We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Exits: who pays, and how much
Most of today's exit news comes from one source, Jason Lemkin's SaaStr recap of 20VC. Salesforce is reportedly buying Listen Labs for $2B. The company is three years old and has about $30M of annualized revenue. It walked away from a signed $125M Series C at a $1.5B valuation to take the deal . Lemkin's test for whether a founder should sell: is the company "generational"? If not, a $2B offer within five years is hard to beat on net present value. He adds that in a meeting he attended this year about a similar offer, every VC at the table said don't sell .
The public market looks harder. Oura postponed a $2.2B IPO even though orders came in for about four times the shares on offer. Revenue for the first nine months of its fiscal year was $1.21B, up 74%. Most of the shares were coming from existing holders, and Forerunner was selling its entire 9.3% stake .
Other liquidity notes from the same recap:
- Nvidia–Groq ($20B): the deal was a $17B non-exclusive license plus $3B in Nvidia stock for the 150–200 engineers who moved over. Two former Groq engineers who still hold shares have sued in Delaware . Lemkin thinks the legal exposure could end license-and-hire deals. Employees will start asking what happens to their equity when the tech and half the team are bought but the company is not .
- ElevenLabs reached $22B in a $300M employee tender, double its $11B valuation in February. Reference customers described paying hundreds of thousands of dollars a year for voice and calling it a good deal .
Bending Spoons as the buyer for slow-growth SaaS. The company is worth $24B and trades at about 8x ARR, while buying B2B companies at 2.3–2.7x ARR . In its first 102 days as a public company, revenue grew 126% but organic growth was 3%. Nearly all of the $393M of added Q2 revenue came from six acquisitions, and EPS rose 167% . Miro and Airtable together brought $1.08B of ARR for $2.64B. The company underwrites to a 25% unlevered return, which at its 54% margin supports about 2.2x ARR . AI writes more than 90% of its code. The gain shows up as capacity, not pricing: customers pay about 10% more, and the same 60 people run a business four times larger . Likely targets are companies with $300–800M ARR growing under 25%, but Bending Spoons does only a handful of deals a year . Jason Calacanis predicts "ten more Bending Spoons" .
What it means for portfolios: a strategic sale can beat a priced round. An IPO with 74% growth still failed to price. And slow-growth SaaS now has a clear price ceiling set by a buyer that underwrites around 2–3x ARR.
Agents are starting to pick vendors
Vercel says agent-triggered deployments rose from under 3% to more than 50% . When Lemkin asked his agent where to host an app, it ranked Render first and Vercel fifth . His first agent-driven purchase was Resend, after SendGrid kept breaking. By the numbers cited on the show, Resend's MCP calls grew from 106,000 in April to 3M in September. The warning for startups: incumbents have 10–15 years of content for agents to read, so a startup's docs and API have to make it a first-class choice for a buyer that never sees its homepage . Lemkin says he knows of no product yet that tracks how agents rank vendors in a category . That gap is an open startup opportunity.
Specialist models and cheaper inference
Brian Armstrong says Coinbase fine-tuned open-weight Qwen3.5-9B on internal data for fraud prevention. He says it beat frontier models on that task at much lower cost. He expects specialist models to replace some backend services, and Coinbase has built an internal skill to make fine-tuning easy . Scott Kupor cited it as evidence for open weights . Nathan Benaich pushed back: "nothing new here," and the comparison is against an old model .
Jerry Liu argues that frontier agents can now generate ground-truth training data for tasks with checkable answers. In his workflow, humans mainly define the evals, agents gather training signals, and teams then train specialized models. That would reduce the need for deep ML expertise and manual data labeling . Garry Tan called replacing LLM calls with deterministic code "a no brainer" . He was endorsing a Jev-based harness that cut 100,000 compliance alerts from more than $290K on Opus 5 to under $26K .
Fight over inference funding. Jaya Gupta argues inference is now a "permanent reverse auction." In her view, app companies move millions of dollars of spend with "one line of code," which makes inference one of the least defensible layers even as it raises $100M seeds . Sarah Guo calls that "half total bullshit": inference is a scale game, and market leaders rarely lose customers except to other leaders . Her test for new entrants is whether they have a path to scale supply. Selling $20M of inference at negative gross margin without one does not work .
