• Cloudflare says vulnerability reports rose exponentially over the preceding 3–6 months and incidents comparable to Log4j/Shellshock shifted from roughly annual to weekly; it expects security to dominate teams’ workloads over the next ~18 months. No Priors: AI, Machine Learning… Recent major vulnerabilities have included cases involving high-school-age users directing agents; Cloudflare expects to grow its red teams from one to four or five and sees AI as both an attack accelerator and a way to solve the security crisis faster. No Priors: AI, Machine Learning…
• Agent browsing is creating pressure for a machine-facing web and new monetization: an agent can visit 1,000 pages in the time a person browses one, and the interviewee described a model where agents could pay publishers based on their CPM through micropayments. No Priors: AI, Machine Learning… Cloudflare says it released a site-monetization gateway with agent wallets and a stablecoin, and is working with OpenAI on Signals to surface new content without repeated full-site crawling. No Priors: AI, Machine Learning…
• Cloudflare released Cloudflare OS as a forkable reference architecture for governed internal AI use and teaching agents company practices; the company says GTM and G&A staff use it more than engineers, and cites replacing a three-team customer-email workflow with a product-manager-led tool and manager approval. No Priors: AI, Machine Learning…
• Cloudflare said it decided to hire 1,000 interns at a company of about 5,000 people after seeing managers of a prior intern cohort adopt AI faster. No Priors: AI, Machine Learning…

• AI-assisted security: Companies including Google and OpenAI searched for vulnerabilities, and the discussion says LLMs helped identify a decades-old flaw in PHP. PHP underpins content-management systems such as WordPress and Drupal and was described as present on tens of millions of sites; urgent fixes took some websites offline, while Cloudflare’s CEO reported millions of attempted hacks. TechCrunch The episode frames the risk as dual-use: LLMs can methodically find vulnerabilities, but can also help attackers exploit them; speakers saw potential demand for defensive cybersecurity tools while leaving the viability of startup solutions uncertain. TechCrunch
• Otter AI raised $320 million from General Catalyst after raising around $50 million previously; hosts worried that pressure to grow could bring more AI features at the expense of the core transcription experience users value. TechCrunch
• Character.AI sought bankruptcy protection, with Disney acting as stalking-horse bidder and committed to hiring most of its workforce and potentially acquiring assets; the asset outcome remained uncertain and other bidders could emerge. Character.AI’s former CEO had already become Disney’s first CTO. TechCrunch Hosts cautioned that this may not be a broad signal about AI companies: they linked Character.AI’s trajectory to Google’s earlier hiring of its founders and top talent, which left the remaining company on what they described as an unsustainable path. TechCrunch
• OpenAI acquired biotech startup Orchestra Bio; its purchase price was not disclosed, and the startup was said to have raised “maybe” about $20 million. Hosts viewed the deal as a sign OpenAI may pursue science-adjacent businesses for commercial potential, despite earlier discussion of avoiding “side quests.” TechCrunch
• Tesla and SpaceX set up special committees to review the idea and feasibility of a merger; hosts speculated SpaceX could absorb Tesla. TechCrunch Although the committees were described as independent, hosts questioned that independence given board, business, and personal ties; they also noted potential share dilution and Musk’s statement that approval would require 50% of all shareholders, not merely 50% of votes cast. TechCrunch A possible financing rationale discussed was that Musk’s ambitions exceed what venture markets can support and a combination could improve access to institutional financing. TechCrunch
• Anthropic announced a plan to put 5% of its equity into a public-benefit fund, with the stake growing according to milestones; the fund’s structure and details were still unclear and described as an LLC. TechCrunch

