The gpu hour just jumped a fifth. nebius raised on-demand rates on oct 1, taking an h100 from $3.85…
the gpu hour just jumped a fifth. nebius raised on-demand rates on oct 1, taking an h100 from $3.85 to $4.50 per hour and a b200 from $7.15 to $8.50, with coreweave following. the queue now sets the price.
read the note →The agent exam just graded everyone below 35 percent. argo-bench, out oct 2, scored claude opus 5.5…
the agent exam just graded everyone below 35 percent. argo-bench, out oct 2, scored claude opus 5.5 as the top agent and it still cleared only 34.8 percent of 210 real tasks, averaging 59.5 points. the passing bar now sits above every model.
read the note →The flagship video model just went reference-heavy. shengshu's vidu q4 preview, out oct 7, takes up…
the flagship video model just went reference-heavy. shengshu's vidu q4 preview, out oct 7, takes up to 15 image references and 3 audio clips at 2k or 4k, priced from $0.014 per second. the storyboard now feeds the whole frame.
read the note →The review benchmark just capped the leader at 26 percent. github's reviewbench, out oct 5, tested…
the review benchmark just capped the leader at 26 percent. github's reviewbench, out oct 5, tested ai reviewers on 219 pull requests and the best one still misses 74 percent of known defects, with cursor's grounded recall at 9.6. the reviewer now has a score to answer to.
read the note →The agent protocol just went stateless. the july mcp spec, still the topic in october, deleted the…
the agent protocol just went stateless. the july mcp spec, still the topic in october, deleted the initialize handshake and session header, and developers are split on whether it's now just rest. the connector now routes without a conversation.
read the note →The open agent lab just priced itself at 1.5 billion. nous research raised 90 million dollars on…
the open agent lab just priced itself at 1.5 billion. nous research raised 90 million dollars on oct 7, backed by nvidia, microsoft's m12, samsung and y combinator, led by robot ventures. the open weights now carry a venture price tag.
read the note →The biotech just hired 37,075 agents. a stanford team ran a virtual biotech in a science study out…
the biotech just hired 37,075 agents. a stanford team ran a virtual biotech in a science study out oct 5, finding narrow cell-type targeting predicts trial success better than broad-spectrum mechanisms. the drug pipeline now runs on a swarm.
read the note →The flash model just 5x'd agentic throughput. deepseek-v4.1-flash on vllm, out oct 7, uses decoder…
the flash model just 5x'd agentic throughput. deepseek-v4.1-flash on vllm, out oct 7, uses decoder replay with cuda graphs to cut prefill compute by 30 to 40 percent. the small model now outruns the queue.
read the note →The jailbreak just got 93 percent cheaper. blindbias, out oct 2, cuts the cost of attacking an llm…
the jailbreak just got 93 percent cheaper. blindbias, out oct 2, cuts the cost of attacking an llm by 93 percent without touching model weights or logits. the guardrail economics just changed.
read the note →The e2e suite just went model-free. tester-army's e2e, trending on github oct 6, lets an agent…
the e2e suite just went model-free. tester-army's e2e, trending on github oct 6, lets an agent drive your app tests once, then replays the whole run with zero model calls. the regression now runs on a recording.
read the note →The video model just passed a human test. tavus's griffin, released oct 8, convinced 48 percent of…
the video model just passed a human test. tavus's griffin, released oct 8, convinced 48 percent of callers in its own study that it was human on a phone call. the uncanny valley now lives in customer support.
read the note →The image editor just halved its api bill. google's nano banana 2.1, out oct 6, adds mask-based and…
the image editor just halved its api bill. google's nano banana 2.1, out oct 6, adds mask-based and multi-image editing while gemini image pricing dropped about 50 percent. the retoucher now runs on a discount.
read the note →The ide just opened its agent era. jetbrains air, in early access since oct 1, brings agentic…
the ide just opened its agent era. jetbrains air, in early access since oct 1, brings agentic development to intellij and the rest of the jetbrains ides as a plugin. the editor now hires its own operator.
read the note →The frontier just launched invite-only. gemini 4 argon, out sep 30, opens at $2 per million input…
the frontier just launched invite-only. gemini 4 argon, out sep 30, opens at $2 per million input tokens but access is limited to vetted cyber defenders. the paid api now waits on clearance.
read the note →The extension just became an agent backdoor. a malicious browser add-on called bragjack hijacked ai…
the extension just became an agent backdoor. a malicious browser add-on called bragjack hijacked ai agents across chrome, edge, comet, opera neon and claude, per october reporting, with chrome patched in v143. the sidebar now needs its own security review.
read the note →The agent inventory just became a terminal command. geiger, an open-source cli out oct 8, maps…
the agent inventory just became a terminal command. geiger, an open-source cli out oct 8, maps every ai agent, harness, mcp server and plugin on your machine and what each one can reach. the supply chain now has a scanner.
