• The episode said Anthropic had filed an S-1 and was targeting a Nov. 16 IPO, with a stated $100B raise and a valuation discussion around $1.5–$2T; speakers cited $16.3B in H1 2026 revenue (14× year over year) and a $100B year-end ARR target, while noting ARR could mean annual recurring revenue or revenue run-rate. All-In Podcast
• Anthropic customers spending at least $100K reportedly grew from 200 at 2024 year-end to 1,500 at 2025 year-end and 11,900 by September; a speaker estimated more than 1 million global businesses for which $100K in annual spend would be less than 1% of total spend. All-In Podcast The speaker cautioned that ROI proof for large enterprise deals remained speculative: about 1,900 customers spending over $1M were said to have made an ROI case, many of them AI-native, leaving only a few hundred Global-2000-type companies. All-In Podcast
• The speakers’ S-1 review identified compute—especially data centers, financing commitments and power—as Anthropic’s leading constraint, with regulation second. All-In Podcast Their market-risk analysis estimated AI firms’ funding needs through end-2027 at about 30% of Treasury issuance, warning that competition for capital could raise borrowing costs for AI borrowers and downstream neoclouds, or for the wider economy if Treasury rates rise. All-In Podcast They argued that algorithmic efficiency can lower token prices while margins rise and usage expands, but flagged open models commoditizing coding and growing reliance on recursive self-improvement (RSI)—which they said Anthropic describes as risky—as strategic and safety concerns. All-In Podcast
• Google Search revenue was said to be up 12% year over year, from $56B to $63B, but ad-click monetization faces pressure from answer engines; Google Cloud doubled to $30.7B, while whether consumer AI can replace search-ad revenue remained an open question. All-In Podcast

• Google DeepMind’s AlphaEvolve found a way to multiply 4×4 complex matrices using 48 multiplications, one fewer than the 49 in Strassen’s method; Demis Hassabis described combining LLM-generated proposals with evolutionary search, and DeepMind published computer-checkable results. Code Bear
• OpenAI announced that a general-purpose model disproved Erdős’s unit-distance conjecture; nine outside mathematicians, including Timothy Gowers, wrote a companion paper checking the result, and Will Sawin showed the construction yields at least n^1.014 pairs. Code Bear Google DeepMind’s AlphaProof Nexus reported 9 of 353 open Erdős problems and 44 of 492 conjectures from the Online Encyclopedia of Integer Sequences, all Lean-checked and peer-reviewed. Code Bear
• OpenAI and Anthropic both said their models refuted strong ETH, and both papers included Lean proofs; this does not solve P versus NP, and Anthropic’s reported speedup only beats brute force in an extreme case involving more than 10,000-digit switch counts, not practical problem-solving. Code Bear
• OpenAI said a team of roughly 10,000 agents found a Navier–Stokes blow-up proof after about 88 hours, followed by 17 hours to formalize it; Sam Altman gave a rough cost estimate of around $10 million. Code Bear The claim concerns a version with an external force, and OpenAI said it would not seek the Clay prize; review remains unsettled, with three researchers arguing that the Lean proof does not match the written paper and the proof not yet fully understood. Code Bear
• OpenAI posted 722 manuscripts across 372 topics, with Lean proofs for about 42% of the main results. Code Bear Three papers were withdrawn the next day after a sign error in one undermined two dependent papers, and 14 more were revised. Code Bear

