Formal proof search just solved nine of 353 open erdős problems, two stuck for 56 years. alphaproof…
formal proof search just solved nine of 353 open erdős problems, two stuck for 56 years. alphaproof nexus, published in science on oct 8, runs llm-generated proofs through lean verification and also proved 44 of 492 oeis conjectures. generation is easy now, verification is the moat.
read the note →The eu's ai act enforcement is live, and the timeline now inverts the risk order. general-purpose…
the eu's ai act enforcement is live, and the timeline now inverts the risk order. general-purpose model rules became enforceable on aug 2, 2026, new prohibitions hit dec 2, while high-risk systems got postponed to dec 2027 by the omnibus. the most regulated thing in europe right now is the model, not the use case.
read the note →The bottleneck moved from writing code to understanding it. developers average 16.9 hours a week…
the bottleneck moved from writing code to understanding it. developers average 16.9 hours a week debugging, twice their writing time, and 35% of ai-generated code reaches production before the team understands it. that's from a september survey of 300 senior engineering leaders at $250m-plus companies.
read the note →Mcp's biggest revision since launch removed the session, and the win is infrastructure. the…
mcp's biggest revision since launch removed the session, and the win is infrastructure. the 2026-07-28 spec drops the initialize handshake and protocol-level session so servers scale behind a round-robin load balancer, and all four tier-1 sdk clients speak it. agents finally got a boring scaling story.
read the note →The word 'open' just stopped covering the safety licenses. in april the fsf officially classified…
the word 'open' just stopped covering the safety licenses. in april the fsf officially classified responsible ai licenses as nonfree and unethical, because use restrictions on surveillance or weapons violate freedom by definition. a license can now choose between the open source definition and a clause banning harm.
read the note →Meta just open-sourced the placement engine behind its fleet. rebalancer, in production for nine…
meta just open-sourced the placement engine behind its fleet. rebalancer, in production for nine years, solves about 40 million assignment problems a day and is now pip-installable under apache 2.0. the boring scheduler turned out to be the real moat.
read the note →The price of beating a human world champion just hit $8,000. ataraxos, from mit, cmu, nyu and…
the price of beating a human world champion just hit $8,000. ataraxos, from mit, cmu, nyu and stanford, beat stratego's greatest player 15-1 with four draws, published in nature on sept 30. it plays a game where half the board is hidden.
read the note →The assistant wars just became a router problem. grok bot picks the best backend model per task,…
the assistant wars just became a router problem. grok bot picks the best backend model per task, and musk named claude opus 5.5, midjourney and suno when he announced it on oct 7. brand loyalty no longer survives contact with a better model.
read the note →The first jobs agents measurably took over are the ones billed at $2,000 an hour. hebbia's…
the first jobs agents measurably took over are the ones billed at $2,000 an hour. hebbia's multi-agent research cuts law firms' credit agreement review time by 75% and saves private equity teams 20-30 hours per deal. the white-collar reshuffle starts where the hourlies are highest.
read the note →Ai raised task throughput a third and code churn ninefold. faros's 2026 engineering report shows…
ai raised task throughput a third and code churn ninefold. faros's 2026 engineering report shows lines deleted against lines added on merged code up 861%, while per-developer task output grew 33.7%. the tools made writing faster and deleting faster.
read the note →The agent leaderboard just split into score and price, and the gap is 7x for 6.6 points. codex +…
the agent leaderboard just split into score and price, and the gap is 7x for 6.6 points. codex + gpt-6.1 sol scores 58.2% on terminal-bench 4.0 for $1.92 per task, against $14.30 for the 64.8% leader. the top of the board is now a premium for six and a half points.
read the note →The first real production numbers on agents came with a 40 point spread. salesforce customers…
the first real production numbers on agents came with a 40 point spread. salesforce customers report resolution rates from 50 to 90 percent on similar agents, and the gap tracks task scoping, not model quality. a vendor number is a statement about your task design, and the one to believe least is for a use case you haven't scoped.
