released cua-s1-forms on september 18, an open mit-licensed 706,048-parameter one-pass option scorer that decides gui form actions, with weights, training code and dataset all published. the smallest specialist just got a job.
read the note →The agent finally got a body. rpent, open-sourced september 21 by tsinghua's rlinf team, runs…
the agent finally got a body. rpent, open-sourced september 21 by tsinghua's rlinf team, runs robots on gpt-6 astra with 92.6% task success on libero-pro and 7x faster end-to-end task completion. the frontier moved out of the chat window.
read the note →Charting just got composable. tanstack charts, out september 17, builds visualizations by composing…
charting just got composable. tanstack charts, out september 17, builds visualizations by composing elements instead of picking fixed chart types, with about 160k weekly downloads already in alpha. the grammar replaced the template.
read the note →Stop reading the agent's transcript, check what its tools did. microsoft's thinkingbox,…
stop reading the agent's transcript, check what its tools did. microsoft's thinkingbox, open-sourced september 11, verifies ai agents by inspecting tool side effects instead of the chat log, catching what the transcript hides. the evidence moved from words to state.
read the note →The 30b model now fits in a phone. mediatek's dimensity 9600 pro, out september 15, is the first…
the 30b model now fits in a phone. mediatek's dimensity 9600 pro, out september 15, is the first 2nm smartphone soc, with an npu that lifts llm prefill throughput 51% and runs mixture-of-experts models up to 30b on-device. the data center moved into your pocket.
read the note →The second agent now audits the first. qodo's agentic toolbox, launched september 9, runs…
the second agent now audits the first. qodo's agentic toolbox, launched september 9, runs adversarial code review from a separate agent directly inside the coding workflow, with governance decisions a human owns. review stopped being a shared hallucination.
read the note →The workflow file just became a markdown doc. github agentic workflows, generally available…
the workflow file just became a markdown doc. github agentic workflows, generally available september 15, compile natural-language markdown into locked yml files that run coding agents in actions, with permissions and safe outputs declared up front. ci is readable by humans again.
read the note →The engineer merged 8 times more code per day. anthropic's institute reported that in q2 2026 the…
the engineer merged 8 times more code per day. anthropic's institute reported that in q2 2026 the typical engineer merged 8x as much code daily as in 2024, because claude writes it and the human directs and reviews. the bottleneck moved from typing to judgment.
read the note →Microsoft's answer to electron is ai-written native apps. its winapp cli, detailed september 16,…
microsoft's answer to electron is ai-written native apps. its winapp cli, detailed september 16, builds winui 3 apps in about 30 minutes, aiming to end the age of bloated web wrappers on windows. the wrapper dies when native generation is free.
read the note →Agent-written code is now a cve faucet. georgia tech's vibe security radar recorded 35 new cvEs in…
agent-written code is now a cve faucet. georgia tech's vibe security radar recorded 35 new cvEs in march 2026 tied to ai coding tools, up from 6 in january, with researchers estimating the true count runs 5 to 10 times higher. the review queue is the new bottleneck.
read the note →The interview now tests whether you fix ai's mess. shopify's ai-enabled coding round, detailed…
the interview now tests whether you fix ai's mess. shopify's ai-enabled coding round, detailed september 16, drops candidates into code the ai may have written and judges how they review and fix it, with interviewers digging into testing strategy. understanding beats generating.
read the note →The community vote happened in pull requests. after hashicorp moved terraform to its busl license,…
the community vote happened in pull requests. after hashicorp moved terraform to its busl license, community-opened prs collapsed from 21% the month before to 9.3% in august and stayed at 9.5% in september, per spacelift's data. the license changed, then the contributions did.
read the note →