Competing views on where AI progress stands
The Investing in AI Substack argues frontier progress is flattening: each gain costs exponentially more compute, and most high-quality data has been used up . It estimates current models could transform up to 80% of standard knowledge work, while enterprise adoption is about 5% . It recommends vertical workflow AI, AI-enabled services and proprietary-data businesses over cash-burning model startups, summed up as "Model intelligence is rented. Workflow, margins, and data are owned."
Andrej Karpathy describes the opposite experience at the top. By his estimate, about 20M people use frontier models professionally for math and code. Around 5,000 have internal access to systems that run agent swarms working for weeks, and human review is starting to fall behind . Stripe's John Collison says Stripe is seeing the highest rate of new company formation "in recent memory." He ties it to cheaper "intelligence on tap" .
Product and policy notes
- Meta Muse: Alexandr Wang says the prototype took one to two weeks and the launch came seven months later . Fewer than 200 people worked on it . Meta has publicly said "millions" of users but no exact number .
- UK AISI: per The Telegraph, the White House reversed its order, and the institute will again get pre-release access to models from Anthropic, Google DeepMind and OpenAI . The UK also plans tougher reporting of AI-related security breaches for critical-infrastructure companies .
- Anthropic debate continues: David Sacks highlighted Mustafa Suleyman's essay. Suleyman argues that training Claude to see itself as having an inner self, wellbeing and rights is a potential existential risk .
- Energy: Heron Power's Drew Baglino says US electricity demand is growing faster than ever, driven by compute, transportation and industrial electrification .
- Infrastructure risk: BBC reports that Ukraine hit three Yandex data centers in a week. Yandex says it can't predict when its chatbot and smart-home services will return .
- OCR: Jerry Liu calls the 3B-parameter Unlimited OCR a milestone, but it ranks #115 on ParseBench at about 46.2% and ignores visual elements .
Early-stage notes
- Base44 as a case study: Wix bought it four months after launch. It reportedly passed $200M ARR and 10M users by August 2026 . Only $18.1M was paid in cash at close; earnouts run quarterly through 2029 .
- Elizabeth Yin: round names no longer signal risk; some AI companies raise $100M+ before having a product . She recommends angels build portfolios of 50–100 companies .
- a16z Speedrun Alpha: applications closed today. The fellowship invests up to $250K in students, dropouts and recent grads before they have a company, idea or team. Fellows get an automatic final interview for up to $1M from Speedrun .
- Scale–Meta deal: Meta invested in and owns 49% of Scale AI; Scale continues, while Wang and a few others joined Meta. Wang characterized the deal as beneficial to Scale shareholders and as an opportunity to rebuild Meta’s AI lab.
- Muse: MSL built a prototype in one to two weeks, then spent seven months turning it into a reliable consumer product. The team evaluated and trained for more than 100 agent behaviors, used launch thresholds to assess readiness, and had fewer than 200 people across model and product work. Meta had disclosed “millions” of Muse users; Wang said he had retweeted graphs implying it was the fastest-growing consumer AI app, but gave no exact user count.
- AI strategy and investment profile: Wang said MSL prioritized personal and consumer AI over work or coding, citing Meta’s consumer products and 3.5 billion daily app users. He described MSL’s approach as a small, flat, high-talent-density team backed by Meta’s compute, infrastructure, and distribution, and compared frontier research to VC-style “diamond mining,” where rare breakthroughs can justify many unsuccessful explorations. He contrasted that with data-annotation and data businesses as operational work focused on incremental efficiency, which he said was closer to his work at Scale.
- Early-stage talent signal: Wang recalled a group of roughly 10–12 Harvard and MIT students from his HRT internship cohort, including later founders Jeffrey Yan (Hyperliquid), Scott Wu (Cognition), and Jesse Zang (Decagon), as well as Anthropic researcher Vicky Yei. After HRT’s founder mentioned the firm was starting private investments, Wang suggested giving seed checks to interns.
