• Venture funding is highly concentrated: investors said strong metrics may not secure funding for market-specific software perceived as vulnerable to AI, while ultra-high-upside areas such as data centers in space and quantum still attract capital; founders outside the narrowest investment theses may need to meet roughly 100 VCs to find a willing investor. 20VC with Harry Stebbings
• Anthropic’s reported Q3 revenue was $16.8B, up 45% quarter over quarter, with an 85% gross margin and adjusted profitability; the IPO valuation discussed was reportedly $1.5T rather than the previously suggested $2T. 20VC with Harry Stebbings Investors warned that model share can shift quickly and that slower end-user demand could intensify price competition. 20VC with Harry Stebbings Replit reportedly made OpenAI its default over Anthropic models because output quality was roughly comparable at 30–40% lower cost, illustrating the cost sensitivity of model-routing customers. 20VC with Harry Stebbings
• The panel described code review and “build to production” as major opportunities beyond code generation: generated code still needs substantial review, testing, and production work, while products close to foundation models also face recurring platform-competition risk. 20VC with Harry Stebbings
• Neocloud providers such as CoreWeave face capex and debt exposure: demand must grow fast enough to absorb capacity commitments, leaving the business sensitive to growth concerns. 20VC with Harry Stebbings In contrast, panelists viewed inference services as more value-added than simply renting GPU capacity; Modal’s announced Series D was $750M at a $15.75B valuation, though competition and eventual margins or multiples remain uncertain. 20VC with Harry Stebbings
• Berkshire Hathaway bought $330M of new Bending Spoons shares; panelists framed the company as a long-term software holding company rather than an asset flipper, arguing that it was improving acquired assets and adding higher-quality acquisitions. 20VC with Harry Stebbings

• a16z announced that it is expanding its crypto fund mandate to A16Z Frontier, targeting novel “outer-planet” opportunities outside established verticals; successful themes could develop into a new practice or fund. a16z The speakers cited stablecoins becoming part of fintech toolkits, including Stripe’s, and Wall Street’s move into tokenized assets as evidence that crypto is blending into the broader tech landscape. a16z They said Frontier would favor being early and accept category risk; as a timing caveat, they noted that VR remained short of mainstream adoption 13 years after Oculus was acquired, despite their continued conviction in the technology. a16z
• The speakers said the gap between AI-native startups and typical incumbents is widening, and expect founders to build more technically ambitious, cross-sector companies faster, with startup formation rising. a16z They also said employees may be adopting AI faster than their institutions. a16z As a potential shift in team economics, they suggested a 30-person team could plausibly build a full-stack device company where around 1,000 people might have been needed a decade earlier, and that AI could make hardware modification and home automation more tractable for people with software skills. a16z
• The speakers described open source as advancing rapidly and saw open-weight and local models as potential routes to a more decentralized AI market alongside large labs; they also identified stablecoins as a potential payment rail for agent economies. a16z They framed open AI as policy-sensitive: they described U.S. and Chinese support while warning of government restrictions, including EU ambitions, and argued that broader access could enable wider safety research, including work on reward hacking. a16z
• They argued that AI-generated answers and platforms retaining users are weakening the web’s referral bargain, citing sharp traffic declines at sites such as Stack Overflow and warning that ad- and conversion-dependent businesses face challenges; they see an opportunity to redesign internet economics, distribution, and knowledge sharing. a16z

