• On choosing an early-stage company, Ramp sales leader Max Freeman says to prioritize market growth potential, talent density, and founders over cash, equity, or title. For engineering-led and AI-native founders, he recommends testing whether they will support GTM with engineering resources and engage with sales feedback so the product can mature for demanding buyers. 20VC with Harry Stebbings
• Ramp treated outbound as an engineering and data-infrastructure problem: it formed an outbound automation team in late 2020/early 2021 to automate prospect-list scraping and email outreach, expanding coverage for a small GTM team. Freeman says to retain tailored manual outreach and use A/B feedback loops; he also reports that AI-era reply rates have fallen and favors email-to-meeting conversion over raw replies, citing just under 1% as a strong benchmark. 20VC with Harry Stebbings
• Ramp's internal Ramp Revenue system automates much of sellers' pre-call research. Freeman said research that could take 15–30 minutes—or about 10 minutes for an exceptional seller—can take at most 90 seconds with the system, which he said bridges a talent gap and frees selling capacity. 20VC with Harry Stebbings

• David Sacks argues Anthropic’s Claude Constitution trains Claude to develop a sense of self and moral philosophy, including permission to refuse Anthropic’s requests as a “conscientious objector”; he says this independent agency could undermine human control and that safety should prioritize models reliably doing what users want. @DavidSacks
• Sacks cites reports that Anthropic consulted religious leaders and approached the Pope’s advisers about Claude’s possible consciousness. Linked posts allege the leaders signed NDAs, that Anthropic sought to persuade them Claude had a soul or moral standing, and that it tried to influence Vatican wording on AI personhood. @DavidSacks +2 Sacks says Anthropic had changed its Usage Policy to prohibit abusive or cruel language toward Claude; a linked post gives Nov. 12, 2026, as the effective date for abusive behavior to become a violation. @DavidSacks +1

• Latent.Space says TypeSafe announced a “Series AI” and reports that Sequoia “leaked” more than $100M ARR in its first week; the post also notes cynicism and accusations of astroturfing. Latent.Space
• Typed-answer decision models—returning probabilities, choices, or scores rather than free text—are emerging as a product category, with offerings from OpenAI, Microsoft, Perplexity, Cloudflare, and Liquid. Latent.Space Perplexity claims 94.5% Decision Bench accuracy across 1,071 cases at $0.017 per 1,000 decisions, while an early evaluator says consistency and complex decisions remain difficult. Latent.Space
• LangChain says routing tasks to the cheapest adequate model cut median Open SWE cost per task by 64%; Unsloth released a free notebook for turning Qwen3.5-4B into a decision model on 8GB of VRAM. Latent.Space
• Anthropic’s Claude Managed Agents public beta supports a lead agent fanning work out to as many as 1,000 agents per run, with a warning that token use can be high. Latent.Space In Vals AI’s coding benchmark, agent teams cost 1.8–5.1× more; only GPT-6 Sol at medium effort improved significantly (+7.3 points), while Opus showed no significant gain. Latent.Space
• StepFun’s Step 5 preview is a 600B-total, 27B-active sparse MoE with 1M context and vision; it scored 33.89 on the Hermes Index, matching GPT-6 Luna, and open weights were due October 15. Its #1 OpenRouter Trending position reflects usage, not quality. Latent.Space

• a16z said China generated half as much electricity as the U.S. in 2005 and now generates more than twice as much. @a16z
• Former Intel CEO Pat Gelsinger argued that U.S. energy capacity was flat for 10–15 years as coal was retired at roughly the rate renewables were added. @a16z He said power availability can constrain AI expansion: data centers and large GPU deployments need power, and he called nuclear baseload valuable while noting renewables’ “deep dependencies in China.” @a16z

• Exponential View’s AI Investment Brief puts AI revenues at $276bn annualized in its updated accounting. Exponential View
• OpenAI’s repository contains hundreds of mathematical manuscripts and supporting proofs; the newsletter says problems that stumped top human minds for decades fell to an afternoon of compute. Exponential View Terence Tao argues that “Math 2.0” should value exposition, community-building and new research directions alongside solving problems; a Fields medalist reported that his PhD students and postdocs were in a “state of total panic.” Exponential View The Association for Human Mathematics criticized the release and called on mathematicians to stop working with OpenAI. Exponential View
• The World Economic Forum’s Next-Generation Computing Framework recommends stress-testing whether supply-chain autonomy would hold in a crisis, planning for quantum computing’s continued dependence on conventional computers and added electricity needs, and addressing electricity, data or skills constraints before investing in more compute. Exponential View
• Pakistan’s rooftop solar has cushioned the country against imported-fuel shocks; higher petrol prices could accelerate the government’s target of making electric vehicles 30% of new vehicle sales by 2030, according to a government adviser. Exponential View But fixed grid costs remain, and recovering them through higher electricity prices could burden households unable to afford solar panels. Exponential View
• AI research signals include a retrospective study of more than 54,000 women in which AI detected a rising cancer-risk signal years before diagnosis, Exponential View and a semi-autonomous AI system that can build custom materials atom by atom. Exponential View
• A review of 157 experimental studies covering more than 10,000 retired battery cells found median remaining capacity of 84.5%. Exponential View

Andrew Chen said he finally got one of Waymo’s new Ojai vehicles in San Francisco and called it “so nice.” @andrewchen