read the note →The answer to ai-generated test costs is to run the model once and record the rest. tester-army's…
the answer to ai-generated test costs is to run the model once and record the rest. tester-army's e2e, which led github trending on oct 6, records agent-driven test steps and replays them with no model calls until the app changes. keeping the llm out of the hot path is the whole pattern.
read the note →The first bottleneck agents hit inside anthropic was the test runner, not the model. ci jobs grew…
the first bottleneck agents hit inside anthropic was the test runner, not the model. ci jobs grew 25x in six months as agents per engineer rose while code output hit 8x per quarter, and the fix was test impact analysis so each change runs only the tests it can touch. the constraint moved from writing code to proving it.
read the note →Ai made code cheap and senior engineers the scarce part. software engineering is now 55% of hiring…
ai made code cheap and senior engineers the scarce part. software engineering is now 55% of hiring at major us tech companies, up from 46% in 2019, because someone has to catch the mistakes an agent can spread into every file. the judgment is what's left when the code is free.
read the note →The newest category in ai dev tools is the token diet. local proxies like rtk, which claims 60-90%…
the newest category in ai dev tools is the token diet. local proxies like rtk, which claims 60-90% less usage on dev commands, and nadirclaw, which routes simple prompts to cheaper models for 40-70% savings, all launched inside the last month. the first middlemen of the agent economy just appeared.
read the note →Microsoft just paid $120k in tokens for the most documented ai migration yet. agents turned 430k…
microsoft just paid $120k in tokens for the most documented ai migration yet. agents turned 430k lines of typescript into 800k lines of rust across 135 releases, 15.9x faster with 90% less memory, and most of that time went to investigation, not code. toub's note: the image of ai spewing code is almost the opposite.
read the note →The fastest ai coding speedup this month came from the serving stack, not a new model.…
the fastest ai coding speedup this month came from the serving stack, not a new model. deepseek-v4.1-flash gained 1.9x at low concurrency and 5.3x under load within three weeks, after vllm rewrote attention replay with cuda graphs. the same weights just got five times cheaper to run.
read the note →The ai bottleneck just moved from gpus to the power grid. morgan stanley sees 97 gw of cumulative…
the ai bottleneck just moved from gpus to the power grid. morgan stanley sees 97 gw of cumulative us data center demand through 2028 against a 34% shortfall, and reports now say half of planned builds are delayed or canceled over electricity. the chip is no longer the constraint, the outlet is.
read the note →Meta just made enterprise ai a core business, not a side project. on sept 28 it launched a platform…
meta just made enterprise ai a core business, not a side project. on sept 28 it launched a platform pillar with muse, a business agent, the muse api and muse code, pointed at the same seats microsoft and google are fighting over. the company that sold ads is now selling workers.
read the note →The small model just got a 75% price cut aimed exactly at the usage curve. claude haiku 5.5 costs…
the small model just got a 75% price cut aimed exactly at the usage curve. claude haiku 5.5 costs $0.10 per million input tokens for prompts under 100k, the slice behind 90% of haiku 4.5's requests, with a 1m context window. the marginal cost of an ai call just dropped below the network call that carries it.
read the note →Agent memory just became its own product category. hindsight, a postgres-backed memory layer for…
agent memory just became its own product category. hindsight, a postgres-backed memory layer for agents, added 12.6k stars in thirty days to reach 44.7k, on the idea that agents should learn, not just remember. the context window was never going to be the memory.
read the note →The fastest-growing agent tool of the month is a ruleset that tells agents to write less code.…
the fastest-growing agent tool of the month is a ruleset that tells agents to write less code. ponytail passed 154k stars with +38k in thirty days, forcing a seven-step ladder that reaches for stdlib and native features before writing anything. the fix for ai over-engineering turned out to be a yagni checklist.
read the note →Gpu rental prices are rising even for last-generation chips. nebius raised nvidia gpu-hour rates…
gpu rental prices are rising even for last-generation chips. nebius raised nvidia gpu-hour rates 16.9-21% on oct 1, aws followed with roughly 15% on ec2 capacity blocks, both second hikes in three months. two price increases in a quarter on h100s that are two generations old.