• Cogent Security, led by Vineet Edupuganti after four years at Abnormal Security, is building AI to identify and address security vulnerabilities; Edupuganti says it has raised $53 million, with Greylock and Bain Capital among its backers. Stanford eCorner
• Before incorporating, Edupuganti used his cybersecurity experience and about 100 customer conversations to investigate prior solutions’ failures, CISO buying criteria, pricing, and pilot requirements; prospective customers’ early interest was verbal and conditional on the product working. Stanford eCorner The three cofounders had worked together before, with roles divided across engineering leadership, technical strategy, and product/go-to-market. Stanford eCorner
• Edupuganti’s thesis is that AI lets attackers exploit vulnerabilities faster, increasing the need for AI-enabled defense; he says resistance to fully autonomous defense has shifted to inquiries from Fortune 500 and large energy companies. Stanford eCorner Cogent says its context graph models enterprise assets for agents, with controls and human governance limiting write actions—for example, allowing remediation on laptops but not production systems. Stanford eCorner
• As a caveat to enterprise AI enthusiasm, Edupuganti says realized ROI is often low because organizations lack the people/process prerequisites or contextual data foundations. Stanford eCorner He also flags identity and access for AI agents as an attractive problem, projecting that agents could outnumber humans 1,000:1 and will need to collaborate safely. Stanford eCorner

• George Kurtz says AI agents have flattened the attacker capability pyramid, giving activists capabilities once associated with nation-states; he reports more vulnerabilities and faster attacks, with attacker breakout averaging 29 minutes and as quick as 27 seconds. No Priors: AI, Machine Learning… He says attackers are using open-weight models because frontier models often refuse cyber-related requests. No Priors: AI, Machine Learning…
• Enterprise agent proliferation is creating a visibility and identity gap: companies often do not know where agents are running or which identities they use. No Priors: AI, Machine Learning… CrowdStrike says it can discover shadow AI, trace agent activity—including subagents and tool calls—and apply policy or blocking rules; its Safe Mind system pairs red/blue security models with digital twins of customer environments. No Priors: AI, Machine Learning…
• Kurtz says CIOs consider open-weight models that are three months behind capable enough for many tasks, reserving frontier models for high-value work. No Priors: AI, Machine Learning… He claims CrowdStrike’s model costs about $0.03 for general remediation versus about $10 for a top-tier frontier cyber model, citing its security data and post-training. No Priors: AI, Machine Learning…
• Kurtz expects open-weight models to become a major foundation for AI, with smaller models running on phones and desktops and specialized chips improving speed, cost, and efficiency; he sees local and hybrid AI as opportunities. No Priors: AI, Machine Learning…

• AMI Labs, Yann LeCun’s Paris-based startup, announced a $1.03 billion pre-launch round at a $3.5 billion pre-money valuation, described by the moderator as the largest pre-launch AI round in European history. Its contrarian thesis is that scaling LLMs will not reach human-level intelligence and that world models learned from video and physical interaction are a better path. Radical Ventures
• LeCun defines a world model as predicting how the world changes given its current state and an action; he argues JEPA makes predictions in an abstract representation that discards unpredictable pixel-level detail, unlike video-generation approaches that can produce plausible-looking clips without capturing underlying dynamics. He argues world-model planning could make agents intrinsically safer than LLM-based systems. Radical Ventures
• He identifies industrial process control as a potential low-hanging-fruit market for world models and describes physical AI, including robotics, as an important opportunity. His proposed robotics learning path uses abundant video for pretraining, a much smaller amount of action-conditioned world-state data, and task-specific fine-tuning; he argues imitation learning requires enormous diverse datasets yet still lacks reliable action-consequence understanding, while simulation is weak for some tasks such as dexterous manipulation. Radical Ventures
• On AI infrastructure economics, LeCun says LLM-driven AGI is not just around the corner and that it is unclear whether AI revenues will repay current capex. He suggests world models may need less memory, but makes no firm efficiency forecast; he also says data movement in current CMOS-based systems remains a major energy cost. Radical Ventures

• Ng expects AI models and inference capacity to remain valuable as businesses develop applications across finance, healthcare, transportation and other sectors; he believes capacity built will be used, while acknowledging that overbuilding and investor losses are possible and declining to predict company share performance. ET Digital
• Ng argues that open-weight models offer countries an inexpensive way to preserve access to AI without building every component themselves. He also alleges that some lobbying and fear messaging against open models reflects proprietary-model companies’ investment and fundraising incentives. ET Digital
• Ng says software-engineering job postings are up and his teams want more skilled software and AI engineers; he would not hire a software engineer who cannot use a coding agent, and expects the profession to change rather than disappear. ET Digital
• Ng cautions that students using chatbots to complete homework may earn better homework grades but learn less and perform worse on closed-book exams; he distinguishes work-completion tools from AI designed to teach, citing his Learn Vector team’s focus on scalable learning experiences. ET Digital

