• Nick Kakonis frames arbitrage as a time-scale gap and sees restaurant pricing as an example: a table’s value differs between a Tuesday at 9 p.m. and a Saturday at 7 p.m., while menus have historically changed only quarterly. He argues localized, dynamic pricing for time-slotted businesses could generate hundreds of billions of dollars, and that AI may make the automation practical for operators. My First Million
• Kakonis is described as having built and sold a software company for hundreds of millions of dollars. During the sale process, he expected the company would not repeat its extraordinary COVID-era growth; management chose to handle diligence without an investment bank and completed a list of 896 questions in three days, offering a 2% fee to address board objections. My First Million
• Kakonis sees in-person experiences as potentially more important as AI agents reduce person-to-person interactions, and is pursuing Strange House, an immersive magic experience intended to foster human connection. He says the project has a potential 90,000-square-foot site beneath Wall Street after six months of lease-LOI negotiations, but may not happen; it cannot be piloted at small scale and needs a multimillion-dollar build of its third room to demonstrate the concept. My First Million

• The panel described OpenAI’s proof release as including the Unique Games Conjecture and a quasi-Riemann-hypothesis result; the latter established a zero-free strip for real parts between 7/8 and 1, not the full Riemann hypothesis, and panelists said reaching 1/2 requires a new idea. Lex Fridman They characterized the quasi-RH proof as chaining known techniques at striking scale, and said the release reportedly included only positive results from thousands of prompts, with about a 10% success rate—so the capability boundary remains unclear. Lex Fridman
• Panelists said Anthropic produced a 13-million-line, sorry-free Lean proof of Fermat’s Last Theorem; Lean checks formal proofs down to their axioms. Lex Fridman They said this proof is not clean or Mathlib-quality and that refactoring and integrating such work into the library remains difficult and labor-intensive. Lex Fridman
• The discussed AI-assisted Navier–Stokes result constructs a finite-time blow-up under smooth forcing in three dimensions, but panelists said it does not make practical fluid engineering easier: simulation and inverse design remain costly and largely unchanged. Lex Fridman They identified a still-open opportunity in finding useful representations or coordinate systems for turbulent flows using data, compute, and mathematics, with potential applications including aircraft drag reduction, artificial hearts, and turbines. Lex Fridman
• Ankur described a fine-tuning experiment in which a model trained on another LLM’s apparently random-number outputs began expressing that model’s hidden preference for owls, despite the preference not being explicit in the training examples. Lex Fridman He linked this to subtle correlations in model representations and said such hidden behavioral transfer matters for alignment and deployment beyond low-stakes, verifiable tasks. Lex Fridman

• Shopify product-design lead Marvin Schwaibold says his workflow has shifted from Figma-first work toward Codex, Claude Code and terminal-based tools for producing usable demos faster; designers still decompose tasks, curate AI output and own quality. He says he is seeing teams of two do work previously requiring many, while clarifying that Shopify’s admin playground prototypes are not production code and cannot enter the main repository. Y Combinator
• Shopify is prototyping Sidekick suggestions that start a conversation with a merchant and can directly change the store—for example, proposing a product-description rewrite and letting the merchant publish it. The intended shift is from navigating the admin interface toward conversational delegation; the demo is still a prototype. Y Combinator
• Schwaibold described Mintlify’s self-updating documentation for startups, enterprises and agents as growing quickly and used by many YC companies and other developer-tool companies. Y Combinator
• Instinct’s site describes the company as building a lunar lander for today’s space industry; in the review, the technology footage was described as showing real activity and the team as building the system. Y Combinator
• Website reviews surfaced a practical early-stage product lesson: Nstack’s broad “AI workspace” positioning left reviewers unsure what to do, while Instaplay asked users to sign in after prompting them to create a game, before they could judge the result. Reviewers recommended a clearer primary use and letting users experience the generated game before signup. Y Combinator

• OpenAI interviewees said computer-use agents were already faster than average human computer users and approaching expert-user speed; they also said lower-priced models were improving at computer use. They framed control over autonomy and auditability as an enterprise requirement, citing app/site sandboxing and human approval for selected actions as available controls. a16z
• OpenAI’s James Sun argued that startups still have room in verticals where AI diffusion is slow: teams with deep problem and industry knowledge, plus expertise in regulation, go-to-market, and customer service, can tackle complex workflows across disparate software. As an example, he cited a company serving restaurant groups that uses storage-room photos and vision to forecast next-day needs, order groceries through procurement software, and update staffing schedules in an ERP. a16z
• The interviewees expect legacy and newly built software to coexist for years, with computer-use agents potentially increasing pressure on software vendors to support agent interaction. They also flagged an open-web monetization tension: in many cases, agent ad impressions do not count toward real revenue; they cited WebMCP as a way for people and agents to use websites together, noting OpenAI added support in its in-app browser. a16z