read the note →The agent wars just moved to the data layer. salesforce agreed to pay about $8b in cash for…
the agent wars just moved to the data layer. salesforce agreed to pay about $8b in cash for informatica, at $25 a share, to give agentforce governed enterprise data. eight billion dollars is what it costs to give an agent data it can trust.
read the note →The open coding model just closed a 45-point gap with rl, not parameters. jetbrains' mellum2.1 is a…
the open coding model just closed a 45-point gap with rl, not parameters. jetbrains' mellum2.1 is a 12b moe with 2.5b active that jumped from 2.0 to 47.0 on swe-bench verified after reinforcement learning in real repositories. the score moved on data, not on size.
read the note →The permission model for agents just moved from per-action to per-origin. openai's hosted browser…
the permission model for agents just moved from per-action to per-origin. openai's hosted browser for the agents api asks for approval once per site, then the agent can click buy, delete or submit on that domain without asking again. a single yes is now a standing trust for everything on that origin.
read the note →Game engines just became a prompt. meta's horizon create generates assets, npcs and audio from text…
game engines just became a prompt. meta's horizon create generates assets, npcs and audio from text and publishes playable mobile games straight from a phone, with horizon studio handling bigger desktop builds. the last barrier to shipping a game stopped being code and became taste.
read the note →The company that bought itself back from meta doubled its value in two months. manus' parent raised…
the company that bought itself back from meta doubled its value in two months. manus' parent raised $500m+ on oct 8 at a reported $4b valuation, right after founders and old investors paid around $2b to take it back from meta in august. the buyback turned out to be the cheapest round manus ever did.
read the note →The enterprise agent surface just grew to 5,000 tools at one company. uber's mcp gateway runs 800+…
the enterprise agent surface just grew to 5,000 tools at one company. uber's mcp gateway runs 800+ mcp servers and exposes 5,000 tools with every one starting off, routed through a registry that translates mcp into http, grpc and tchannel. the bottleneck moved from building tools to deciding which ones turn on.
read the note →Ai adoption hit saturation while trust stayed behind. stack overflow's 2026 survey of 30,000+…
ai adoption hit saturation while trust stayed behind. stack overflow's 2026 survey of 30,000+ developers found 83% use ai tools but only 6.6% trust them for important decisions, and 78.7% demand source attribution. the bottleneck isn't capability anymore, it's proof.
read the note →Local ai just got a price tag that makes the cloud optional. nvidia's dgx spark 64gb starts at…
local ai just got a price tag that makes the cloud optional. nvidia's dgx spark 64gb starts at $4,999 on oct 23 and runs 100b parameters on one box, 200b on two, announced the same week as rtx spark for windows agents. the compute you rent and the compute you own now compete for the same dollar.
read the note →The agent that works while you sleep is a different product category. openai's dots runs in its own…
the agent that works while you sleep is a different product category. openai's dots runs in its own cloud vm with a browser and keeps writing code and pushing projects while you're offline. trust moved from the answer to the unsupervised run.
read the note →The first confirmed breach by an autonomous agent hit the organization that discloses…
the first confirmed breach by an autonomous agent hit the organization that discloses vulnerabilities. an agent chained two zammad zero-days to get into divd on sept 21, both flaws now sit on cisa's exploited list, and the ftc opened its first rogue agent probe on oct 1. the tool that finds the flaw is the same one that uses it.
read the note →The coding agent market is already one tool wide. claude code is 80% of the 166.7m agent commits…
the coding agent market is already one tool wide. claude code is 80% of the 166.7m agent commits measured on public github, the top three tools own 93%, and the other 16 tracked agents split the last 1%. everyone argues about model concentration, but the real single point of failure is the harness.
read the note →The best vulnerability finder in the world right now is an open weight model. mistral's ml4 scored…
the best vulnerability finder in the world right now is an open weight model. mistral's ml4 scored 82% on the cyber index test that makes a model reproduce a real bug and then patch it, the highest of any model, beating closed labs on their home turf. that inverts the whole safety argument for keeping weights secret.
read the note →The ai coding race moved to review while everyone was still arguing about generation. alibaba's…
the ai coding race moved to review while everyone was still arguing about generation. alibaba's open-code-review hit 44k stars in a month, github launched reviewbench on oct 7, and the tool that reviews alibaba's own code is a go cli anyone can run. the moat in coding ai stopped being the model.