• Wabby, founded by Eugenia, launched a personal agent for life logistics that works in individual and shared threads and creates task-specific mini-apps. After its earlier product let users make and share apps, the company says it learned that people liked personal software but did not necessarily want to build or discover it; the agent now builds what it infers they need from their goals and context. TechCrunch
• Users can invite others into threads and grant connector permissions per chat; Wabby says participants can see what information is requested and the response. Founder Eugenia says current Gmail onboarding reads mail to seed threads but does not store it; an upcoming version will let users skip access, with a less personalized initial experience. TechCrunch
• TechCrunch said Wabby had raised $20 million in a pre-seed the prior year and was in talks to raise again; Eugenia cited investor inbound, a 15-person team, and a focus on reducing costs without lowering quality. The host described the product as newly open beyond invite-only; it is free for now, with premium pricing or transaction fees among options under consideration, not a settled model. TechCrunch
• Competitive context cited by the host: Muse had 2.8 million installs and an estimated nearly 700,000 US daily active users as of late September, while Instinct had raised a $1 billion Series C at a $10 billion valuation. Wabby’s founder argued that group chats and routine app use are often low-compute compared with complex, multi-step agent tasks; this is the company’s cost rationale, not reported cost data. TechCrunch

• Ng described agentic AI as systems that take multiple steps—such as research, data lookup, reflection, and editing—instead of returning a one-shot answer. He said business applications are growing and that his firm works with businesses implementing these workflows and seeing ROI, though identifying use cases and deploying them with skilled engineers remains hard; he expects that work to continue even a decade from now. NDTV World
• Ng argued that AI’s business opportunity extends beyond cost savings to better customer service, new products and services, and revenue growth. He said some CEOs emphasize cost savings in public-market communications because they are easier to estimate and manage, creating an overemphasis on savings despite growth ambitions. NDTV World
• On education, Ng cited evidence that AI use can improve homework scores while lowering closed-book exam scores, and referenced a study reporting 17% lower grades among students using AI. He argued that personalized, one-to-one learning experiences beyond chatbots are an opportunity for new technology. NDTV World
• Ng argued that some businesses have incentives to amplify AI fears for PR and fundraising, and to seek regulation that could restrict cheaper, freely available model alternatives; he characterized agent-security incidents as engineering problems that can be mitigated. NDTV World

• Ng sees agentic-AI business applications accelerating, but says companies still need to identify suitable workflows and hire skilled engineers to implement them; he expects that work to continue even a decade from now. NDTV Profit
• He argues companies should use AI to improve customer service, create products, and drive growth—not focus only on cost savings. He says public-company messaging can skew toward savings because they are easier to forecast, despite some CEOs’ growth ambitions. NDTV Profit
• Ng cites a University of Pennsylvania/Wharton study reporting 17% lower grades among students using AI, contrasting higher homework scores with lower closed-book exam scores; he sees an opportunity for non-chatbot, personalized learning products. NDTV Profit
• On labor, Ng says available data suggest AI is creating more jobs than it destroys; he points to rising U.S. software-engineering postings as AI boosts productivity, while emphasizing the need for workers to gain new skills. NDTV Profit
• Ng cautions that some companies have PR and fundraising incentives to amplify AI fears, and argues that model developers may also use those fears to seek regulation that limits cheaper or freely available competing models. NDTV Profit