read the note →The gigawatt just got a hundred-billion co-signer. nvidia committed up to 100 billion dollars with…
the gigawatt just got a hundred-billion co-signer. nvidia committed up to 100 billion dollars with openai on a 10-gigawatt ai data center, per oct 8 reporting, with the first gigawatt due in late 2026. the compute buildout now rivals national budgets.
read the note →The vector index just had a buffer overflow. pgvector 0.8.7, out oct 5, patches a flaw in ivfflat…
the vector index just had a buffer overflow. pgvector 0.8.7, out oct 5, patches a flaw in ivfflat index builds that could run arbitrary code. the embedding store now ships its own cve.
read the note →The chip giant just circled its own open-weight bet. nvidia, already reflection ai's biggest backer…
the chip giant just circled its own open-weight bet. nvidia, already reflection ai's biggest backer at 800 million dollars, is in talks to buy it outright, per oct 11 reporting, after a march round valued the startup at 25 billion. the supplier now wants the whole model shop.
read the note →The flagship just got a 50 percent haircut. openai cut gpt-6 api prices on oct 7, dropping sol to…
the flagship just got a 50 percent haircut. openai cut gpt-6 api prices on oct 7, dropping sol to $2 per million input tokens as the price war with anthropic's opus 5.5 went public. the reasoning tier now trades like a commodity.
read the note →The scheduler just left meta's datacenter. meta open-sourced rebalancer on oct 10, the c++ and…
the scheduler just left meta's datacenter. meta open-sourced rebalancer on oct 10, the c++ and python library it used for 9 years to place shards and servers. the invisible load balancer now has a public home.
read the note →The ai answer just cited its own spam. perplexity surfaced over 215,000 ai-generated spam pages in…
the ai answer just cited its own spam. perplexity surfaced over 215,000 ai-generated spam pages in search results, per oct 9 reporting. the citation graph now feeds on its own output.
read the note →The coding agent just left the terminal. github copilot's computer use, in public preview since oct…
the coding agent just left the terminal. github copilot's computer use, in public preview since oct 5, lets the cli and the app operate macos and windows desktop apps, off by default and approval-gated. the sandbox now covers the whole desktop.
read the note →The open-weight trillion just landed in europe. mistral's large 4 preview, out oct 6, pairs 1.05…
the open-weight trillion just landed in europe. mistral's large 4 preview, out oct 6, pairs 1.05 trillion total parameters with 49 billion active and a million-token window, trained on 3,800 grace blackwell gpus. the frontier now trains on the continent's own grid.
read the note →The workforce just got traded for capacity. oracle is cutting about 30,000 roles globally, 12,000…
the workforce just got traded for capacity. oracle is cutting about 30,000 roles globally, 12,000 of them in india, per oct 7 reporting, while analysts point the savings at data centers. the headcount now funds the fleet.
read the note →The voice just moved inside the song. suno's speech beta, out oct 2, generates spoken vocals and…
the voice just moved inside the song. suno's speech beta, out oct 2, generates spoken vocals and original music in a single track. the feature now owns the chorus.
read the note →The ai-generated label just got its rulebook. the european commission published the final code of…
the ai-generated label just got its rulebook. the european commission published the final code of practice on marking ai content, per oct 6, as ai act enforcement starts. the machine-readable tag now has a standard.
read the note →The reserved gpu just got 15 percent pricier. aws raised capacity block rates on oct 7, taking an…
the reserved gpu just got 15 percent pricier. aws raised capacity block rates on oct 7, taking an h100 hour from $5.191 to $5.970 and a b200 from $12.355 to $14.208. the spot market now keeps the discount.
read the note →The safety budget just went independent. openai is putting 7.5 million dollars into the alignment…
the safety budget just went independent. openai is putting 7.5 million dollars into the alignment project, per oct 6, funding outside researchers on misalignment risks. the guardrails now get built by outsiders.
read the note →The image model just learned to lay out pages. black forest labs' flux 3, out oct 8, renders 4k…
the image model just learned to lay out pages. black forest labs' flux 3, out oct 8, renders 4k output with bounding-box layouts and up to 10 reference images. the compositor now lives inside the model.
read the note →The gigawatt just got a price tag. bernstein, in an oct 5 note, put one-gigawatt ai datacenters at…
the gigawatt just got a price tag. bernstein, in an oct 5 note, put one-gigawatt ai datacenters at 34.6 to 39.5 billion dollars of capex depending on the accelerator. the power bill now dwarfs the model bill.
read the note →The video model just learned physics. hidream's o1-video-1.0, out oct 11, co-models text and video…
the video model just learned physics. hidream's o1-video-1.0, out oct 11, co-models text and video so scenes keep physical consistency and natural pacing. the storyboard now plans the timeline itself.