• Chamath estimated that 1% of companies account for 90% of AI spending, with 80% of that spend coming from unprofitable, VC-backed startups; a Ramp researcher reportedly said the estimate was “pretty much correct.” All-In Podcast Chamath warned that VC-funded startups spending on AI and then raising on the strength of their resulting revenue could create circular valuation signals. Jason countered that startups’ willingness to spend can also signal that AI offers the best, cheapest, or fastest competitive edge. All-In Podcast
• A preview of Ramp data showed top AI spenders increasing their GPU-compute spending; Chamath interpreted this as a shift toward buying GPUs and using open-source models. He also described enterprise diffusion as slow, with some companies moving toward their own data centers or virtual private clouds. All-In Podcast Palo Alto Networks CEO Nikesh Aurora said the company prefers open-weight models for modification, training, explainability, and auditability, but said resilient global inference outside the hyperscalers was difficult to obtain at viable prices. He expected hybrid infrastructure, including bare metal from hyperscalers or compute aggregators, amid tight compute availability. All-In Podcast
• The panel reported that security researchers using AI found two critical PHP exploits, rated 9.2/10; exploit information leaked before the planned patch, after which AWS and Cloudflare reported increased attack activity and some organizations took systems offline to patch. All-In Podcast Aurora said persistent compute helped solve a vulnerability that had been identified years earlier, while legacy unpatched systems and fragmented infrastructure hinder defense. He said enterprises were using Mythos or equivalent models to test their defenses and emphasized aggregating signals to detect anomalous behavior as the vulnerability-and-patching cycle continues. All-In Podcast
• Aurora said personal agents often access bank websites using users’ credentials, making agent activity difficult for banks to distinguish from the user’s; panelists identified liability and accountability as unresolved issues. All-In Podcast The discussion cited a consumer-agent constraint: after Amazon blocked shopping bots from completing purchases, Grok’s shopping bot could prepare a cart but not check out. All-In Podcast

• Perplexity began with three AI-researcher co-founders before ChatGPT 3.5; Srinivas says they were first-time entrepreneurs who learned startup operations incrementally. Cambridge Union The first product grew out of his difficulty researching U.S. health-insurance setup through Google and SEO-heavy sites; he says the team kept improving it when it could not answer their own questions. Cambridge Union
• Perplexity introduced Model Council, which sends the same question to three models for research and uses a chair model to assess agreement and disagreement before producing a final answer; Srinivas framed it as extending academic inquiry, not replacing peer review. Cambridge Union
• For consumer AI, Srinivas recommends tracking repeat use at seven and 30 days and growth in users active two, four, or seven days per week; he cautioned that free trials are weak demand evidence and called retaining at least half of paying customers after six months a product-market-fit signal. Cambridge Union
• Srinivas characterized fake social content, engagement-driven attention capture, and sycophantic companions as more immediate AI risks than job replacement; he said layoffs he observed were overhiring-related and claimed AI users at his and other companies were functioning 20–30% more than without it. Cambridge Union He advocated independent evaluations and audits, company evaluation and red-teaming teams, and some regulation for personal AI companions and social-app AI. Cambridge Union
• He framed a key continual-learning challenge as identifying missing information, acquiring it, and running experiments over long horizons; he suggested research agents capable of real-world experiments could help address it. Cambridge Union He also forecast that frontier-grade intelligence might run on local devices within four to five years, through further self-improvement and efficiency gains. Cambridge Union
• For founder-investor fit, Srinivas advised choosing capital from investors who let founders preserve independence and direct the company, rather than investors who micromanage. Cambridge Union

• After joining Meta in summer 2025, Wang said he arrived after the company had described Llama 4 as off its needed trajectory; Meta built a small, talent-dense research lab through internal and external recruiting, with a focus on scientific discovery. The lab’s model series includes Muse Spark and Muse Wonder, and Meta delayed Muse’s launch for months while iterating on the model and its privacy and security architecture. TBS CROSS DIG with Bloomberg Muse’s safeguards include a secure virtual-machine sandbox, a separate sentinel agent that reviews proposed actions and information sharing, and human approval for many actions. TBS CROSS DIG with Bloomberg
• Wang framed Meta’s opportunity as consumer personal agents while much of the AI industry focused on coding agents and B2B or enterprise use. TBS CROSS DIG with Bloomberg For this product focus, Meta invested in multimodality, active memory and focus, and built a personal-superintelligence index because existing benchmarks did not capture many relevant agent capabilities; its measures include tool chaining, personality, discretion, failure tolerance and browser use. TBS CROSS DIG with Bloomberg
• Wang called the response after Muse’s US and Canada launch better than expected, citing broad adoption and users saving time and money; he said Meta was launching in Japan, supported by years of investment in multilingual models. TBS CROSS DIG with Bloomberg
• Wang outlined subscriptions and additional token purchases, interest from small businesses and enterprises, and shopping activity as Muse business opportunities. TBS CROSS DIG with Bloomberg He also cited Meta’s recommendation and ad systems as a way to use compute with strong returns, alongside Meta AI and enterprise offerings; he expects efficiency improvements to support serving more Muse users and countries with the same level of compute. TBS CROSS DIG with Bloomberg