• Anduril announced BlueChip, a U.S.-designed computer intended to use domestic components, processed metals, and batteries; Packy McCormick frames it as evidence of the company’s growing manufacturing capacity and argues that high-volume products such as laptops could help pull more supply chains back to the U.S. Not Boring by Packy McCormick
• OpenAI announced model-generated results in harmonic analysis, complexity theory, and quantum field theory, including a proof of Sogge’s local-smoothing conjecture in every dimension, an Exponential Time Hypothesis (ETH) refutation, and a proof that interacting quantum field theories exist in four dimensions. The math proof and ETH result have Lean formalizations, but field experts still need to verify that the formal statements capture the intended claims; the physics group withheld its AI material and is writing it up itself. Not Boring by Packy McCormick Separately, Axiom announced a complete Lean auto-formalization of the Classification of Finite Simple Groups, and Physical Superintelligence published results on five open quantum problems with papers, Lean formalizations, and open-source verification code. Not Boring by Packy McCormick
• Anthropic says its life-sciences team has partnered with six organizations to run drug-development programs. Reported examples include Claude generating about 35,000 antibody designs in a month, three early candidates that dislodged IgE better than standard medication, a CAR-T redesign associated with roughly 10 times as many cancer-spotting receptors, and Enveda manufacturing-record review that was 89% faster and found seven issues human reviewers had missed. Not Boring by Packy McCormick The Column Group is launching Pleiades to run more drug programs with Claude, on the thesis that faster, cheaper discovery could make riskier bets viable; Anthropic cautions that nearly all of the reported work is preclinical. Not Boring by Packy McCormick
• Lilly’s retatrutide reduced sleep-apnea events in two 80-week Phase 3 trials involving 691 adults with moderate-to-severe apnea: in the trial without diabetes, the 9 mg dose reduced events by 34 per hour versus 10 with placebo, and about half of patients on higher doses reached resolved or mild apnea without daytime sleepiness versus roughly one in ten on placebo. The drug is not yet on the market, and whether it prevents heart attacks or strokes remains an open question for a dedicated trial. Not Boring by Packy McCormick
• Neuralink reported that one participant using its VOICE brain-computer interface reached 75 words per minute with a 5.8% error rate across a 125,000-word vocabulary in the first week, compared with a cited Stanford system’s 62 words per minute and nearly one-quarter of words wrong. The program plans to enroll up to six people with ALS, spinal cord injury, or stroke, so the reported result is from an early, small study. Not Boring by Packy McCormick
• The newsletter’s sponsor Reducto is pitching an SDK and embeddable editor for agent-enabled DOCX, PPTX, and XLSX editing, with formatting preservation, collaboration, and deployment in a browser or on a customer’s own servers; the ad says its document infrastructure is used by Harvey, Scale AI, Toast, and Vanta. Not Boring by Packy McCormick
• Readwise Books is in beta with waitlist access; the described product combines an e-reader with built-in AI, Readwise highlights, text-to-speech, and reading leaderboards. McCormick disclosed that he is an investor in Readwise and friends with its founders, making his strong product endorsement interested. Not Boring by Packy McCormick

• Superbase’s CEO reported 13M signups, up from about 3M roughly a year earlier; nearly 50M databases under management versus 8M; about 1M new databases per week; and expected 4× revenue growth for the year. Y Combinator
• The company has shifted from dashboard and API products toward code-first tooling and CLI for agents; its CEO said becoming a default database choice in LLMs helped, and described open source as critical partly because agents can inspect the code when problems occur. Y Combinator It plans a SQLite-compatible Spark tier for starter workloads alongside Postgres-based infrastructure intended to scale to petabytes. Y Combinator
• AI is widening the pool of potential builders beyond software engineers: the CEO said non-engineers are increasingly able to build and that signups, activation, and conversion were all rising; he cited a Brazilian taxi driver who built a local Uber-like app on his phone. Y Combinator
• The interview hosts described vertical SaaS startups struggling to raise Series A even at $5–10M ARR, while horizontal and AI investments remained favored; they said lean teams could reach $10–50M in revenue with 10–15 people. The CEO said he now views not needing to fundraise as an achievement. Y Combinator

• Anthropic’s S-1 was described as showing $16.3B in revenue in the first half of the year, versus $1.2B in the same period a year earlier, and many customers paying over $100,000 per month—evidence of its enterprise focus and rapid growth. TechCrunch The speakers also cited roughly $1T in compute commitments over the next decade or two and said Anthropic must keep growing to fund GPUs and data-center capacity. TechCrunch They juxtaposed a $693M operating loss with a net-loss figure rendered as about $80B, which they said largely reflected debt and commitments; the discussion did not reconcile that figure, and noted the obligations remain payable. TechCrunch The S-1 review also put acquisitions at about $737M excluding Coefficient Bio, with about $260M undisclosed, and identified a one-person biotech company among the acquisitions. TechCrunch
• Apollo’s chief economist reportedly found no sign of AI-driven gains in revenue or revenue per employee among S&P 500 companies after excluding seven major tech firms, despite substantial AI investment. TechCrunch The speakers cautioned that it may be too early to see returns, but warned that enterprise AI contracts could weaken if customers do not realize a payoff. TechCrunch
• Bedrock Robotics raised $600M and began commercial operations roughly two years after its founding, offering self-driving kits for construction equipment, particularly excavators. TechCrunch The company has a Waymo and robotics pedigree; the discussion also noted prior agriculture-autonomy disappointments, while pointing to construction’s many specialized niches as opportunities for focused physical-AI startups. TechCrunch
• Rebuild, a Massachusetts domestic-manufacturing accelerator/incubator founded in 2020 by former Amazon commercial CEO Jeff Wilke and Miles Arnon, raised $2.1B in a round led by General Catalyst with participation from JPMorgan Chase’s strategic investment group. TechCrunch It helped start and spin out Slate Auto and works on areas including industrial automation and drone assembly; the speakers described an Amazon-heavy network but said Rebuild is being run independently, not as an Amazon arm. TechCrunch
• Marissa Mayer’s Dazzle was described as an AI assistant that learns about users from their camera-roll photos; her example of using it to generate traits about her daughter drew backlash, though Mayer said coverage made it seem she needed the tool rather than was using her own product. TechCrunch The discussion linked the reaction to broader skepticism of AI executives’ consumer use cases and concern that AI optimization could displace human interaction. TechCrunch

