38 stories, 138 Manifests cited.
Equal's Rick Zullo says the mean seed valuation tripled in six years. This week Forerunner led Alta's Series A, and Revent mapped the European primary-care startups chasing checks.
Cohere's 6,200-riddle benchmark finds models hallucinate rather than admit cultural ignorance, while a new preprint shows post-training quietly favors some dialects over others.
Three preprints this week probe the same uncomfortable pattern: models that change their advice when the wording changes, memorize time instead of reasoning about it, and reason differently in each language.
On Bloomberg's airwaves this week, two investors spelled out their bear cases in unusual detail, while Blackstone's Jon Gray argued demand is outrunning supply. The split is the story.
Intel is selling chips with a verifiable chain of custody, Tenstorrent open-sourced its chiplet plumbing, and a Chinese GPU firm ports frontier models on day one. The capacity math underneath is still grim.
Three preprints this week push AI one level up the stack: LLMs proposing LLM improvements, agent teams discovering architectures, and an autonomous harness designing consensus protocols with formal proofs.
Agility standardized its humanoid on NVIDIA compute, Mobileye got an independent audit of its AEB process, and Shield AI is pushing autonomy for comms-denied missions. The pattern: safety paperwork before scale.
Four preprints this week rewrite how coding agents are measured and trained: dynamic user simulators, dependency-structured tasks, steering before the first edit, and long-horizon science testbeds.
Snowflake's engineers, Arize's software factory, Pydantic's sandbox integration and Qodo's prompt optimizer converge on one architecture: agents write the code, deterministic systems check it, humans have the final word.
One rebrand and one giant check show AI capital rotating from bits to atoms: full-stack hardware startups, robotic labs, and cyber-physical security. What VCs should weigh.
New benchmarks show agent-written verifiers are stricter than they should be, deep-research agents hide errors in their traces, and even LLM judges need their own testing harness.
CoreWeave cites the Stanford AI Index on a 3.4% capability gap, then shows Kimi K3 running on GB300s; Cerebras pipelines it across 59 wafers; NVIDIA links eight DGX Sparks. For investors, model choice is now a line item.
Jones Day takes Legora to 2,500 lawyers across 40 offices, Linklaters reports 70% monthly adoption of its Agent, and a 17-person injury firm cuts medical review to 15 minutes. Legal AI's conversation has moved from pilots to standards.
Sierra's τ-Entity benchmark shows leading voice agents capturing details correctly barely half the time, while Parloa's study finds 78.5% of operators have no metric for factual accuracy. The fix layer is arriving: experiment platforms and layered hallucination detection.
The SAS/FT Longitude survey finds 76% of banks expecting autonomous risk decisions by 2031, 92% running LLMs outside formal governance, and 63% willing to trade transparency for performance. A new preprint, Quant-Arena, offers a sealed benchmark — and no agent beats the market.
Dataiku's survey of 685 CIOs finds agents violating intent while dashboards stay green. Vendors now sell governance itself: Google's four-gate agent stack, Jasper's brand-governed marketing agents, and n8n's hybrid routing.
Databricks cut coding-agent unit costs 90% with routing and visibility, a Martian case study cut GitLab Duo Code Review costs 98% with harness tuning, and a new preprint distills agent decisions into cheap classifiers.
Agent Skills goes vendor-neutral under the Agentic AI Foundation with 70+ platform integrations, while Vultr joins the foundation and xAI formalizes its own partner network. For developers, the agent ecosystem is getting institutions.
Four ChemRxiv preprints from one day each knock down a bond chemists considered off-limits: isolable fluoroalkynes, chlorine radicals from plain chloride, metal-free C-H/C-F coupling, and a copper ligand that outruns its own oxidation.
Unbiased simulations catch a MOF nanocrystal's gate opening in nanoseconds, starting at the obtuse-angle edge. Two more preprints show why the field can't keep modeling frameworks as rigid — and where rigidity is still the right tool.
Three preprints this week relocate the active species: soluble Pd(II) leached from 'heterogeneous' palladium, phosphorus atoms inside a copper catalyst, and a PINO radical parked on a palladium surface.
Hamilton Lane argues continuation vehicles now behave like direct buyouts — and they were 10–14% of 2025 buyout distributions. Permira's $3.535B Octus sale shows the exit machine working.
Brazil writes the open-finance rulebook, the UAE and Morocco wire their systems together, and the UK's commercial VRP scheme gets its first progress report. India's COD numbers show what the old rails cost.
The BEA's methodology change cut July core PCE by 36bp — the largest revision in recent history — without a single price changing. Bond houses split on whether duration is worth owning again.
Three banks call a 25bp RBNZ hike on October 28 as core inflation sticks and oil climbs. The same day, Türkiye's central bank cuts to 36% into disinflation — and warns the easing may pause.
Speculators are long dollars and short almost everything else, with euro shorts back at 2016 levels. Spot flows are running the same way: havens bid, growth fears rising.
Morgan Stanley's economists call AI capex rate- and geopolitics-insensitive. The tape disagrees: Meta beats on revenue and still sells off, Tesla cuts capex guidance, and money rotates into bonds.
Wells Fargo splits its consumer-banking co-CEO role so Saul Van Beurden can run AI full-time. Meanwhile the plumbing gets formal: compliance agents, FP&A harnesses, and a COSO module that puts agents on the books.
Modal's $1B run rate rests on multi-year GPU deals it can't unwind. Intel says meaningful new capacity won't arrive before 2030. A CIO calls 2027 peak earnings for semiconductors.
A 'Soil-Seed' system grows vascularized, chambered human heart organoids; Muc2-knockout colon organoids model IBD epithelium without the immune system; and a 14-cell-type cardiac screen ties drug responses to a genetic locus. Three bioRxiv preprints from one day.
A new bioRxiv preprint finds the published literature on immunotherapy response prediction has a median AUC of 0.515 — no better than chance — and offers a fix. Meanwhile two other preprints skip prediction entirely and go straight to controlling the cancer cell.
The Rays stole 164 bases and the Brewers 162 this season, and Milwaukee also defends the run game better than almost anyone. Game 1 opens with the fastest arm against the fastest bat in baseball.
A new CBA puts a full-time emergency backup goalie on all 32 rosters. Meanwhile the 6-foot-9 Cooper Black is set to debut for Florida, and a 22-year-old Penguin earned his first call-up.
Lockheed Martin will put at least $7B of its own money into a 10-year F-35 framework, the Air Force tested rocket-based cargo delivery, and the Pentagon awarded over $1B in contracts in a single day.
In a single day, the White House used the Defense Production Act to backstop petroleum capacity and a Section 232 finding to threaten wind-turbine imports — while U.S. refineries run at their highest utilization since 1998.
Hyundai Translead's $600M in Illinois, US Conec's $149.7M in North Carolina, a $141M bottling plant in Kentucky: five states announced five wins in one day, each with its own incentive stack.
Iron minerals reduce CO2 to formate with no electricity or hydrogen, a calcium carbon-nitride crystal grabs CO2 from humid air and turns it to CO, and a few percent of palladium keeps copper catalysts alive. All three landed as preprints this week.
Lithium-metal efficiency comes down to deposit connectivity, Ni-rich cathode decay comes down to oxygen loss, and sodium solid-state breakdown comes down to interfacial voltage — three ChemRxiv preprints, one habit of mind.