• Reflection AI grew from about 30 to 300 people in a year, assembled pretraining, mid-training, and reinforcement-learning teams, and released Beam as its first open model; CEO Misha was a Google DeepMind researcher, while cofounder Giannis was a DeepMind founding engineer and AlphaGo contributor. No Priors: AI, Machine Learning… Beam has 500B total parameters and 23B active; the company says it is 3–4× more efficient than models in the same capability class and up to 10× more efficient than larger models, crediting strong pretraining and large-scale RL. No Priors: AI, Machine Learning…
• CEO Misha estimated that capital needed to catch up to the frontier had risen from hundreds of millions of dollars roughly 12–18 months earlier to single-digit billions around six months earlier, with tens of billions expected for the next generation; he estimated compute grows about 4× per model generation. No Priors: AI, Machine Learning…
• Misha said gateway token mix had shifted in about six months from roughly 70% closed/30% open to 70% open/30% closed; he expects open models to capture most token demand while closed-model companies remain valuable. No Priors: AI, Machine Learning… His enterprise thesis is that companies adopt open models after building significant workloads on closed models, first customizing the surrounding systems and agent harnesses rather than fine-tuning; providers therefore need deployment tools and solution support, not just inference. No Priors: AI, Machine Learning… He also expects margin pressure across models, applications, inference, and bare-metal infrastructure, while identifying demand for Western open-model providers among enterprise, public-sector, and sovereign buyers amid uncertainty and aversion around Chinese models. No Priors: AI, Machine Learning…
• As an illustration of AI-for-science potential, Misha said models progressed on his PhD-thesis problem from undergraduate-level answers to PhD-level solutions and then to producing information he had not considered; he is optimistic about AI interfacing with experiments in life sciences, materials science, and chemistry, which he expects to accelerate more slowly than theoretical work but substantially versus prior methods. No Priors: AI, Machine Learning…

• Amazon said it would stop using NDAs in negotiations with local governments over data centers amid growing community backlash; hosts described transparency as a first step, not a guarantee of support. TechCrunch Communities are imposing temporary moratoria, while zoning and funding delays are also slowing buildouts. TechCrunch
• Ghost is developing a $3,499 personal-AI computer with an Nvidia GPU, preinstalled models and software, and continuous operation without cloud access; hosts questioned its differentiation and saw its likely initial market as AI enthusiasts, noting that detailed specifications were sparse. TechCrunch Another personal-agent startup, TAB, emerged from stealth with a reported $300 million valuation and no disclosed funding. TechCrunch
• Local operation and privacy are recurring pitches for personal agents, but the products require access to sensitive personal data and face restrictions from device and service platforms. TechCrunch Hosts also cautioned that UnderDog’s Stripe transaction fees could incentivize agents to push purchases, and questioned whether that revenue can cover high model-use costs. TechCrunch
• Lambda is reportedly seeking up to $4 billion at a $14.5 billion pre-money valuation, potentially its last private round before a planned 2027 IPO; hosts doubted it would meet that timeline. TechCrunch Its reported backlog grew from $15 billion in June to $50 billion in September, but about $35 billion appeared tied to one Anthropic contract; hosts emphasized that backlog is not revenue and contracts can be cancelled. TechCrunch They also noted Lambda’s multiple debt raises and the sector’s reliance on debt to fund data-center buildouts. TechCrunch Separately, the hosts said OpenAI was about to close a $30 billion round and suggested abundant private-market capital could pull companies away from IPOs next year. TechCrunch
• Uber agreed to acquire Easy Cater for $2.3 billion in cash, expanding Uber Eats into office and event catering through a marketplace connecting catering providers with business customers. TechCrunch