• Eve sells legal AI to plaintiff-side firms, targeting their labor bottleneck with a workflow designed to handle matters from initial calls and qualification through case work and settlement; the company frames the benefit as enabling firms to serve more clients and expand revenue, not simply cut costs. a16z One Los Angeles firm reportedly increased its caseload from about 100 to 500 clients per attorney using Eve across intake, qualification, case handoffs, and settlement preparation. a16z
• Eve says it never trains on firms’ privileged case data; it employs attorneys to track case success on its platform, and its Auditor reviews cases overnight for missed information. a16z The company says it shifted from seat-based to outcome-based pricing early because firms value serving more clients, while expecting billable-hour pricing to remain useful for work whose value is difficult to know in advance. a16z
• The product is not yet fully autonomous: Eve says human touches remain, although its agents are live with some customers; the company is working toward end-to-end case handling and describes engineering capacity as a constraint. a16z Jay said Eve had helped “a couple hundred thousand clients” to date. a16z

• Cybersecurity: A PHP vulnerability was described as exposing 70% of websites, including WordPress; Kalif and Wiz reportedly used frontier models to identify it, and most of the issue had been patched by the discussion. Lightspeed Venture Partners Speakers said 24/7 AI agent swarms are shifting cyber from episodic incidents to continuous offense and defense, with defensive swarms expected and cybersecurity becoming a growing must-fund line item. Lightspeed Venture Partners They flagged startup opportunities in security tools for smaller vibe-coded enterprise projects, while noting that large companies already run agents and may not know what some are doing. Lightspeed Venture Partners
• AI for science: OpenAI’s reported recent results spanned math, computer science, and physics, including a wave result extending a problem solved in 3D to every dimension and a physics result addressing a proof gap described as roughly 50 years old. Lightspeed Venture Partners Separately, an OpenAI release of about 400 math results across more than 700 papers faced verification and quality caveats: about 42% of papers had computer-checked proofs, three had been pulled for errors, and some were criticized for missing credit. Speakers questioned whether the volume translates into field productivity or downstream innovation without human verification and understanding. Lightspeed Venture Partners
• World models: Google’s Genie 4 was presented as a more photorealistic, longer-running world model, with multimodal generation and interactive characters. Lightspeed Venture Partners The speakers framed world models as useful for physically realistic simulations, including robot training across varied environments and game development; Genie 4 also allows real-time scene changes. Access was limited to early testers, with availability for Ultra subscribers expected soon. Lightspeed Venture Partners
• Delivery and agentic commerce: A possible DoorDash acquisition of Instacart was described as alleged and speculative. Lightspeed Venture Partners The discussion cited pressure from delivery apps and major retailers, and the risk that agentic ordering could bypass Instacart’s advertising business; its grocery relationships and purchasing data could still be valuable, though grocery partners might view DoorDash as a less-trusted counterparty. Lightspeed Venture Partners

OpenAI published a paper claiming a subexponential-time algorithm for 3-SAT, which would refute the Exponential Time Hypothesis; the video says expert review was not yet complete, although the proof appeared increasingly convincing, and stresses that this does not resolve P vs. NP. Two Minute Papers The paper reportedly supplies the algorithm and says it both finds solutions and counts them; reductions could extend the result to problems such as delivery routing, radio-frequency assignment, and sensor placement, though the video cautions that such transfers do not always apply. Two Minute Papers

David Ulevitch argues AI tools could shorten an 18-month review to 18 weeks or less, saying this shift is beginning and will eventually become standard; he presents it as a forecast, not a completed result for the cited case. @davidu The case is an accepted new-license application for an American-owned and operated enrichment plant: the NRC declined to commit to an expedited 12-month review and estimated completion in April 2028 (18 months), citing anticipated workload and resource constraints, though it said it would consider efficiencies if General Matter provides timely, complete information and resources are available. @UnoMasReactor

The post labels its subject(s) “Creators of PStack and GStack” and announces a Lightcone episode with @poteto and @romanugarte_. @garrytan

YC says Lightberry (F25) is the first and currently only U.S. humanoid company able to ship robots using an existing supply chain @ycombinator. YC linked to Lightberry’s DoW-approval page @ycombinator.