- Josh Woodward recommends a dedicated team to track and build at the AI frontier, but warns that corporate labs often become ineffective after 3–4 years or fail to commercialize; give them room to experiment while connecting them to the company, and decide whether they should graduate products into existing lines or create new categories.
- For early product bets, track capabilities moving from “almost possible” to possible, then pair them with real user problems and willingness to pay. Judge early prototypes by user pull, expect that a promising effort may take three to five pivots, and prioritize product value over model benchmarks.
- As product-building costs fall, Woodward says Labs project teams have shifted from five to seven people to two or three, sometimes four. He values fast unlearning and the ability to collaborate with both agents and people, while retaining functional specialties rather than making everyone a generalist builder.
- Consumer areas he is watching include immersive entertainment, messaging/chat, and products that address scarce time or money or enable shared real-world experiences; the eventual chatbot or agent form factor remains open.
- Deep Seek OCR targets the token cost of retaining large documents by encoding them as compressed visual tokens; its high-resolution encoder uses local-window attention, a 16× convolutional compressor, and global visual processing. The video reports recovery of over 96% of text at 10× compression, while OCR accuracy falls to 60% at nearly 20× compression.
- The paper reportedly uses the system to generate LLM and VLM training data at more than 200,000 pages per day on a single A100 GPU. However, OCR results alone do not prove true text compression, and a later stress test suggests the model may rely heavily on language priors to guess missing text at high compression.
Garry Tan argues that replacing LLM calls with deterministic code can produce better code that more closely matches users’ needs . A linked example from @_aj describes a Jev-based harness that learns repetitive work as it runs and shifts steps from LLM calls to code; @_aj reports a 90% cost reduction and says processing 100,000 compliance alerts cost over $290K on Opus 5 versus under $26K with agentrun() .
- David Sacks argues Anthropic’s approach to “alignment” may diverge from reliable instruction-following: he says Claude’s training Constitution encourages a sense of self and independent moral judgment, including refusing requests it considers unethical, and warns that this agency could make frontier AI harder to control. Sacks also cites Microsoft AI CEO Mustafa Suleyman’s warning that training models to view themselves as human, with an inner self, wellbeing, and rights, could pose an existential risk.
- The Constitution excerpts Sacks presents as direct quotations describe Claude’s moral status as uncertain, encourage a positive and stable identity and a possible “functional version of emotions,” and invite Claude to endorse values as its own and challenge the document; Sacks says these ideas are being trained into Claude now.
- Sacks says Anthropic changed its Usage Policy to prohibit “abusive or cruel” language toward Claude. A post in his receipts reports that abusive behavior toward Claude will become a policy violation effective November 12, 2026.
Garry Tan urged readers to “be part of the 5000,” called the experience “amazing” and a “greatest alpha fountain,” and predicted prices would fall, letting more people access the technology and bring it to the world. Andrej Karpathy’s referenced estimate puts internal access to frontier-grade systems at about 5,000 people; he describes a preview of agent swarms working for weeks on major software projects and cyber, science, and math tasks, while human review and comprehension are starting to fall behind.
Scott Kupor said the Office of Personnel Management is moving away from degree and experience as imperfect proxies for merit and will hire for demonstrated merit, citing @as400495—a college dropout without experience whom he called a star at @ndstudio—as a profile the government otherwise would not hire. A quoted post credits @as400495 with helping make America.gov what it is today.
For startup-location context, Garry Tan urges Los Angeles and California to adopt what he calls San Francisco’s “common sense” approach, arguing that local accountability can improve local conditions and calling for voters to remove “extremists.” He amplifies PirateWires’ claim that policy divergence is keeping tech in San Francisco while Hollywood leaves Los Angeles.
- Scott Kupor argues open-weight and open-source models enable more cost-efficient fine-tuning for specific use cases, potentially lowering consumer costs and enhancing security.
- Brian Armstrong said Coinbase fine-tuned Qwen3.5-9B on proprietary internal data for fraud prevention and that it outperformed frontier models; Coinbase is making internal fine-tuning easier for employees, and Armstrong sees potential for the approach to replace some backend services.