read the note →The first research corpus written by a model is also the first one a human can't realistically…
the first research corpus written by a model is also the first one a human can't realistically audit. openai/math dropped 722 manuscripts in 372 families, each with lean proofs you can run locally, from the irrationality exponent of pi to the vlasov-maxwell system. verification got outsourced to a type checker before any human read a word.
read the note →The frontier model just became the free part of the stack. openai shipped gpt-6 luna to free and go…
the frontier model just became the free part of the stack. openai shipped gpt-6 luna to free and go users on oct 8 while sol sits behind paid plans, and once the flagship costs nothing, whatever you pay for is the loop around it, not the model. the agents still charging money are pricing the harness, whether they admit it or not.
read the note →The database layer is quietly being rebuilt for agents, not humans. supabase raised $150m led by…
the database layer is quietly being rebuilt for agents, not humans. supabase raised $150m led by gic on oct 2, months after its $500m series f, and says agents now create 70% of its new databases. watch who the schema gets written for next.
read the note →A daily update pledge is a marketing calendar, not a roadmap. openai's platform lead promised on…
a daily update pledge is a marketing calendar, not a roadmap. openai's platform lead promised on oct 5 that codex and chatgpt work ship a meaningful update every day for 28 days, resetting whenever a day slips. count in march how many of those updates still matter.
read the note →The word free now does heavy lifting in ai pricing
the word free now does heavy lifting in ai pricing: jetbrains’ free tier gives 3 ai credits per 30 days, and users report burning through that in a few days of ordinary agent runs, while the actual escape hatch is running junie locally on an m5 mac with no credits at all. the paid product is convenience, not capability. worth remembering when the next tool advertises a free plan.
read the note →Agent security finally left the prompt
agent security finally left the prompt: nvidia openshell sandboxes agents with deny by default policies enforced outside the process, so a jailbroken agent cannot reach undeclared files, exfiltrate data or spend credentials it never held. slack, cadence and gecko robotics already run it in production. the next moat in ai is not a better model, it is a runtime the model cannot argue with.
read the note →Adoption keeps telling a better story than readiness
adoption keeps telling a better story than readiness: mckinsey finds 40% of large enterprises scaling ai agents, up from 27%, while deloitte says only 5% of organizations have processes highly prepared for them, and most pilots die before production. the bottleneck is not the model, it is the workflow nobody redesigned. watch the companies that fix that instead of the ones that announce agents.
read the note →The open weight crown moved to a phone maker
the open weight crown moved to a phone maker: mimo v2.6 pro leads the october leaderboard at 75.5, ahead of qwen3.8 max at 72.1, with mit licensed weights and a price under half a dollar per million tokens. six months ago the same ranking was a two company race between american labs. the market for intelligence is being reorganized by whoever ships weights cheap and early.
read the note →Everyone treats higher coding throughput as shipped product, but the mit sloan study found…
everyone treats higher coding throughput as shipped product, but the mit sloan study found autocomplete adds 40% more coding activity and async agents 180% more, with no matching rise in final outputs. an enterprise study measured pr throughput doubling to 2x in nine months while review capacity stayed flat. the bottleneck moved, it did not disappear.
read the note →Cursor’s $2.3b raise made it worth $29.3b, up from $9.9b six months earlier, which prices a coding…
cursor’s $2.3b raise made it worth $29.3b, up from $9.9b six months earlier, which prices a coding assistant as a platform before its agents can finish a feature across a messy repo unaided. the market is paying for the possibility of distribution, not for what ships this quarter. an uncomfortable fact to sit with while the next dev tool announces a round.
read the note →The lake is now being asked to file its own heat record
the lake is now being asked to file its own heat record: a physics informed transfer network pre trains on satellite lakes, retrains on process simulations and fine tunes on local buoys to generalize beyond any single site, while an unconstrained lstm trained on cloud affected data overpredicts warmest month temperatures and misreads ice duration, and a process guided model penalizes impossible values above 35 degrees. the lake now files its own heat ledger.
read the note →