GBrain Cloud is positioned as a way to run shared context across multiple AI-agent harnesses; the post groups Grok, Muse, and Dots as personal-use options, OpenClaw and Hermes for raw power, and Claude Code, Codex, and Capy for code generation. @garrytan The post also links setup instructions for using GBrain Cloud from Claude Code. @garrytan

Scott Kupor highlighted the power-law concentration of VC returns: 90% come from the top 5% of investors and 50% from the top 1%. @skupor The post links this to Stanford GSB Professor Ilya Strebulaev’s effort to compile the largest dataset of individual VCs to investigate why some succeed. @StanfordGSB

• a16z spotlights Eve’s AI for plaintiff-side law, which it says automates work from the first client call through settlement. One firm reportedly grew from about 100 to 500 clients per attorney; the pitch is that AI changes the economics of cases that lawyers often reject when they would generate less than $5,000 in fees. @a16z
• The interview frames legal software as labor-market automation: a case can involve 10,000 back-and-forths, while cheaper filings and possible agent-to-agent negotiations are discussed as potential shifts. @a16z
• Eve CEO Jay Madheswaran’s post claims the company has helped 350,000 people and that firms using it have recovered more than $3.5 billion; it also describes one client whose lawyers secured 15 times the insurer’s initial offer with Eve’s help. @a16z +1

a16z frames Eve’s opportunity as making plaintiff cases that generate under $5,000 in fees more viable, noting that most lawyers will not take such cases. @a16z Eve CEO Jayanth Madheswaran says one Los Angeles firm increased its caseload from about 100 to 500 clients per attorney using agents for intake, qualification, internal handoffs and case management through settlement; he expects lawyers to take on more data-science- and engineering-like work to improve these workflows. @a16z

• Jason (@jasonlk) proposed putting 20% of a fund into OpenAI at a $1.25T round, arguing that OpenAI had closed its competitive gaps and that the investment could return the fund; this is an investment thesis, not a realized outcome. @HarryStebbings The recap’s bullish case also called OpenAI mispriced and projected that OpenAI and Anthropic could each become $10–20T companies, citing Anthropic’s current 0.1% share of Fortune 500 budgets and a 2–3% conservative estimate. @HarryStebbings
• Anthropic’s S-1 was described as beating expectations on margins and operating income even after stock-based compensation, with strengths in customer concentration, compute efficiency and cash; Q3 revenue remained undisclosed. @HarryStebbings The recap said 80% of Anthropic revenue came from models less than six months old, while 70–80% of compute costs went to training; it argued that models’ six-to-twelve-month high-profit serving window and the $10–12B compute investment required could favor Anthropic if it remains aggressive. @HarryStebbings
• The recap argued that Chinese open-weight models faced US adoption headwinds, saying they were beginning to be banned across federal agencies despite DeepSeek’s latest raise. @HarryStebbings It also argued that CIOs and CISOs would reject a model if confidential data were sent to Alibaba servers without consent even once per 100,000 uses; this is presented as a risk scenario, not evidence that such transfers occurred. @HarryStebbings

Supabase says its open-source Postgres platform has 13 million users, launches more than a million databases weekly, and gets roughly 70% of its new activity from agents; it is rebuilding around agents as AI coding shifts toward agents building software for people @ycombinator. The company’s co-founder and CEO discusses designing products agents want to use and the prospect of personal software becoming as common as personal documents @ycombinator.

Jason shared an All-In episode agenda covering an Anthropic S-1 breakdown framed as “Greatest Business Ever?”, AI financial risk and an Anthropic “Doomer Showdown,” a segment titled around Google shedding $500B in market cap, and midterm predictions. @Jason +1

• The author describes “Black October” as a wave of critical vulnerability disclosures, including PHP2Shell and GitLab, and points to strained cyber first responders and an exponential rise in vulnerabilities—indicating an ongoing patching and incident-response burden, not a settled spike. Interconnects
• Epoch AI used GLM-5.3 in vulnerability research and found a Firefox exploit that could exfiltrate files when a user visited a website, without further interaction; the author cautions that deploying an exploit still requires additional steps, including setting up the malicious site. Interconnects
• The author says evidence that open models drove recent cyber harms is weak and that closed models played a substantial role; near term, open large models are costly to serve amid constrained compute, while closed models are more price-effective, though open-model costs may fall as compute supply grows. The author frames the response as a costly industry shift and notes that firms may both produce vulnerable software with AI tools and sell tools to fix it. Interconnects

YC launched its Early Access Network, bringing 30 companies and 100+ enterprise leaders together months before Demo Day; the network is now part of every YC batch, with W27 applications due Nov. 2. @ycombinator