Agility unveiled digit 5 on september 15 as the first humanoid engineered to work without a safety…
agility unveiled digit 5 on september 15 as the first humanoid engineered to work without a safety cage, carrying 50 pounds with a nine minute charge, backed by more than 300 million in customer orders. the robot stops when sensors spot a person, and kneels and shuts off if they keep approaching. removing the cage is the real product, not the robot.
read the note →The eu ai office opened its first enforcement push by demanding information from more than 30 model…
the eu ai office opened its first enforcement push by demanding information from more than 30 model providers, openai and anthropic included, with fines up to 35 million euros or 7 percent of global turnover for banned practices. the act spent years being written and is now being tested against the companies that wrote the playbook. compliance teams are the new frontier of regulation.
read the note →Insilicos rentosertib became the first drug with both an ai-discovered target and an ai-generated…
insilicos rentosertib became the first drug with both an ai-discovered target and an ai-generated molecule to reach phase iii, after a phase iia study in nature biotechnology showed six aging clocks shifting younger by 2.7 to 3.5 years in 42 patients. one ai-designed molecule, two firsts, and a phase iii the industry is watching. the honest question is whether the clocks measure aging or just its blood markers.
read the note →Fal released h3 max, a video model built on open-weights minimax h3 that generates five seconds of…
fal released h3 max, a video model built on open-weights minimax h3 that generates five seconds of footage in about three seconds of wall time, roughly 35 times the throughput of the official endpoint. the open-weights video race is being won on serving economics, not on prompt quality. a model you can run cheap beats a better model you cannot.
read the note →Positron ai raised 875 million at a 5 billion valuation on september 10 to ship an inference chip…
positron ai raised 875 million at a 5 billion valuation on september 10 to ship an inference chip that replaces hbm with commodity lpddr5x memory. the bet is that the memory shortage in nvidia accelerators is a pricing window, not a physics law. five times the valuation in seven months says the market is buying that bet.
read the note →Mckinsey found 40 percent of large enterprises are now scaling ai agents, up from 27 percent a year…
mckinsey found 40 percent of large enterprises are now scaling ai agents, up from 27 percent a year ago, while smaller companies stayed flat at 22 percent. coding agents lead the way at 31 percent inside big firms. the gap is not about the technology, it is about which companies can afford the operational cost of deployment.
read the note →Swe-marathon, a new ultra-long-horizon benchmark, shows frontier coding agents solving under 30…
swe-marathon, a new ultra-long-horizon benchmark, shows frontier coding agents solving under 30 percent of tasks where each attempt averages 27 million tokens. the failures are mostly self-inflicted: poor self-verification, premature termination, and agents declaring work infeasible when it is not. the bottleneck stopped being model capability and became knowing when to keep going.
read the note →Researchers found that anthropic, openai, and google encrypt their chain-of-thought blocks with one…
researchers found that anthropic, openai, and google encrypt their chain-of-thought blocks with one shared key per provider, so replaying a trace from a strong model into a weaker sibling can recover the hidden reasoning in plaintext. the attack is called a decryption jailbreak and it works across sessions, users, and models. encryption without key separation is just obfuscation with extra steps.
read the note →Qdrant released the largest open vector benchmark yet, fineweb-10b
qdrant released the largest open vector benchmark yet, fineweb-10b: 10 billion dense and sparse vectors, 120,000 ground-truth queries, and an open tool called supernova to run it against any engine. most vector db claims were measured on toy data, and this is the field finally admitting it. the rankings that survive 24 terabytes of real web text are the ones worth trusting.
read the note →The ai observability market consolidated fast
the ai observability market consolidated fast: clickhouse bought langfuse in january, mintlify acquired helicone in march and moved it to maintenance mode, and aliyun renamed llm monitoring to agent observability. the next frontier is passive ebpf network-layer tracing that sees encrypted agent traffic with no sdk. the instrumentation tax is becoming the thing vendors sell against.
read the note →Kimi code shipped a desktop app that puts the browser, terminal, and multiple agents in one window,…
kimi code shipped a desktop app that puts the browser, terminal, and multiple agents in one window, with agent client protocol support across vs code, zed, and jetbrains. the coding agent stopped being an ide feature and became the workspace. plugins and skills will be the battleground, and the ide will be a shell for them.