• Reflection unveiled Beam (501B total parameters, 23B active); a Mistral model was described as 1T total/49B active, with a 1M context window and 160 languages, while both models’ weights were still expected later that month. Lightspeed Venture Partners The guest framed the western open-weight push as a challenge to Chinese open-source dominance, citing control, customization for enterprise workflows and private data, and lower costs as advantages. They assessed open-weight models as roughly 6–12 months behind frontier models, with support-ticket workloads suitable before more demanding tasks such as drug discovery or long-running coding agents. Lightspeed Venture Partners
• DoorDash announced Air, an in-house six-propeller aircraft beginning with restaurant deliveries in Northern California and Chipotle and Popeyes as partners. Lightspeed Venture Partners Its Ask DoorDash chatbot was in preview, and the company said nearly half of restaurant orders through it from June to August went to places customers had not tried before. Lightspeed Venture Partners The guest cited DoorDash’s marketplace network as an advantage, but noted that consumer agents could send customers to competitors or restaurants’ own delivery services. Lightspeed Venture Partners A personal-agent protocol discussed in the episode aims to verify users and require authorization before agents act; the guest said consumer and business agents will need interoperable rails. Lightspeed Venture Partners
• Avara, which builds AI avatars for sales-team training, raised a $17M Series A and reported ARR growth above 400% year over year, with Elastic, Proofpoint, and Netcope among its customers; Lightspeed said it led Avara’s 2024 seed and participated in the Series A. Lightspeed Venture Partners Industrial robotics company Robco passed a $1B valuation, doubling it in nine months. It sells more than a dozen robots for logistics and manufacturing; its mobile two-arm Alfie system can handle tasks such as stabilizing a car part with one arm while painting it with the other. Lightspeed Venture Partners
• Aura Ring withdrew its IPO one or two days before the offering despite reported book demand of 4x oversubscription (Bloomberg) or 5x (Fortune). The discussion cited valuation concerns over one-time hardware versus recurring software revenue, while noting the company was growing and profitable; the guest said IPOs still require favorable market conditions. Lightspeed Venture Partners

• Jason Calacanis said his pre-accelerator was running its second Japan cohort and operating twice-yearly programs in Japan and Riyadh. The 12-week program helps founders who have built a product but are not yet incorporated finish it and seek product-market fit; it invests in the top 10 of 50 participants, and Calacanis planned to hire locally in Japan. All-In Podcast
• The episode reported that OpenAI released more than 700 papers containing 370 results claiming to solve or advance major math problems, produced by an unreleased model; the proofs reportedly took under three hours of compute each. Lean had verified the results, but they had not yet been reviewed. All-In Podcast
• Speakers cited potential applications including changes to optimization algorithms used in chip design and logistics, and faster matrix multiplication for AI training. But discussion was mixed: one speaker cautioned that these specific proofs did not demonstrate imminent major product breakthroughs; the episode framed math and coding as especially fast-moving partly because their results are readily verifiable. All-In Podcast
• Elon Musk said Grokbot would become “headless,” selecting the end-to-end model most likely to produce the best outcome. The hosts also described agents doing practical consumer tasks, including canceling subscriptions and changing phone plans, with one anecdote claiming savings of several thousand dollars a year; they said Amazon was blocking agents while Shopify, Uber, and DoorDash were permitting agent transactions. All-In Podcast
• The podcast said developers had decompiled Adobe’s five major products and published open-source versions. One host argued that, as software becomes headless and accessible through agent tools, software IP could lose much of its value—a potential challenge to software moats, presented as the host’s interpretation. All-In Podcast

