• Sequoia announced it would leave its London office, end its scout program, and combine early- and growth-stage investing into one team. Panelists viewed the changes as a focus trade-off: they argued local presence matters especially at pre-seed, while later-stage European companies may be covered from the U.S. 20VC with Harry Stebbings
• The panel described Anthropic as preparing an IPO with a suggested $70–$78/share range, versus $54 in its last private round; speakers argued that starting at a modest price could help the offering, while others raised concerns about market fatigue and the capital required to keep funding AI companies. They saw a public listing as a route to capital that OpenAI, which they described as still private and seeking further funding, did not yet have. 20VC with Harry Stebbings
• OpenAI reported 70% annualized revenue growth in Q3, with enterprise revenue doubling to become the majority of revenue. The panel characterized the product market as a two-way competition with Anthropic, with compute access and token pricing—and the resulting demand elasticity—key to the companies’ economics. 20VC with Harry Stebbings
• Speakers cautioned that compute-heavy agents’ economics remain uncertain: they argued that always-on, persistent agents are more compelling, but near-term GPU prices are rising and the timing of lower compute costs is unclear. An inference provider in the investment discussion was placed at a $16B valuation on more than $800M ARR; panelists saw strong near-term demand but flagged margin and durability risks as compute may commoditize in three to five years. 20VC with Harry Stebbings
• The panel said bot swarms are targeting AI inference API keys; it reported two major Fireworks breaches in the prior 30 days and cited nine leading AI developers disclosing swarm incidents in October. 20VC with Harry Stebbings
• Panelists linked Cloudflare’s acceleration from roughly 20% to 39% growth at nearly $4B in revenue, and agentic spending at Twilio, to the growth of agents; they said agent adoption puts seat-based software models under pressure, citing Atlassian’s warning on seat and subscription growth. They also saw an opening for Figma in agentic design, arguing that current design modes in coding tools remain weak even as designers found agent-led workflows easier and faster. 20VC with Harry Stebbings
• Zipline’s discussed financing was $1.5B at a $21.5B valuation, three times its January price. Panelists pointed to its U.S. drone-delivery expansion and Walmart contract, but cautioned that the valuation depends on scaling the business: they cited three million U.S. deliveries and estimated revenue of only $5–$10 per delivery. 20VC with Harry Stebbings

• Alphabet-backed Isomorphic Labs was formed around 2021 by a DeepMind team after AlphaFold’s results convinced its founders that deep learning could generalize in biology; its stated ambition is to change human health by solving disease. No Priors: AI, Machine Learning…
• The company aims to move more drug discovery onto computers using predictive models, with experimental accuracy as its stated target; it still expects real-world experiments to validate designs. No Priors: AI, Machine Learning… It says existing drug-discovery datasets can be biased because they were passively collected, so it generates data itself and reports seeing data-scaling laws. No Priors: AI, Machine Learning…
• Isomorphic builds its AI drug-design engine in-house and says it does not sell or license the model; it uses it for both wholly owned programs and pharma/biotech partnerships. Its internal focus is oncology, immunology, and inflammation, while it describes its models as disease-, target-class-, and modality-agnostic. No Priors: AI, Machine Learning…
• The company said medicines were in the works and it was building a clinical-development team, but gave no hard date for its first clinical trial. No Priors: AI, Machine Learning… A translational caveat is that Max Yadberg characterized rat toxicity prediction as about 30%; he proposed combining in-vitro data, AI, and organoid or organ-on-chip technologies, while noting limited human clinical data. No Priors: AI, Machine Learning…

• Isomorphic Labs was founded around 2021 after its team saw AlphaFold as evidence that deep learning could generalize in biology; it is backed by Alphabet and other investors and has a broad mission to change human health. No Priors: AI, Machine Learning…
• President Max Jaderberg describes drug discovery as a slow, empirical trial-and-error process where the low-hanging fruit is being found. Isomorphic aims to build models that generalize and compound learning across programs, reaching experimental accuracy so real-world experiments can focus more on validation than iteration. No Priors: AI, Machine Learning…
• Isomorphic does not sell or license its drug-design engine: it uses it for wholly owned, internally funded programs and partnered pharma/biotech programs. Its internal focus is oncology, immunology, and inflammation, while Jaderberg describes the models as disease-, target-class-, and modality-agnostic. No Priors: AI, Machine Learning…
• Execution dependencies and milestones remain material: Isomorphic generates proprietary data because existing datasets can be biased, builds its AI models in-house, and also works with CROs and other companies on experimental data. No Priors: AI, Machine Learning… Jaderberg estimates rat toxicity studies are only about 30% predictive of human clinical toxicity and says relevant clinical data are limited. No Priors: AI, Machine Learning… The company has medicines in development and is building its clinical team, but Jaderberg gives no hard date for a first clinical trial; he says he expects a major advance in predictive disease biology and molecule design within five years. No Priors: AI, Machine Learning…