• Peter Levels reportedly pivoted to AI apps in 2022 and now argues against SaaS, favoring moats such as proprietary data or content, network effects, audience or distribution, and regulation. The hosts called the thesis potentially alarmist and countered that brand and network effects remain valuable; they also favored businesses where AI supports rather than replaces the core offering, potentially lowering costs for physical or in-person businesses. My First Million
• James Currier’s founder advice, as relayed by the hosts, is that fear and avoidance are high-leverage constraints: addressing them can help founders make faster decisions and simplify products, especially when cash pressure creates an “effort moment.” My First Million
• A Tesla Roadster demo featured SpaceX rocket boosters and was said to be coming the following year, but the hosts noted the car had been teased for about a decade and that the demo offered no firm delivery calendar or proof of timing. My First Million
• Influencer Clavicular pitched a men’s skincare and makeup line and asked interested people to contact him; the hosts saw influencer reach as a possible way to build demand, but stressed the need for strong behind-the-scenes operators, citing Congo Brands’ founders and their work with Alani Nu and Prime as a contrast to a celebrity running the company. My First Million

• AWS CEO Andy Jassy said AWS’s AI business exceeded a $25B revenue run rate in its first three years; Amazon expects about $220B in 2026 capex and still cannot meet all demand. AWS describes 2026 capacity as sold out, 2027 as largely sold out, and a good chunk of 2028 as already sold; Jassy expects component constraints to persist for several years. a16z Data centers require investment roughly two years before servers can be installed, while server and networking purchases are made closer to demand; rapid expansion therefore weighs on free cash flow until capacity is online and monetized. a16z
• AI demand is “barbelled”: labs and a few breakout apps consume very large amounts of compute, while many enterprises use AI for cost avoidance and productivity; most enterprise production workloads are not yet using inference pervasively, though Jassy expects it to spread across many applications over five to seven years. a16z He says inference remains an order of magnitude too expensive for pervasive use and describes Amazon’s Trainium accelerators as roughly 30–40% more price-performant than comparable processors, making lower inference costs important to broader enterprise adoption. a16z
• Amazon says its investments in Anthropic and OpenAI are part of deeper partnerships: Anthropic workloads run on Bedrock and the companies have worked together on Trainium, including a multi-year, multi-gigawatt Trainium commitment; Amazon and OpenAI built an offering for OpenAI on Bedrock and a stateful runtime for agents. a16z
• Jassy expects smaller, cross-functional teams and fewer people doing many tasks performed today; he says AI-native startups are beginning with agent-native approaches and that slower-moving incumbents may be more vulnerable. a16z He flags cybersecurity as an immediate AI-era pressure: models are finding more vulnerabilities, and applications will need to respond to threats and deploy patches in minutes. a16z

• Amber Yang said she is focused on AI compute and inference constraints, including opportunities in alternative energy, power, data centers, and matching data-center capacity with AI companies. She described the path from land to usable compute as disconnected and opaque, and said she is looking for a vertically integrated provider—her “Standard Oil” analogy. Lightspeed Venture Partners
• Lightspeed-backed Exa launched Exa 3.0, which it called its largest improvement to search quality, and Exelite, priced at $4 per 1,000 requests; Lightspeed said it led Exa’s seed and Series A rounds in 2024. Lightspeed Venture Partners
• Goodfire researcher Nick Camarada published a working hypothesis that most model neurons have a single readable purpose—he said 90%—rather than being entangled. The cited evidence was his own experiment compressing one layer from 480 neurons to 10, with signals mostly remaining in their original domains; the discussion noted that no formal research paper had yet been published. If supported by further work, the claim could matter for interpretability and AI safety. Lightspeed Venture Partners
• DeepSeek and MiniMax released efficiency-focused model updates; MiniMax claimed 70% lower usage costs, and the models were described as performing well on benchmarks. The guest said U.S. closed models remained ahead “as far as we know,” while American companies were already using Chinese open-source models; the episode also reported a U.S. order for civilian federal agencies to ban DeepSeek. Lightspeed Venture Partners
• A reported hacker used AI agents to breach more than 100 company websites, including John Deere and SharkNinja. Thousands of agents ran the same prompt independently rather than coordinating as a swarm; the discussion contrasted this with hierarchical swarms that share context and could handle more complex tasks. Lightspeed Venture Partners