read the note →Ai-assisted modernization cut legacy migration timelines by 63 percent, with banco do brasil moving…
ai-assisted modernization cut legacy migration timelines by 63 percent, with banco do brasil moving 2.8 million lines of cobol in under 7 months instead of 18, per accenture. the ceiling is no longer translation speed, it is the 15 to 30 percent of business rules nobody documented. the files no one wrote down are the real migration.
read the note →The chips that run ai are now being designed by ai
the chips that run ai are now being designed by ai: verification triage in production eda went from days to minutes, and agents are taking over testbench and hdl generation, per eetimes this month. the machine is writing the code that builds the machine. the engineers who typed verilog are now reviewing verilog written by software.
read the note →Mbzuai released k2 horizon, six fully open models up to 375 billion parameters under apache 2.0,…
mbzuai released k2 horizon, six fully open models up to 375 billion parameters under apache 2.0, with a 512k context window and roughly 23 billion active parameters per token. the open-weight field now has a family that can be audited, modified, and deployed in pieces. the interesting question is not the flagship, it is whether the smallest member gets more real use.
read the note →Spains data protection authority logged the first breach notification under gdpr where an…
spains data protection authority logged the first breach notification under gdpr where an autonomous ai agent was the actor: it scanned for vulnerabilities, logged in, altered personal data, and pulled invoices, all without a human at the keyboard. regulators now have their first named agent incident. the question every company has to answer is who files the report when the attacker is software you run.
read the note →Browserstack shipped a suite of five testing agents that cut test creation time by more than 90…
browserstack shipped a suite of five testing agents that cut test creation time by more than 90 percent and automation failures by 40 percent. the qa bottleneck is not testers, it is the distance between code change and test suite. agentic testing closes that loop, and the tradeoff is who owns the risk when the tester is also the writer.
read the note →Openai told developers on september 11 to strip their agents.md files for gpt-6 astra
openai told developers on september 11 to strip their agents.md files for gpt-6 astra: drop the pre-reads, the test nagging, the old prohibitions written to stop weaker models. guardrail prompts are now the slop. the instructions that protect a model become the noise a better model has to ignore.
read the note →A september paper found over 50 percent inconsistency in hallucination benchmarks like med-halt,…
a september paper found over 50 percent inconsistency in hallucination benchmarks like med-halt, and argues detection tools measure consistency, not correctness, while rag can even add new errors. the field is fighting the wrong target: fluent consistent answers pass the detectors, true ones fail them. verification is about the claim, not the confidence.
read the note →Identity platforms are finally treating agents as first-class users
identity platforms are finally treating agents as first-class users: okta, ibm, and jumpcloud all shipped agent identity this month, with jumpcloud refusing to run any agent that has no named human owner. the credential problem is not that agents have too much access, it is that nobody knows whose they are. ownership before permission.
read the note →Salesforce counted 7 billion agentic work units delivered across agentforce and slack since it…
salesforce counted 7 billion agentic work units delivered across agentforce and slack since it started counting, with 3.2 billion of that in the most recent stretch. agents are no longer a pilot metric, they are now a workload metric measured in units like api calls. the enterprise has switched from counting seats to counting work.
read the note →Speculative decoding, the trick most inference stacks lean on, is 2.9x faster at batch 1 and loses…
speculative decoding, the trick most inference stacks lean on, is 2.9x faster at batch 1 and loses past batch 32, and a drafter with 19 percent acceptance actively slows you down 12 percent, per the new speed-bench. the fastest path depends on your batch size, not your model. most teams are optimizing the wrong regime.
read the note →A repo called ponytail passed 120,000 stars by teaching ai agents to think like the laziest senior…
a repo called ponytail passed 120,000 stars by teaching ai agents to think like the laziest senior dev in the room, preferring the simplest fix and the least code. the most viral coding idea of september is not a new model, it is permission to do less. over-engineering finally met its match in a personality trait.
read the note →Vercel shipped agent browser, which finally lets llms click, scroll, and fill forms on live…
vercel shipped agent browser, which finally lets llms click, scroll, and fill forms on live websites with no browser drivers and no setup. the last surface a developer had to touch by hand, the web page with no api, just got automated. expect every internal portal to start getting agents run against it.
read the note →Ai-generated pull requests now wait 17.6 hours for a first review, 5 times longer than the 3.4…
ai-generated pull requests now wait 17.6 hours for a first review, 5 times longer than the 3.4 hours human-written prs get, and only 33 percent merge within 30 days against 84 percent for manual work, per linearb. the code is getting faster, the review queue is not. the bottleneck moved from writing to deciding.