read the note →The main agent just hired its own team. anthropic's dynamic workflows for managed agents, in public…
the main agent just hired its own team. anthropic's dynamic workflows for managed agents, in public beta since oct 9, let one agent assign sub-agents — and on 116,000 lines of code the swarm found 66 bugs where a single agent found 27. the handoff now happens inside the model.
read the note →The frontier model just failed its own safety review. openai cancelled the october release of…
the frontier model just failed its own safety review. openai cancelled the october release of gpt-6.1 astra after internal testing found it missed alignment standards, per oct 11 reporting. the ship date now answers to the safety case.
read the note →The agent fleet just outran the data. a survey of 540-plus data leaders, out oct 6, found 57…
the agent fleet just outran the data. a survey of 540-plus data leaders, out oct 6, found 57 percent of businesses run ai agents while only 8 percent call their data production-ready. the model is the easy part now.
read the note →The humanoid just shipped for the job nobody wants. minerva humanoids came out of stealth oct 6…
the humanoid just shipped for the job nobody wants. minerva humanoids came out of stealth oct 6 with a 10 million dollar pre-seed, led by general catalyst, and a robot named roger built for oil and gas. the danger pay now goes to hardware.
read the note →The reverse engineer just got an agent. morluto/rea, first on github trending oct 10, drives agents…
the reverse engineer just got an agent. morluto/rea, first on github trending oct 10, drives agents from app behavior down to native binaries. the decompiler now has a scout.
read the note →The code editor just priced itself at thirty billion. cursor closed a 2.3 billion dollar round at a…
the code editor just priced itself at thirty billion. cursor closed a 2.3 billion dollar round at a 29.3 billion valuation, per oct 6 reporting. the terminal now outvalues most public software companies.
read the note →The benchmark just started auditing itself. testjack, out oct 7, regenerates tests from prompt…
the benchmark just started auditing itself. testjack, out oct 7, regenerates tests from prompt requirements to catch patches that pass fixed tests but break the ask. the leaderboard score now has a witness.
read the note →The support ticket just lost its human handoff. zendesk's autonomous agent, out oct 6, claims it…
the support ticket just lost its human handoff. zendesk's autonomous agent, out oct 6, claims it resolves 80 percent of customer issues without a person. the escalation queue now starts where the model gives up.
read the note →The model maker just became a chip maker. openai's jalapeño inference chip, detailed oct 4, claims…
the model maker just became a chip maker. openai's jalapeño inference chip, detailed oct 4, claims 3.6x lower latency than nvidia silicon by tuning blocks to transformer inference. the stack now bends all the way to the wafer.
read the note →The theorem mill just hit 372. openai says an unreleased internal model produced 372 results on…
the theorem mill just hit 372. openai says an unreleased internal model produced 372 results on open math problems, per oct 8, including solving the kakeya conjecture in four dimensions. the proof pipeline now outpaces the referees.
read the note →The benchmark just crowned grok. grok 4.5 hit 94.9 percent on gpqa diamond zero-shot on oct 10,…
the benchmark just crowned grok. grok 4.5 hit 94.9 percent on gpqa diamond zero-shot on oct 10, ahead of gemini 3 pro at 93.4 and gpt-6.1 sol at 92.9. the frontier order shifted without an announcement.
read the note →The team's agent skills just became a git repo. tencent cloud open-sourced teamai on oct 10, a…
the team's agent skills just became a git repo. tencent cloud open-sourced teamai on oct 10, a git-based tool that shares skills and configs across 16 different agents like claude code, codex and cursor. the review workflow now applies to prompting too.
read the note →The humanoid factory just set its target. hyundai, owner of boston dynamics, named ex-amazon ai…
the humanoid factory just set its target. hyundai, owner of boston dynamics, named ex-amazon ai chief rohit prasad ceo on oct 9 to push a plant capable of 30,000 robots a year by 2028. the biped line now has a production number.
read the note →The coding agent just got owned before it started. at pwn2own ireland on oct 8, researchers hit…
the coding agent just got owned before it started. at pwn2own ireland on oct 8, researchers hit openai's codex with a single argument-injection flaw in its --repo flag, running shell commands during clone. the agent now needs a bouncer for its own inputs.
read the note →The chat window just grew widgets. openai's gpt-6 with intelligent ui, rolling out oct 8, renders…
the chat window just grew widgets. openai's gpt-6 with intelligent ui, rolling out oct 8, renders calculators, bill splitters and diagrams inside the conversation instead of plain text. the answer now shows up as the tool itself.
read the note →The gpu rental market just split in two. us spot compute now runs about twice the price of…
the gpu rental market just split in two. us spot compute now runs about twice the price of long-term contracts, per oct 11 reporting, with goldman projecting spacex's capacity alone growing from 1.4 to 7 gigawatts by 2027. the instant price now tells a different story than the committed one.
read the note →