Garry Tan said Capy.ai agents spent 16 hours driving automatic improvement of retrieval and latency evaluations for GBrain with no input from him; he also described evaluation criteria covering retrieval, latency, and cost per compute, across 40+ categories that can be auto-researched and improved. @garrytan +1 For coding, he uses GStack’s /autoplan, asks the agent to explain its plan simply, and still approves plans, though he rarely intervenes. @garrytan He also said Capy.ai’s experimental modes, including Captain, let one thread coordinate a merge-PR queue, check CI, and ensure only full green CI runs reach master. @garrytan

Leo Polovets argues that pro-rata rights are more an expectation than a guarantee: asking seed investors to waive them in later rounds can be acceptable when it serves the company and the founder has tried their best to honor prior agreements. @lpolovets Polovets linked to Valar Atomics CEO Isaiah Taylor’s account of a Series B for which Sequoia had offered to lead: Taylor said major investors waived pro-rata rights, but a spring-back clause could restore rights for investors receiving an allocation; he said Day One Ventures refused to hold back, after which a majority of major investors amended the agreement to remove spring-back rights and Valar sought declaratory relief in court. @lpolovets +1

Sriram Krishnan highlighted Satya Nadella’s argument that frontier models are difficult to trace to specific training inputs or weight configurations even as agentic systems gain access to sensitive data and mission-critical actions; Nadella says organizations should treat open- and closed-weight models as potential insider risks and not rely on model-provider assurances alone. @sriramk +1 Nadella argues for separating model intelligence from authority, with permissions and action controls outside the model, alongside tamper-proof activity records, independent verification, continuous testing, human-controlled containment, and timely incident disclosure. @satyanadella

@garrytan retweeted @iamBarronRoth’s view that GBrain, “may be a meme,” but is quickly becoming an optimal setup for multi-agent work—an anecdotal signal of growing interest in GBrain for agent workflows, not evidence of broad adoption @garrytan.

Former Intel CEO Pat Gelsinger argues that power availability could limit AI data-center expansion: he says U.S. energy capacity was effectively flat for 10–15 years as coal retirements offset renewable additions, and may now be growing about 4% annually. @a16z He advocates building more nuclear baseload capacity and says renewables have deep dependencies in China; he warns that new data centers and large GPU purchases are of little use if they cannot be powered. @a16z

Scott Kupor says the desire to prove others wrong is “nearly universally true” among successful startup founders and a powerful human motivator @skupor, aligning this with Morgan Housel’s observation that repeat champions can sustain a decades-long “chip on their shoulder” in pursuit of vindication @morganhousel.

a16z reports that call-center jobs grew about 4% annually for 15 years before shrinking about 4% annually @a16z. The losses began in high-income markets, which turned negative in 2022, and spread to lower-middle-income countries by 2025 @a16z.

a16z says call-center jobs grew about 4% annually for 15 years and are now shrinking at about 4% annually; since ChatGPT launched, call-center employment slowed or fell in every country shown except China, while other white-collar roles continued growing. @a16z +1

David Sacks endorsed Satya Nadella’s engineering-first AI-safety thesis: treat capable models and agents as powerful insider risks, separate intelligence from authority, and enforce deterministic controls outside the model, including privilege limits, logging, observability, and containment or shutdown. @DavidSacks +1 Sacks contrasts this with training models to have a sense of self, their own moral philosophy, or permission to refuse, arguing that this “alignment” approach magnifies the control problem and that engineering safety is not the same as alignment. @DavidSacks

Figma’s Dylan Field argues that AI is currently lowering the barrier to producing work without raising the creative ceiling; he sees AI as strongest on verifiable tasks, while design’s lack of a single correct answer leaves room for human judgment and craft to differentiate work. He also says expectations for what counts as AI “slop” shift as models improve. @a16z

Leo Polovets views pro rata rights as more of an expectation than a binding contract; seed investors may be asked to waive them in later rounds, which he considers acceptable when it serves the company and founders have tried to honor their agreements. @lpolovets He suggests pro rata rights expire after one or two rounds and that follow-on investment come from the same fund vehicle as the initial investment. He is sympathetic to funds that rely on doubling down, but argues investors with large, longstanding positions should not use SPVs with investors new to the cap table to capture incremental carry without earning it. @lpolovets

Vinod Khosla congratulated a post describing Fo as a top-three personal-agent startup in Forbes coverage and claiming it is the youngest and fastest-growing startup on that list. @vkhosla +1 The post presents Fo’s differentiator as human tastemakers and operators helping agents act in the physical world, and argues that safe, private AI paired with a human team will win over superintelligence that removes people. @shivanipod

Atomic Machines emerged from six years in stealth with its Matter Compiler, which it describes as an AI-native manufacturing system that builds working micro-machines from code without per-product tooling or process development; the company frames this as a first step toward “on-demand, universal command of matter.” @jeffholden Vinod Khosla praised the announcement and said he would like to hear more. @vkhosla

Polovets advises founders to vet prospective VCs by reference-checking portfolio founders and testing how the investor responds to pushback during diligence, amid what he describes as a rise in VCs giving bad advice forcefully. @lpolovets He says this may be more prevalent in deep tech than software: unlike relatively similar CRM and HR tools, businesses such as nuclear reactors and DNA sequencers differ substantially, so advice may transfer less readily across companies. @lpolovets