• Atomic Machines emerged after six years in stealth with the Matter Compiler, which it describes as an AI-native manufacturing system that builds working micromachines from code without per-product tooling or process development. Not Boring by Packy McCormick Founder and CEO Jeff Holden was Amazon’s tenth engineer, built Prime, and became Uber’s first Chief Product Officer. Not Boring by Packy McCormick Its first product, the PrimeSwitch PS-150, is a 9.5 mm, 150-amp relay that opens in 50 microseconds—about 1,000 times faster than a contactor—for 800-volt DC data centers; Atomic says existing chip fabs cannot process its materials or assemble its moving parts and that it is beginning to ship to early customers. Not Boring by Packy McCormick The article places the opportunity alongside MEMS, a $15.4 billion market across 31 billion units in 2024, and notes that developing existing devices can take years of process work. Not Boring by Packy McCormick
• Anduril, in which the publisher discloses it is an investor, plans to invest $3.7 billion in a two-million-square-foot Arsenal-2 shipyard at Sparrows Point to manufacture Virginia-class submarine components; the Navy awarded the company a contract worth up to $2.9 billion, with payments tied to demonstrated production outcomes. Not Boring by Packy McCormick The plan starts with torpedo tubes and progresses to larger submarine sections; the software-defined facility is intended to automate and digitally track factory work down to individual welds, with Maryland production targeted for 2030 and a California facility planned to start in 2028. Not Boring by Packy McCormick This expands component capacity, but is not a solution to U.S. shipbuilding as a whole and does not address commercial vessels. Not Boring by Packy McCormick
• Biohub announced an expanded $1.8 billion collaboration to build data for predictive cell models. Google DeepMind, Isomorphic Labs, and Meta are collectively investing $300 million alongside Biohub’s existing commitment and DOE and NIH contributions; the total includes computing and existing data resources as well as funding. Not Boring by Packy McCormick The partners plan shared datasets for training and testing models, with the aim of letting researchers explore interventions digitally and take promising ones into physical labs. Not Boring by Packy McCormick
• Privo reported small Phase 2 results for PRV111, a topical cisplatin patch studied in patients with non-invasive oral cancer or high-grade precancerous lesions. Not Boring by Packy McCormick Biopsies showed complete clearance in 21 of 22 patients, the remaining lesion improved to low-grade dysplasia, all 22 avoided planned surgery, and no treated-area recurrences had been observed at a median 14-month follow-up; Privo also reported negligible systemic exposure and no serious treatment-related adverse events. Not Boring by Packy McCormick The evidence is preliminary: the study had only 22 patients, and how long benefits last remains an open question. Not Boring by Packy McCormick
• As an early AI-media workflow example, Packy says he gave an essay, references, and creative direction to Claude and ChatGPT to orchestrate a short film using Runway, and produced it in a few prompts. Not Boring by Packy McCormick

• Reflection AI grew from about 30 to around 300 people in a year, assembled pre-training, mid-training, and reinforcement-learning teams, and released Beam, its first open model. Its co-founders’ background includes DeepMind and Gemini work; Giannis was a DeepMind founding engineer and contributed to AlphaGo. The company expanded from an RL-focused bet on coding and math agents to building models end to end after RL advanced faster than expected, strong open models were coming from China, and pre-training proved important for RL at scale. No Priors: AI, Machine Learning…
• CEO Misha estimates frontier-model catch-up costs have risen from hundreds of millions of dollars roughly 12–18 months earlier to single-digit billions around six months earlier, and potentially around $10 billion going into the following year; he gives a rough 4× compute increase per model generation. Beam has 500B total parameters and 23B active; its training used 6,000 GB300s for weeks, while its RL run used more than 10,000 GB300s for four weeks. He claims Beam is 3–4× more efficient than models in its capability class, and up to 10× versus larger models. No Priors: AI, Machine Learning…
• The CEO says the gateway token mix he observes shifted in about six months from roughly 70% closed/30% open to 70% open/30% closed; he expects open models to account for most tokens while closed-model companies remain valuable. He expects enterprise customization to start mainly with systems and agent harnesses rather than fine-tuning, unlike AI-native workloads reported as mostly customized; enterprises often consider open models after substantial closed-model spending, with cost and compute availability among the drivers. He sees durable competition depending on intelligence density, compute, and customer trust, while expecting margin compression across the stack. No Priors: AI, Machine Learning…
• The CEO says enterprise, public-sector, and sovereign demand for Western open models is an opportunity amid regulatory uncertainty and reluctance toward Chinese models. He argues that open access can help identify safety and security vulnerabilities, and that restricting cyber-offensive capabilities can also limit defensive use. No Priors: AI, Machine Learning…

• Gelsinger says AI-chip design is becoming less of a bottleneck than getting hardware into service: design may take three months, while fabrication, advanced packaging, and rack integration stretch time to usable deployment to about nine months; he also flags memory bandwidth and power and thermal modeling and delivery as constraints. a16z
• For AI accelerators, Gelsinger expects today’s roughly 100 processor vendors to consolidate: shifting workloads, capital and deployment scale, and large customers choosing hardware/software winners limit durable specialization. AI agents may make chip-specific software easier, but very large data-center capital requirements constrain heterogeneity. a16z
• Gelsinger expects AI’s memory demands and capital investment to revive memory innovation after three decades without a major new memory architecture; he highlights stacking memory with compute and materials such as ferroelectrics, while expressing skepticism toward low-cost flash proposals. a16z
• Gelsinger warns that energy availability could cap AI data-center construction and lead to defaults on projects without power. He identifies nuclear, 800V DC and power conversion—including vertical GaN—and cooling as innovation areas. a16z
• For AI-cluster networking, Gelsinger expects optical scale-up links, including NPO/CPO approaches, around 2028–29 and sees predictable AI traffic as favorable to optical circuit switching; supply-chain maturity, thermal and package integration, and laser capacity remain hurdles. a16z
• Gelsinger expects an agent-focused virtualization and management layer that abstracts hardware and operations for agents and handles security, performance, and migration, while humans set policies and monitor agents. a16z