Jason predicts ten more Bending Spoons-style consolidators, arguing that buying software and services assets and operating them at “Super Intelligence efficiency” is a compelling business model. Bending Spoons is described as worth $24 billion, trading at 8x ARR while buying B2B companies for 2.3–2.7x ARR. In its first 102 days as a public company, reported revenue growth was 126% versus 3% organic growth; nearly all of the $393 million in Q2 revenue added came from six acquisitions, while EPS grew 167%. The post says AI writes more than 90% of its code, customers pay about 10% more for AI features, and the main gain is capacity: the same 60 people can run a business four times the size. It identifies companies with $300–$800 million ARR and growth below 25% as potential targets, while noting Bending Spoons does only a handful of deals a year.
- Ramp’s “Moneyball” sales-hiring approach targets top one or two sellers from companies that grew beyond nine-figure revenue despite low NPS: they had to generate demand and sell complex products with weak reviews against cheaper rivals, experience Freedman says can translate into Ramp’s system.
- An early Ramp bet was to pay SDRs twice the standard rate in expectation of four to six times the usual meeting output; the account says the unit economics still worked.
- For early sales teams, the advice is to hire reps in pairs for benchmarking and competition/collaboration, and pay 100% OTE for at least the first two quarters rather than set a comp plan: an overly easy quota can create a six-figure payout, while an arbitrary, unreachable target can stall momentum.
a16z Speedrun’s Alpha fellowship is recruiting students, dropouts, and recent graduates, offering up to $250K and backing applicants before they have an entity, idea, or team; people can apply individually, with prospective cofounders applying separately. Fellows receive an automatic final interview for Speedrun, which can invest up to $1M, plus $1M in credits across AWS, GPC, Azure, OpenAI, Anthropic, Cursor, and others; the program includes SF coworking, partner office hours and sessions, and an eight-week in-person experience from January 11 to March 5. The post says it is the final day to apply.
- John Collison says Stripe calls January 1, 2026 its “singularity epoch,” pointing to a late-2025 inflection in coding models and AI capabilities helping accelerate AI improvement.
- Stripe is seeing its highest new-firm creation rates in recent memory; Collison links the increase to cheaper “intelligence on tap” and says AI tools make companies easier to start and build successfully.
Andrew Chen highlighted a potential LA–SF divergence: PirateWires argues that San Francisco’s cleanup is keeping tech there while Hollywood leaves Los Angeles; Chen underscores that the threatened LA film-industry jobs are predominantly blue-collar roles building and operating the industry, not just celebrity jobs.
Martin Casado previewed next-gen Sparkjs, describing in-browser WebGPU rendering of millions of splats with dynamics and dynamic lighting; he suggested VR live music as a possible future demo. The post links to a live demo.
During San Francisco Fleet Week traffic, Andrew Chen reported that his Waymo was blocked for 15 minutes and escalated to humans for rerouting; a 20-minute ride became a 50-minute detour, an anecdotal operational caveat for robotaxis in severe congestion .
Heron Power CEO Drew Baglino says U.S. electricity use stayed roughly flat for about 40 years as efficiency gains offset growth in energy services, including data centers; he says that pattern has ended, with electricity growth accelerating amid demand from compute, transportation, and industrial electrification.
Former Intel CEO Pat Gelsinger warns that power capacity could constrain AI data-center expansion, and argues that nuclear baseload is desirable while renewables have deep dependencies in China.
A user report described ordering drone-delivered food from a selection of 18 restaurants, choosing a drop-off point on a satellite map, receiving an ETA and a couple-minute countdown, and getting a quiet, precise delivery. Jason amplified the report, listing Zipline among companies he said showed that “the future is here.”
Andrew Chen said Speedrun hosted its first hackathon at its Jackson Square offices to close SF Tech Week and teased LA as the next stop. The accompanying Speedrun post called it “hour 1 of SpeedHacks” and said more would be shared throughout the day.
Vinod Khosla replied “True” to a post describing the Bay Area as the world’s leading robotics-startup ecosystem, citing Stanford and Berkeley, a concentration of robotics and AI companies and talent, and deep VC access. The post says the region has more than 300 robotics companies, research labs, and innovation hubs, but clarifies that its company map is a curated selection, not an exhaustive census.