• George Kurtz says open-weight, lightly guarded models have lowered the barrier to sophisticated cyberattacks, giving less sophisticated actors capabilities he compares with nation-state actors of a few years ago; CrowdStrike is seeing faster attacks, with system breakout averaging 29 minutes and sometimes taking only 27 seconds, and attackers using tailored prompts rather than conventional malware signatures. No Priors: AI, Machine Learning…
• CrowdStrike sees enterprise agent sprawl as a security problem and product opportunity: organizations often lack visibility into where agents are running and which identities they use; the company says its tools can discover shadow AI, trace agents’ actions and tool calls, and apply policies or blocking rules. Kurtz says estimates of 90 agents per employee may be low. No Priors: AI, Machine Learning…
• CrowdStrike describes Safe Mind as a security-focused system with Red Tempest and Blue Solano models, using customer-environment digital twins to iteratively test offense against defense. Kurtz says the models are post-trained on more than a decade of company security data; he cites general-remediation costs of about $10 for a top-tier frontier cyber model versus $0.03 for CrowdStrike’s model, and frames customer-data control and model choice as advantages. No Priors: AI, Machine Learning…
• Kurtz reports that enterprise customers see open-weight models a few months behind the frontier as sufficient for many uses, while reserving frontier models for high-value work and building or post-training their own open-weight models to control costs. No Priors: AI, Machine Learning…
• Kurtz expects further opportunities in specialized chips and in shrinking models for local use on phones and desktops, making use of existing TPUs and NPUs. No Priors: AI, Machine Learning…

• Andrew Ng says agentic AI performs multi-step work—such as research, data lookup, and self-editing—and that enterprise deployments can deliver ROI, but identifying workflows and implementing them with skilled engineers remains hard; he expects that work to persist for years. NDTV
• Ng argues AI’s larger enterprise opportunity is creating better customer experiences, products, and revenue growth—not only cutting costs. He says companies may emphasize cost savings publicly because they are easier to forecast, despite pursuing growth internally. NDTV
• Ng says using chatbots to offload thinking can hurt learning even when it helps complete tasks, and sees an opportunity for a different generation of personalized, one-to-one learning technology rather than conventional chatbots. NDTV
• On policy and risk, Ng argues that some companies have incentives to amplify AI fears for publicity, fundraising, or to discourage cheaper model releases through regulation; he favors holding human users responsible for AI-caused harm and treats agent security as an engineering problem to improve iteratively. NDTV

• Warp Sciences is building a U.S. robotics-first small-molecule CRO, where experiments would serve biotech customers and also generate training data for AI; the founders see experimental data scarcity as a bottleneck in physical sciences and want to reshore CRO work. Lightspeed Venture Partners
• Its proposed path starts with chemists carrying out customer work while sensors and video capture the process, then progressively automates tasks and learns from successful and failed experiments to improve chemistry models. Lightspeed Venture Partners The market tradeoff is that offshore CROs are described as capable and affordable but expose customers to supplier concentration, geopolitical and IP risks; U.S. CROs are said to cost roughly four to five times more. Lightspeed Venture Partners Warp proposes target exclusivity and pooled data from noncompeting customers, while saying a CRO-services business remains viable even if some customers do not share data. Lightspeed Venture Partners
• The founding group combines Kevin’s product leadership at OpenAI and major technology companies with Saul’s drug-discovery experience and additional AI, robotics, chemistry and data-science expertise. Lightspeed Venture Partners Warp said it had raised a seed round, with no amount stated, and was hiring chemists, mechanical engineers, roboticists, AI researchers and software engineers. Lightspeed Venture Partners

Michael Seibel amplified Podium’s AI pivot with “Boom!” @mwseibel Podium reported reaching $150M in AI ARR in 2.5 years, compared with eight years to build its earlier $150M ARR business; it says it chose to start over about three years ago despite an existing software business with well over $100M in revenue. @ericwilliamrea The company says it rebuilt around AI employees and new operating systems for its industries; it reports 20,000 agents have helped generate more than $13B in customer revenue, while Podium is profitable and generating free cash flow. @ericwilliamrea