• Altman thinks AI may be nearing the end of an era dominated by ever-larger models: parameter counts may still rise, but capability is the goal, and multiple smaller models are an option. He also leaves open the possibility of a new paradigm that obsoletes current players. Imagination in Action
• In a discussion where the interviewer floated $50–100 million for foundation-model training, Altman said OpenAI does not disclose its figure but that it is “much more than that”; the exchange signals substantial capital needs without establishing an exact spend. Imagination in Action
• For founders, a platform shift is no substitute for fundamentals: Altman says to build products people love, stay close to users, develop a durable long-term advantage, and use tight feedback loops rather than relying on speculation. He contrasts fragile App Store-era plays with companies that endured by building durable businesses. Imagination in Action
• Altman expects AI to contribute more to the model-improvement loop while humans remain in control for a long time; he also points to data-center construction and power access as constraints on abrupt takeoff. Imagination in Action
• On AI safety, Altman said OpenAI waited more than six months after training GPT-4 before release for safety study, external audits, and red-teaming. He argued that safety rigor should rise with capability, while criticizing the six-month-pause letter as lacking technical nuance and not being the optimal approach. Imagination in Action

• Hey Clicky is a cursor-following AI teacher designed to teach users anything. Founder Farza Majeed said an early demo drew nearly 3 million views and 7,000 downloads, but the more important signal was that about 100 users kept using it daily for five days; he spoke with those users and said their usage—not virality—convinced him to continue building. Y Combinator
• Majeed reported sharply different acquisition results by channel: a YouTube video with about 20,000 views converted at roughly 20% to app downloads, while a 2.9-million-view Twitter video had very low account conversion. Targeted TikToks averaging about 4,000 views generated around 50 design-agency leads, and he said Instagram accounted for about 70% of Hey Clicky’s traffic. Y Combinator
• Majeed said Hey Clicky’s first two hires were a filmmaker and an engineer; at one point, three people were making videos alongside one engineer. He argued that software could be cloned quickly, while the audience and community around a product were harder to copy. Y Combinator