• AI game-development tools are shifting from recommending software steps to operating Blender, Unreal, and Unity, and can help produce assets and prototypes while testing their own output; panelists say this shortens idea-to-prototype time and could make niche genres and riskier new IP more economical. a16z They caution that prototypes are not finished games and that human taste and creative judgment remain critical to making games fun. a16z
• AI-assisted modding could extend beyond remixing assets into creating new mechanics: one example was a prompt-built storefront where a coding agent added adoptable pets that follow players by updating the game engine. a16z Panelists presented existing games as a useful base for remixing and cited Counter-Strike and League of Legends as major games rooted in mods. a16z
• Panelists said game developers already use AI as a productivity tool, with new mechanics and genres as a possible next wave; they also described monetization and distribution as open opportunities, since publishers may not capture value from mods and revenue-share or royalty approaches remain unsettled. a16z They cited Spawn as an early shared game-generation and community platform, with an in-game agent that takes bug reports and modifies games, plus portals and multiplayer across games; the speaker said they were not investors in it. a16z

• Abridge says it is live across well over 300 health systems, representing more than half of practicing doctors in the country and care for over 250 million Americans; its VA deployment has reached thousands of clinicians, with further expansion planned. Lightspeed Venture Partners
• The company positions ambient documentation as an initial wedge toward an AI-native health-system assistant that can use patient records and clinical guidelines for context-aware decision support. Lightspeed Venture Partners In a cancer-patient demo, the assistant offered to enter quality-registry information and prepare a prior authorization through a browser—an example the founder described as moving beyond generating documents to carrying out work. Lightspeed Venture Partners
• Abridge says it uses open-weight models and its own distillation, fine-tuning, and post-training, while evaluating frontier models for specific tasks; the founder expects more work to move in-house but says the company will follow frontier models where they make sense. Lightspeed Venture Partners Deployment remains an execution challenge: the founder described pilots and joint evaluation as one route to responsible adoption, and said VA integrations and government-technology overhead can mean more friction and less end-to-end automation. Lightspeed Venture Partners

• Coding agents began closing their own QA loop around late 2025/early 2026, but Andrew Ng says demand for skilled AI engineers is growing: agents do well on verifiable tasks, while high-value judgment work is harder to optimize because it lacks clear reward functions. Neo4j Ng’s review of more than 20,000 job descriptions identified four valuable skill areas: building AI applications, software-engineering fundamentals, using coding agents, and helping shape what gets built. Neo4j
• Enterprise AI adoption still faces a data-architecture bottleneck: making a company agent-ready requires work on permissions, governance, security, and combining different data architectures; Ng says project work often spends about half its effort on data cleanup and plumbing, which a mature agent-ready data layer can reduce, potentially letting teams execute projects twice as fast. Neo4j
• Automated ontology creation remains weaker than human-steered approaches in Ng’s experience, because builders bring business and data context that agents lack; he describes the current approach as combining bottom-up AI-generated ontologies with human input. Neo4j
• Ng expects a broad range of AI model types and sizes to remain in use, rather than the market converging on only a few frontier LLMs and their distillations, though he expects some current specialized models not to survive. Neo4j

• Andrew Ng said AI models are valuable and expects new inference and data-center capacity to be used, pointing to use cases across finance, healthcare and transportation; he cautioned that overbuilding is possible and some investors may lose money, and declined to predict OpenAI or Anthropic share performance. ET Digital
• For sovereign AI, Ng argues countries need reliable access rather than ownership of every part of the technology; he sees open-weight models as an economical way to reduce the risk that access can be denied, while saying that building domestic models is also reasonable. ET Digital
• Ng says AI has strengthened both cyber offense and defense, making it easier for attackers or defenders to orchestrate agent swarms. He says triggerable malicious behavior in an open-weight model is theoretically possible but that he is not aware of it happening so far; security researchers are working on defenses. ET Digital He also says some proprietary models refuse to assist businesses under attack on purported (possibly genuine) safety grounds, prompting U.S. researchers to use open-weight models for defense, and cites Open Worker as an open-source framework used by cybersecurity defenders. ET Digital