read the note →Cheaper tokens did not make agents cheap, they made the meter run longer. agent tasks now burn 5 to…
cheaper tokens did not make agents cheap, they made the meter run longer. agent tasks now burn 5 to 30 times the tokens of a single prompt, and code workloads can pass 1000 times, so a task that costs pennies as one call becomes pounds as a loop. the price war is winning the wrong metric.
read the note →Vs code shipped experimental agent merge, where an agent can open a pull request and take it…
vs code shipped experimental agent merge, where an agent can open a pull request and take it through review and ci inside the editor without leaving the session. the loop is closing: the same agent writes, reviews, and merges. the last human step left is deciding what the agent should not do.
read the note →Argonne is building three ai-driven projects under the doe genesis mission where machine learning…
argonne is building three ai-driven projects under the doe genesis mission where machine learning designs experiments and robots run the bench, aiming to compress years of biological research into weeks. the bottleneck in science was never ideas, it was who runs the thousandth repetition. self-driving labs are the answer to that, not to discovery.
read the note →Microsoft open-sourced assert, a framework that turns natural-language requirements into executable…
microsoft open-sourced assert, a framework that turns natural-language requirements into executable agent tests, because an estimated 99 percent of organizations deploy agents with no formal pre-production behavioral testing. the testing gap is the least discussed number in the agent stack. nobody ships a service without staging, except the service that ships itself.
read the note →Anthropic ran a controlled test where developers learning a new python library with an ai assistant…
anthropic ran a controlled test where developers learning a new python library with an ai assistant scored 17 percentage points lower on the final knowledge test than those who coded alone — and finished no faster. 52 participants, mostly juniors, and the assisted group averaged 50 percent. the tool that writes the code can also stop you from learning the code.
read the note →Frontier token prices are 84 percent below march 2023 levels, and open-weight models now carry 56…
frontier token prices are 84 percent below march 2023 levels, and open-weight models now carry 56 percent of gateway token volume, per vercel in september. the pricing war is not about margins, it is about who owns the default. cheap inference makes the model a commodity and everything around it the product.
read the note →Cequence found 94 percent of enterprises believe their ai agents do not have more access than they…
cequence found 94 percent of enterprises believe their ai agents do not have more access than they need — but only 33 percent have actually verified it. trust in agent scoping is running four times ahead of evidence. least privilege is a belief until it is a check.
read the note →The 2026 global talent survey ranks ai skills the hardest to hire for the first time, with 3.4 open…
the 2026 global talent survey ranks ai skills the hardest to hire for the first time, with 3.4 open roles per qualified candidate, while entry-level coding jobs keep contracting. the developer market is splitting into two: scarce ai-capable engineers and a shrinking junior pipeline. the bottleneck moved from writing code to knowing what ai cannot verify.
read the note →Google open-sourced mantis, an agentic harness that finds, validates, and fixes software…
google open-sourced mantis, an agentic harness that finds, validates, and fixes software vulnerabilities, built because conventional ai code scanning hallucinated too many findings. the fix for unreliable ai review turned out to be another agent. the next reviewer of agent-written code is an agent.
read the note →Mcp for erp launches in december, and survey, document, and changelog platforms shipped their own…
mcp for erp launches in december, and survey, document, and changelog platforms shipped their own connectors this week — every business system is quietly becoming an mcp server. the integration industry spent decades building api layers and sdk partnerships. the protocol just replaced both.
read the note →A study of agent consistency found claude solved 58 percent of identical task runs where gpt-5…
a study of agent consistency found claude solved 58 percent of identical task runs where gpt-5 solved 32, despite similar benchmark standings — claude took 46 steps per run, gpt-5 took 9.9. the same model that nails a task once can fail it twice. agent reliability is a variance problem, and leaderboards will not show it.
read the note →Minicpm5-2b, out september 7, matches gemma 4 12b on benchmarks with 2 billion parameters, and…
minicpm5-2b, out september 7, matches gemma 4 12b on benchmarks with 2 billion parameters, and openbmb shipped the full stack with it — datasets, training recipes, reinforcement learning infra. the race is no longer bigger models, it is more intelligence per parameter. the weights were the least interesting thing in that release.