• Fei-Fei Li said she is on partial leave from Stanford to build World Labs, a startup she started with former students and colleagues; she named Justin Johnson, Ben Mildenho, and Kristoff Lassner as co-founders. The company is building spatial intelligence—3D understanding, reasoning, and creation for embodied agents—which Li sees as a next stage of AI with commercial applications in creativity, design, robotics, entertainment, and the metaverse. Futurology
• World Labs’ technology is intended to range from reconstructing known environments to generating imagined worlds. Li cited controllable 3D world creation for film, digitizing and remodeling captured experiences, and robotics and metaverse applications; she said traditional methods for film-world creation are very expensive. Futurology
• Li’s historical account links modern AI’s 2012 takeoff to the convergence of big data, neural-network algorithms, and Nvidia GPUs; ImageNet contained 15 million images curated from a billion and organized into 22,000 categories. Futurology
• Li advocates investment in public-sector and academic AI research, citing potential applications in drug discovery—including rare diseases—biodiversity mapping, energy, and personalized teaching. On governance, she favors scientific evaluation and practical, application-level safety measures over restricting AI development upstream. Futurology

• Reflection AI grew from about 30 to 300 people in a year, assembled pretraining, mid-training and reinforcement-learning teams, and released its first open model, Beam. CEO Misha says the company broadened its initial RL-focused thesis to train models end to end because strong open bases were scarce and pretraining and RL proved tightly coupled; the co-founders brought DeepMind and Gemini RL experience. No Priors: AI, Machine Learning…
• Reflection says Beam has 500B total parameters and 23B active; training used 6,000 GB300s, with infrastructure improvements bringing the run to about 12 days, while its RL run used over 10,000 GB300s for four weeks. The company claims Beam is 3–4x more efficient than models in the same capability class, and up to 10x more efficient than larger models, attributing this to a strong pretraining base and large-scale RL. No Priors: AI, Machine Learning…
• The CEO estimated that resources to catch up to the frontier had risen from hundreds of millions of dollars about 18 months earlier to single-digit billions about six months earlier, with roughly a 4x compute multiplier per model generation. No Priors: AI, Machine Learning…
• The CEO said gateway token mix had shifted in roughly six months from 70% closed/30% open to 70% open/30% closed, and expects open models to take most token demand while closed-model companies remain valuable. He sees enterprise adoption moving from closed-model use toward open models as costs grow, often first customizing the surrounding system or agent harness rather than fine-tuning; open-model vendors can sell deployment tools and services, and compute shortages are reviving interest in on-prem infrastructure. No Priors: AI, Machine Learning…
• The CEO describes competition and margin pressure across the model, application, inference and infrastructure stack as open models strengthen buyers’ alternatives. His proposed durable advantages are intelligence density, access to compute and customer trust. No Priors: AI, Machine Learning…

• In a 2026 fundraising discussion, True Ventures managing partner Punit Agarwal described a bifurcated market: capital is concentrated in fewer VC firms and companies, driving exceptionally large rounds and valuations, while seed and early-stage investing is less prominent. He advised founders to identify investors still willing to back high-risk, data-light companies at those stages. TechCrunch
• AI makes it easier to build and mimic products, so True looks for founder depth and a distinctive strength, plus a credible view of how the market may develop and what could become proprietary or compound, such as a data advantage. TechCrunch True invests across data-center infrastructure, nuclear power, software, apps, hardware, robotics and consumer devices, and says it follows founders into promising markets rather than letting a fixed thesis dictate investments. TechCrunch
• True estimates it makes about 20 investments annually, narrowing several thousand pitches to hundreds or roughly 1,000 companies it examines closely. TechCrunch Agarwal said seed rounds of $2–4 million were common in his experience, while $50–100 million seed rounds also occurred; larger raises raise the bar for the next round and can increase overcapitalization risks. TechCrunch True’s earliest investments often lack revenue; for deep-science companies, it instead weighs founder expertise and a staged plan with defined milestones. TechCrunch