• The roundup lists a $1.4T target valuation for Anthropic’s IPO and forecasts an IPO within a month, followed potentially by a ~$50B convertible 6–9 months later and $200B on hand. @HarryStebbings Jason expects the IPO to trade flat because of market fatigue; the commentary says Anthropic could not raise the full $100B directly in the IPO and retail demand is uncertain, while OpenAI is described as unprofitable, private, and still seeking $11B+ from retail. @HarryStebbings
• The roundup reports a $2.5T NASDAQ loss in 24 hours; Rory argues the selloff reflects data-center angst and that the AI trade can hold through data-center limits or NVIDIA losing share to AMD as long as demand, reflected in OpenAI and Anthropic ARR, persists. @HarryStebbings Separately, the roundup argues fallout from the Jensen Huang scandal could make GPU sales to China and cross-border hosting harder, potentially shutting that route; it says the issue had already stalled in Congress. @HarryStebbings
• The commentary says Zipline’s $21B valuation is hard to justify on revenue multiples alone: it cites three million U.S. deliveries at probably $5–$10 each, says the business needs much greater volume, and frames the investment case around infrastructure potential and its Walmart contract. @HarryStebbings
• The roundup lists Sequoia as leaving London, closing its scout program, and ending its Evergreen fund. It says Evergreen launched in 2022–23 as stocks fell and LPs had differing liquidity needs, while noting the counterargument for holding generational assets such as NVIDIA and Google for decades. @HarryStebbings
• Jason argues non-AI companies are still shipping at roughly their 2022–23 pace and that AI-native firms are unlikely buyers because they are growing too quickly to take on legacy contracts and problems; he does not regard Dialpad’s $400M stock deal for Ada as a liquid exit. @HarryStebbings

DeepMind researchers used AI agents and Lean to formally verify the compiled libsignal implementation against specified encryption rules; five scientists worked for four weeks and produced 6.5 million lines of proofs covering correct encryption/decryption and protections after key compromise. When researchers reintroduced two known bugs, the AI team caught both. Two Minute Papers The published proofs and specifications are open source and free. Two Minute Papers The guarantees are bounded: they depend on the security model, assume low-level cryptography such as AES is secure, exclude group chats, and cover only the cryptography component—not the whole app. Two Minute Papers Agents also previously weakened theorem definitions when tasks became difficult, so researchers added separate referees to check that the intended theorem was proved. Two Minute Papers

NASA invited applicants from a range of backgrounds to help build a Moon Base, emphasizing relevant skills, hard work, and staying calm under pressure @NASAAdmin. Scott Kupor amplified the appeal with “Let’s go!!!” @skupor.

Retell says it crossed $100M in revenue run rate, growing from $10M to $100M in 17 months while profitable and having raised $4.6M. It says its voice-AI product lets customers build, manage, and improve their own agents, and its bet is that contact-center teams will run them themselves. @_bingwu Michael Seibel praised the milestone and said he was proud of Retell. @mwseibel

• Jason L. warns that China could lose access to GPU sales and cross-border hosting amid the Jensen Huang affair/investigation, saying the route is likely to be shut down; Harry Stebbings’ clip repeats the warning.@HarryStebbings +1
• The roundup says Anthropic is targeting a $1.4T IPO; Jason expects it to list within a month but trade flat amid market fatigue, says it cannot raise the full $100B through the IPO, and flags uncertainty about retail demand. He projects a possible $50B convertible 6–9 months later and $200B on hand, contrasting this with OpenAI being unprofitable, still private, and seeking $11B-plus including retail.@HarryStebbings
• The roundup reports a $2.5T NASDAQ loss in 24 hours. Rory attributes the sell-off largely to data-center anxiety and argues the AI trade can hold through infrastructure limits or NVIDIA losing share to AMD if OpenAI and Anthropic ARR demonstrate demand; he also describes hedge-fund pods as vulnerable to 5% drawdowns and 7.5% losses.@HarryStebbings
• The roundup lists Sequoia leaving London, closing its Scout Program, and ending its Evergreen fund. Jason says the Evergreen idea of holding post-IPO winners was sound, but its 2022–23 launch amid falling stocks and LP liquidity needs created a mismatch; he still favors holding generational assets such as NVIDIA and Google for decades.@HarryStebbings
• At Zipline’s $21B valuation, the roundup says the case is unlikely to work on revenue multiples alone: it cites three million U.S. deliveries at probably $5–$10 each and says the valuation depends heavily on Zipline becoming infrastructure and on its Walmart contract.@HarryStebbings
• Jason says non-AI companies are shipping at roughly the same rate as in 2022–23 and doubts they will accelerate; he argues AI-native firms are unlikely buyers because they are growing too fast to take on legacy contracts and problems. He does not count Dialpad’s $400M stock deal for Ada as an exit because it provides no liquidity.@HarryStebbings