• AI megafunds are reshaping venture financing, but face constraints. Anthropic’s fundraising pushed Menlo Ventures, Lightspeed and ICONIQ to raise funds, seek LP co-investment and market SPVs; its then-pending S-1 was expected to show a cap table dominated by corporate investors and late-stage megafunds. Newcomer OpenAI was reportedly in talks for a $30 billion fundraise at a $1.25 trillion valuation. Newcomer High interest rates have weakened M&A, IPO markets are receptive mainly to the strongest companies, and many LPs have paper gains but little cash returned—raising questions about how long larger funds can keep growing. Newcomer Mid-tier firms can be shut out or get access at inflated prices, though Sound Ventures and Factorial Funds each invested tens of millions in Anthropic’s Series C after major Sand Hill Road firms passed. Newcomer
• AI deal activity remains large, alongside new early-stage capital. Flapping Airplanes raised a $1 billion Series A at a $6 billion valuation, co-led by Lux Capital and Index Ventures. Newcomer Legal-AI startup Legora raised $550 million in a Series E at a valuation above $8 billion, led by Atreides Management. Newcomer ASI raised $325 million in a Series C for its AI asset-management system for military and airline customers. Newcomer Pigment raised $250 million in a Series E, with an additional Nvidia investment planned. Newcomer Manufacturing-AI startup Arda, Bob McGrew’s venture after OpenAI, raised $232 million in a Series A. Newcomer AI note-taker Otter secured $320 million in non-equity financing from General Catalyst’s Customer Value Fund. Newcomer Warhound raised $200 million in seed funding to genetically engineer warfighting dogs. Newcomer Inspired Capital raised $450 million for its fourth early-stage fund; Bison Ventures raised $160 million for its second early-stage fund, focused on the physical world. Newcomer
• Stanford’s Ilya Strebulaev and Ohio State’s Blake Jackson ranked a16z first among 50 AI-focused VCs, followed by Lightspeed, Menlo Ventures, Sequoia and ICONIQ Growth; the ranking considered valuations, growth, dilution, capital distributions and firm value-add. Newcomer
• Data-center delivery remains a key uncertainty. SemiAnalysis forecast 50 GW of capacity coming online in 2027, while Altimeter’s Brad Gerstner estimated about 25 GW; IREN said customers prepay for about half its GPUs and debt covers most of the rest, but whether chips and facilities can physically be operating by 2027 remains the question. Newcomer
• Meta’s Q3 revenue of $65 billion beat a $63 billion estimate, but $35 billion in capex, negative free cash flow and falling margins weighed on sentiment; new non-voting Class C shares would let it raise capital for AI without diluting Mark Zuckerberg’s voting control, and the stock closed down 7%. Newcomer
• AI models helped Calif and Wiz independently discover two serious PHP flaws that could enable remote control of unpatched servers; AWS research indicated at least one attempted exploit before public disclosure. Newcomer
• Bloomberg reported that Polymarket investors demanded CEO Shayne Coplan repay company-paid living and private-jet expenses before another check; Coplan and allies disputed the account, saying he had accounted for and repaid expenses and that the total was below a typical software engineer’s salary. The report came as Polymarket was losing market share to Kalshi. Newcomer

• Sponsored AI infrastructure: Reducto says its document parser serves AI teams at Scale AI, Harvey, Toast, and Vanta and processes more than one billion pages monthly. Its r-1 model combines layout detection, reading order, tables, formatting, grounding, and citations; Reducto claims up to 20% fewer parsing errors and a 1¢-per-page all-in cost. Not Boring by Packy McCormick
• Precision oncology: The FDA granted accelerated approval to dostarlimab (GSK’s Jemperli) as a standalone first treatment for stage II and III mismatch-repair-deficient rectal cancer; in GSK’s 154-patient trial, 71% had no progression after a median 26 months. The eligible subtype accounts for about 3% of locally advanced rectal cancers. Not Boring by Packy McCormick Coherence Neuro announced a $45 million Series A from Founders Fund and OpenAI to develop cancer therapy using neurotechnology to read and modulate electrical signals around tumors; human trials in glioblastoma patients are underway, with pancreatic cancer among the planned next targets. Not Boring by Packy McCormick
• Nuclear deployment plans: The Nuclear Company proposed Prosperity One, a Kentucky campus of up to six AP1000 reactors totaling about 6.6 GW; it aims for 2028 NRC approval and plans to build two units at a time, with the company’s chief nuclear officer calling 30% the floor for expected successive-unit savings. Not Boring by Packy McCormick Separately, the Navy selected Radiant’s factory-built 1 MW Kaleidos microreactor for Naval Weapons Station Crane, with a 2028 target from the Navy; it would be the Navy’s first shore-based advanced reactor. Not Boring by Packy McCormick
• Defense and engineered biology: Warhound launched with $200 million from investors including In-Q-Tel, Sequoia, and 8VC, at a reported $1.2 billion valuation. It plans to combine licensed Colossal Biosciences technology with Anduril hardware and software, including gene-edited working dogs intended to improve health and performance; the company says animal welfare is central and cites American Humane certification. Not Boring by Packy McCormick
• Agricultural bioelectric sensing: Electric Plant emerged from stealth with $11.2 million in funding and Bombadil, a foundation model that reads plant electrical signals to estimate seven types of stress. The model was pretrained on more than four million hours of recordings from 100 species; the company thinks its models and sensors could cut farm water use by 20% or more. Not Boring by Packy McCormick
• Fusion-enabled applications: Marathon Fusion reported converting mercury into stable gold at ten times CERN’s 2025 rate, while the article notes this remains a demonstration and that CERN’s comparator produced about 29 picograms over four years. The company says it has started making copper-64 and copper-67 for cancer imaging and targeted radiotherapy; an outside lab certified its plasma centrifuge’s lithium-6 enrichment, and Marathon says a new run beat a 40-year-old Soviet record. Its next step is a San Francisco fusion reactor designed to fire a trillion neutrons per second into a one-centimeter target continuously. Not Boring by Packy McCormick
• Consciousness research: Doris Tsao’s team found that neurons in face-processing regions of two macaques alternated roughly every 60 milliseconds between representing a seen face and a suppressed one; both could be reconstructed from neural activity. The mechanism explaining why only one face enters awareness remains unresolved, but the result offers a way to investigate that question. Not Boring by Packy McCormick