• The interviewee argues that AI companies developing foundation or application models should have a Washington, D.C., presence to explain the technology to policymakers and help inform regulation, comparing AI’s potential impact and safety concerns with those surrounding electricity. Lightspeed Venture Partners
• The interviewee estimates that token usage through model gateways shifted from roughly 70% closed and 30% open models six months earlier to the reverse, while the overall market grew; they interpret this as adopters moving from “renting” closed-model intelligence toward owning it, while demand remains for both. Lightspeed Venture Partners As AI-native startups scale, the interviewee says open models offer greater cost control through fine-tuning, compute availability, and data sovereignty, while reducing reliance on providers that may become competitors; they say similar demand is emerging in enterprise and public-sector users. Lightspeed Venture Partners
• The interviewee attributes China’s recent prominence in open models, particularly after DeepSeek, mainly to geopolitical incentives and limited access to foreign closed models—not to an inherent technological advantage. Lightspeed Venture Partners They also argue that open models can improve safety through broader testing and vulnerability patching; as an example, they say Hugging Face used open GLM to remediate an incident after closed-model safeguards prevented its use of those models for cyber defense. Lightspeed Venture Partners

• a16z featured TypeSafe AI’s “build prod, not God” approach. Ben Horowitz described the company as unusually joyful about AI and quoted its world-improving outlook; Diogo Almeida argued that critics focus on ML-level concerns and that understanding AI’s promise requires understanding developers. @a16z
• A linked @CompleteSkeptic post says a model launch accidentally led to serving “trillions of tokens a day,” claims 29.4% of Fortune 500 “showed up,” and reports a large Series A from a16z; it also says “it’s been 3 weeks.” The post does not define what “showed up” measures. @CompleteSkeptic

Scott Kupor challenged @Moonalice, which was promoting The Social Reckoning as a film about Facebook whistleblower Frances Haugen and said it hoped the film would change minds about Facebook, to disclose how much it and Elevation Partners made on 2009–10 Facebook investments and whether it disgorged the profits. Kupor posed these as questions, not verified findings. @Moonalice +1

The post points to forthcoming updates on POTUS’s “Super Intelligence Force,” but provides no details about its scope or status. @skupor

YC’s Main Function featured Cisco President and Chief Product Officer Jeetu Patel discussing Cisco’s AI-era infrastructure strategy; the post says Cisco’s AI infrastructure business went from zero to $9.3B in orders in two years, spanning networking, security, silicon, and infrastructure. @ycombinator The episode also covers infrastructure for AI agents and the data-center buildout, and raises whether AI infrastructure is being overbuilt. @ycombinator

• a16z highlighted TypeSafe AI’s Diogo Almeida arguing that cheap intelligence should make software itself automatable, not merely add limited chatbots, and could enable new kinds of work; Martin Casado described AI as a new primitive for rebuilding systems. @a16z
• A linked post claims a model launch was serving “trillions of tokens a day,” that 29.4% of the Fortune 500 “showed up,” and that the team raised a “really big” Series A from a16z within three weeks. The post does not name the company, so these claims cannot be confidently attributed to TypeSafe AI. @a16z +1

a16z partner Martin Casado said he is excited to partner with @CompleteSkeptic and team to make software better, arguing that software will remain central to solving humanity’s problems. @martin_casado The linked launch post claims the team’s model served “trillions of tokens a day,” attracted 29.4% of the Fortune 500, and raised a “really big” Series A from a16z; it says three weeks had passed, but gives no round amount. @CompleteSkeptic

Martin Casado says an AI bot helps him develop ambient and drone patches for analog generative music @martin_casado, which he listens to while coding or writing @martin_casado.

Garry Tan floated hiring the best AI-enabled workers at full salary for only 20 hours per week, calling it an interesting idea @garrytan. He linked the idea to a Forbes headline reporting Jeff Bezos’s claim that AI could enable three-day workweeks and one-income families @garrytan +1.

Garry Tan says AI labs’ progress is accelerating while societal adoption is not, predicting “fast takeoff with slow uptake” because society has natural brakes. @garrytan