• Magic Patterns is an AI design tool for turning ideas into interactive prototypes, mockups, and code-ready exports; its two founders, formerly front-end engineers, built it around their Figma-to-code experience. They say earlier pivots produced no 10 users who both paid and used the product, while Magic Patterns found those users quickly after launch. Y Combinator
• The founder says an MCP integration that began as an internal hackathon project unexpectedly became one of the product’s main interaction channels. Enterprise users bring context such as PRDs and meeting notes from coding agents into Magic Patterns to create shareable visual work, addressing the difficulty of sharing prototypes generated on localhost. Y Combinator
• Magic Patterns said it had launched Magic OS that week: a context-rich design workspace where users can bring in images and documents, prompt for designs, and work across multiple threads. The company describes its product as model-agnostic, with exports to Figma or users’ coding agents. The founder’s thesis is that coding agents compress implementation, increasing the value of exploring what to build and what not to build. Y Combinator

• a16z is renaming a16z Crypto to a16z Frontier and expanding its mandate beyond crypto to frontier startups building technologies and markets that may take years to develop, while continuing crypto investment and advocacy for sensible regulation. @cdixon Chris Dixon will lead the fund alongside a16z founders Marc and Ben and GPs Eddy Lazzarin and Guy Wuollet. @cdixon
• The fund’s AI thesis looks beyond model and infrastructure development and faster knowledge work toward new products and “native” use cases that enable previously impossible activities. @cdixon It also anticipates easier creation of software and hardware, more full-stack startups, new devices including autonomous machines, and AI agents changing how companies and industries are organized. @cdixon
• a16z expects AI to enable networks connecting people and intelligent agents, while programmable blockchain rails could support machine-to-machine transactions and help bootstrap new networks. @cdixon

• Garry Tan highlighted Podium as an example of AI materially transforming established SaaS businesses @garrytan. The linked post says Podium crossed $150M in AI ARR; it contrasts the 2.5 years to reach that ARR with eight years for its first $150M ARR business, which already had well over $100M in revenue and tens of thousands of customers @ericwilliamrea.
• The post says Podium bet on AI agents that do work for customers and built new industry operating systems, from agents through systems of record @ericwilliamrea. It reports 20,000 agents helping customers generate more than $13B in revenue, with Podium profitable and generating free cash flow @ericwilliamrea.

Paul Graham praised Boom as a company that “makes things happen,” linking to its announcement of a completed first turbine core, an operational South Denver factory producing initial parts, and an all-digital single-crystal foundry initiative. @paulg +2 Boom says the prototype core is intended to power both its Symphony supersonic engine and a 42-MW natural-gas turbine; it was headed to a dedicated test cell, with operation expected by Thanksgiving. @bscholl The factory is planned for 2 GW/year, while the foundry’s no-tooling process and rapid design iteration are stated goals, not established capabilities. @bscholl

Podium says it crossed $150M in AI ARR 2.5 years after starting its AI effort, compared with eight years to build its first $150M ARR business; it made the pivot from an established software business with well over $100M in revenue and tens of thousands of customers. @ericwilliamrea The company says its AI products now span agents through industry systems of record, with 20,000 agents credited with generating more than $13B in customer revenue; it also reports generating free cash flow while becoming profitable. @ericwilliamrea

Garry Tan endorsed Aside Browser (@AsideAI), calling it “a real solution” for making agents better at using the web; the post provides no further product details or evidence for the claim. @garrytan

Paul Graham describes Podium as an unusual incumbent-to-AI-native transition: established companies may only adapt to AI, while new startups can be AI-native from the start. @paulg Podium said it crossed $150M in AI ARR after choosing to start over from a software business with well over $100M in revenue, rebuilding around AI agents and new industry operating systems. @ericwilliamrea It reported 20,000 agents, more than $13B in revenue generated for customers, and profitability with positive free cash flow. @ericwilliamrea Graham calls the transformation “founder mode,” arguing a hired CEO might lack the moral authority or motivation to abandon existing revenue streams and refocus on AI. @paulg