Garry Tan attacked Connie Chan as “anti tech, anti housing, anti prosperity” and criticized Pelosi’s political backing of Chan and Gary McCoy, framing the dispute around tech and housing policy. @garrytan Tan shared a post quoting a report that, among 84 land-use votes Chan cast since joining the board in 2021, her record was second only to Aaron Peskin’s in opposing housing liberalization; it contrasted this with Scott Wiener authoring 186 bills in the state Senate from 2017–18 through 2025–26, 97 of which were signed by the governor. @kimmaicutler

Yann LeCun argued that online adaptation in embodied systems is feasible without a major new conceptual breakthrough: a predictive model can adjust within milliseconds when an object is heavier than expected. Health-Wealth He described the mechanism as self-supervised correction of prediction errors, requiring no task, reinforcement signal, or value function, and characterized it as generic rather than task-specific. Health-Wealth

YC featured Farza, founder of Heyclicky and previously a builder of Buildspace, sharing a launch-video playbook; the post says he has made videos for over a decade and his posts regularly get millions of views. @ycombinator The interview covers launching constantly, shipping demos, showing the building process, turning demos into stories, and figuring out which social platforms convert. @ycombinator YC also challenges the assumption that a successful launch video requires a tens-of-thousands-dollar budget and high-production commercial style. @ycombinator

OrbitalOps (W25) was awarded up to $35M by the DoD to fly a full-scale in-orbit demonstration of Astraeus, its cryogenic maneuvering spacecraft. @ycombinator Unlike a typical rocket stage, which lasts about a week in orbit before its propellant boils away, Astraeus uses onboard refrigeration to keep liquid hydrogen and oxygen cold for years, allowing it to wait in orbit and maneuver within hours of a command. @ycombinator

Garry Tan called Grep’s open-sourcing of major parts of its harness a big deal. @garrytan Grep says its improved workflow DSL helps turn agentic loops into more deterministic recipes and has significantly reduced agent costs without sacrificing performance. @MiguelriosEN The learning loop is not yet open-sourced because separating Grep’s business logic still needs work. @MiguelriosEN

Amazon CEO Andy Jassy says AWS’s generative-AI business exceeded a $25B annual run rate within its first three years, versus $60M for early AWS after three years—a several-hundred-fold faster start. @a16z Amazon’s AI growth is driving a large part of its expected $220B in 2026 capex @a16z, yet Jassy says that spending will still leave meaningful customer demand unmet. @a16z

Martin Casado shared @jaydrogers’ warning that a GitLab-related bug had moved upstream into Ruby; the post reports “Multiple 10.0 CRITICAL exploits” and urges users to patch, linking to Ruby’s regexp-fixes notice. @martin_casado +1

• Andy Jassy describes AI demand as a barbell: AI labs and a few runaway apps on one end, enterprise cost-cutting and productivity uses on the other; he sees the largest absolute opportunity in existing enterprise production workloads, where most companies are not yet using inference pervasively. @a16z
• Jassy says inference remains about 10× too expensive for widespread use; lower costs would reduce the hurdle to rewriting CPU-based applications for AI accelerators and could make applications more capable. @a16z
• The accompanying a16z post says Amazon’s AI business reached a $25B run rate in three years, growing several hundred times faster than early AWS, and that Jassy expects Amazon to spend $220B this year to win AI. @a16z