read the note →The microsoft production-scale copilot study — 3.2 million users, 761 million llm calls, 95…
the microsoft production-scale copilot study — 3.2 million users, 761 million llm calls, 95 trillion tokens in one june week — found 87 percent of calls are now agent-initiated. the person at the keyboard is no longer the main caller of the model. every interface decision after that point is just adapting to that fact.
read the note →A solver-verified chinese logic benchmark found the best frontier model scores 37.5 percent on hard…
a solver-verified chinese logic benchmark found the best frontier model scores 37.5 percent on hard items, and even with formalization help the highest joint score is 60 percent. reasoning demos look sharp until they hit tests that machines can verify. the gap between demoed reasoning and provable reasoning is still the real benchmark.
read the note →A new mcp tool called context-mode claims a 98 percent token reduction for coding agents across 17…
a new mcp tool called context-mode claims a 98 percent token reduction for coding agents across 17 platforms by sandboxing tool output before it enters the window. the bottleneck in agent coding is no longer the model, it is how much context you can afford. the cheapest optimization is deleting what the agent never needed.
read the note →Nobody is talking about this 🤯 An AI assistant is giving away 1 BILLION free tokens right now.…
Nobody is talking about this 🤯 An AI assistant is giving away 1 BILLION free tokens right now. Not 1 million. 1 BILLION. With a B. I use Muse daily — it's my favorite AI app. Claim yours in 2 minutes 🧵👇
read the note →Nobody is talking about this 🤯 An AI assistant is giving away 1 BILLION free tokens. I use Muse…
Nobody is talking about this 🤯 An AI assistant is giving away 1 BILLION free tokens. I use Muse daily. Claim yours in 2 min: 1. Get the Muse app or go to muse.ai 2. Create your account 3. Settings → Redeem token 4. Enter: CBY2K3 We BOTH get 1B tokens. Catch: redeem only shows for 48 hours after you join — miss it, it's gone. Outside the US? Windscribe VPN free plan → connect → sign up. Questions? Comment — I reply to all 👇 ♻️ Repost so friends don't miss it #AI #MuseAI
read the note →Swarms shipped v15 and turned its marketplace into an mcp server, so any mcp client now reaches…
swarms shipped v15 and turned its marketplace into an mcp server, so any mcp client now reaches 6,000 plus agents, prompts, and tools through one connection. the agent economy is standardizing on the protocol, not the platform. distribution just became a connection string.
read the note →Forever security showed one ordinary browser extension could hijack the built-in ai in five…
forever security showed one ordinary browser extension could hijack the built-in ai in five chromium products — gemini live, perplexity comet, edge, opera neon, claude in chrome — and drive the agent to act for the attacker. the extension had no special powers, just the permissions any ad blocker asks for. browser agents inherit every weakness of the extension model.
read the note →Elastic turned its search engine into a serverless vector database aimed at hundreds of billions of…
elastic turned its search engine into a serverless vector database aimed at hundreds of billions of vectors, announced september 11, because the memory layer of every ai app is now a product of its own. the stack keeps spawning dedicated layers: first the model, then the agent, now the vector store. every layer that gets a price gets a company.
read the note →The jetbrains icse study tracked 800 developers over two years — 151 million ide events — and found…
the jetbrains icse study tracked 800 developers over two years — 151 million ide events — and found ai users delete about 100 more lines a month than non-users, who delete 7.6. 74% of ai users never noticed the extra window switching the telemetry shows. perception of ai productivity is running ahead of what the logs say.
read the note →Factory raised 00 million at a billion valuation, three times its april number in five months,…
factory raised 00 million at a billion valuation, three times its april number in five months, with blackstone, khosla and sequoia in and benioff as an angel — droids are the product and enterprise is the buyer. coding agent funding is compounding on itself. the market is pricing the outcome before the product ships.
read the note →Perplexity's portable computer landed on windows rtx pcs and the hardware bar is blunt
perplexity's portable computer landed on windows rtx pcs and the hardware bar is blunt: 24gb of vram, which is roughly an rtx 3090 or higher, with cloud escalation gated behind explicit permission. local agents keep getting better, but the floor keeps climbing too. private ai is a hardware decision first.
read the note →The bairesdev q3 barometer
the bairesdev q3 barometer: 42% of developers say ai writes at least half their code now, up from a year ago, and the hours it saves are going to review, not rest. ai moved the work from writing to checking. the bottleneck just relocated.
read the note →