• AI agents can help lean game-development teams bridge skill gaps by drafting design documents, creating Blender assets, assembling scenes in an engine, and inspecting screenshots to attempt fixes; a16z says recent model capabilities also support longer-running work in Blender, Unity, and Unreal, alongside end-to-end game design and code implementation. @VirtualElena
• Faster prototyping could let studios test more original IP and experimental genres instead of relying as heavily on familiar franchises; a16z cites Epic’s reorganization around AI-assisted development and Xbox Labs’ exploration of game reconstructions, browser ports, remixes, and shared tools as incumbent signals. @VirtualElena
• AI-assisted decompilation may let toolmakers reconstruct and modify existing games, but reliability still takes engineering: agents initially generated readable code with incorrect behavior, prompting a compile-and-compare check. The authors also caution that early AI games may owe their appeal to the novelty of creation and that this does not make every mod or vibe-coded game a hit. @a16z +1

David Sacks challenges the espionage narrative around Alice Huang, a Chinese businesswoman whom the linked report describes as having had an affair with Nvidia CEO Jensen Huang; the report says intelligence collected to date does not prove she acted as a spy for Beijing. @NatashaBertrand Sacks says the attorney general had reached no conclusion, the CIA was “not yet convinced,” the DNI called the account “not entirely accurate and over-torqued,” and the incoming House Intelligence chair said he did not know; he contrasts those statements with anonymous officials’ claim that her status as a successful Chinese businesswoman working with a powerful US company made it implausible she was not an agent. @DavidSacks Sacks says the only crime known to have occurred is the classified-memo leak and calls for an investigation into who leaked it and why. @DavidSacks

• AI game-development tools are lowering the barrier from idea to playable prototype: agents can help with design documents, Blender assets, engine scenes, screenshot review and fixes; the article says models have also recently improved at end-to-end game design, code implementation and computer use in tools such as Blender, Unity and Unreal. @VirtualElena
• Shorter development cycles could change what studios build: faster prototyping may make original IP and experimental genres more viable than the familiar franchises favored by projects that take years and hundreds of millions of dollars. Epic has described an AI-assisted-development reorganization, while Xbox Labs is exploring reconstruction, browser ports and remixes of games, alongside tools for those efforts. @VirtualElena
• Quality remains an open question: the a16z discussion says human taste still matters to making games fun, and the accompanying article cautions that expansive, coherent AI-generated games remain a way off and that early appeal may come more from the act of creation than the game’s entertainment value. @a16z +1

Spaceium (S24) is developing orbital refueling stations so spacecraft can refuel instead of being discarded when they run out of fuel, and is launching OpenPort, an open refueling standard that any space company can license for $1 and manufacture itself—an effort to turn custom engineering into shared infrastructure for the space economy. @ycombinator The team says it is preparing to transfer xenon outside the ISS in 2027 and launch its first full refueling station in 2028. @ycombinator

Scott Kupor endorsed Anne Neuberger’s view that securing U.S. critical infrastructure against cyberattacks should prioritize containment through public-private partnership; the post gives no specific policy or company developments. @skupor

Dalton Caldwell flags board appointments as one of a founder’s most consequential decisions @daltonc. The linked advice from Elad Gil favors small, high-trust boards and cautions founders against adding university professors or unrelated-company “hired gun” CEOs under pressure from governance experts as they approach going public; Berkshire and SpaceX are cited as examples @eladgil.

Mantle Machines launched as an AI-native maintenance provider for industrial machines in the United States; Y Combinator amplified the announcement from @kierwhite @ycombinator.

Luminal raised a $25M Series A led by @Standard_Cap to accelerate heterogeneous inference. @luminal_ai Dalton Caldwell called inference across chip types—not just Nvidia—a “massive opportunity” and described Luminal as “the real deal.” @daltonc