• Nadella no longer expects alignment to emerge automatically from scaling; he distinguishes model-level safety from an external zero-trust containment layer with agent identities, tamper-resistant activity/tool/model-call logs, policies, and behavioral monitoring. 20VC with Harry Stebbings
• He expects a market for frontier models to persist but says most token usage will be commodity: open-weight or cheaper closed models can serve stable tasks, especially when trained on a customer’s own traces. He sees price competition and favors horizontal specialization with multiple winners and composable layers, while acknowledging that vertical integration could emerge. 20VC with Harry Stebbings
• Nadella argues firms should retain more value from their knowledge and AI-use signals, and expects each firm to have a private evaluation benchmark as core IP, with models improving against it; he predicts this practice could spread broadly within about five years. 20VC with Harry Stebbings An enterprise agentic system for trade claims began with a frontier model and moved to a lower-cost MAI model after training on company traces; after initial FTE support, the customer’s IT team was managing it. He describes this kind of deployment as a combination of business and technical consulting. 20VC with Harry Stebbings
• Nadella says software-development demand is elastic and reflects large IT backlogs, while extending similar gains to other knowledge work is harder because verifiable rewards are less clear; he points to contact centers, supply chains, marketing, science, and engineering as areas of potential. 20VC with Harry Stebbings He also cites power and silicon shortages and long lead times, says the timing of supply and demand is uncertain, and argues that large training investments ultimately depend on revenue and broader GDP growth. 20VC with Harry Stebbings

• Replit CEO Amjad Masad argues that agents can handle organizational coordination in the background, giving people more autonomy; he warns that companies anchored in old ways of working risk falling behind as the shift accelerates. @a16z
• Replit reports that engineers nearly tripled code output in the first half of 2026 while review times, reversions, and product incidents stayed flat; it says designers and support staff also began using agents. @amasad
• Replit is building Replit for Business around a controlled agent environment with access limits, human-approval rules, safeguards against mistakes, and shared company knowledge. @amasad In Replit’s account, Sumitomo Mitsui Finance and Leasing said a development process that had required six people for six months was completed by one person in one week; the project’s expected lifecycle cost fell from approximately JPY 3 billion to under JPY 1 billion. @amasad

• Marc Andreessen backs Replit’s Agent Habitat concept for enterprise agents, which he says currently run in chat windows, personal accounts and one-off scripts with unclear permissions and visibility. He argues companies need to rein in that chaos without extinguishing bottom-up initiative. @pmarca
• He treats chaotic employee adoption as a signal of real value and argues that expanding who can contribute code can strengthen organizational alignment; he says Replit and its customers recognize this opportunity, while most companies do not yet. @pmarca
• Andreessen says he believes Replit has an example at Sumitomo Mitsui Finance and Leasing where one employee completed in one week a development process planned for six people over six months. @pmarca

Yann LeCun trained a network of shared, sliding image detectors at Bell Labs in 1989 to recognize handwritten ZIP codes; backpropagation tuned the detectors from examples rather than having them drawn by hand. Machine Theory His networks later read checks at American banks, and he shared the 2018 Turing Award. Machine Theory

• Anthropic has filed an initial S-1; Clouded Judgement estimates roughly four weeks from filing to trading under a typical schedule (2–3 weeks to a roadshow, then about 1.5 weeks to trading). Clouded Judgement The post reports more than $19B in LTM revenue for the period ended June 30, 2026, and 1,368% year-over-year revenue growth in Q2. Clouded Judgement
• Reported margins were 85% gross for the six months ended June 30, 2026, −34% LTM GAAP operating margin versus 8% in Q2 2026, and 11% LTM FCF margin versus 21% over the latest six months. Clouded Judgement Anthropic cautions that margins may fluctuate or compress because compute expense can precede the revenue it supports; most compute is committed 12–18 months before availability, and it expects greater near-term use of more expensive short-term compute. Clouded Judgement
• Token-metered pay-for-intelligence is the primary revenue model, supplemented by subscriptions; enterprise minimum-spend contracts are increasingly annual or multi-year and primarily consumption-based. Clouded Judgement Anthropic describes an integrated model, platform, and first-party-app stack, with four model classes—Mythos (including Fable), Opus, Sonnet, and Haiku—spanning different performance, cost, and speed needs. Clouded Judgement Mythos is offered through trusted-access programs for vetted cybersecurity and life-sciences users, while Fable is generally available with production safeguards; Claude Code and Cowork are first-party examples of agentic coding and knowledge-work products. Clouded Judgement

YC said one week remained to apply for Winter ’27, with a Nov. 2 deadline @ycombinator. Its advertised package included $500K in funding, $12M+ in tokens, credits and deals, founder-partner support, a founder community, and Demo Day before thousands of investors @ycombinator.

Graham said YC often continues advising founders 10 years after investing, rather than handing them off to later investors; he attributed this to the difference made by being there at the beginning and compared YC to old friends @paulg +1.

a16z amplifies Marc Andreessen’s view that technologies such as PCs, spreadsheets, and email often entered workplaces chaotically before organizations developed guardrails that preserved bottom-up use while restoring control. @a16z Replit’s Agent Habitat addresses a similar problem for agents, which may run in chats, personal accounts, or one-off scripts with unclear permissions and visibility; the proposed balance is to let employees create useful code while providing an organizational environment for it. @a16z +1

Andreessen argues that code created by more employees can embody business logic and strengthen organizational alignment. @pmarca As a qualified productivity example, he says he believes Replit has a case at Sumitomo Mitsui Finance and Leasing where one employee completed in a week a development process planned for six people over six months. @pmarca

Garry Tan urged followers to “Publish your stack,” amplifying Theo’s announcement that he was publishing his skills; the posts provide no further detail about the skills or stack @garrytan +1.

OpenAI announced GPT-6.1 Terra as a fast, capable computer-use model at a lower price; it can understand screens, act across apps, and write naturally, with efficiency aimed at handling more of a workday. @OpenAI Sam Altman said OpenAI is especially excited about this new era of computer use. @sama

YC said its Winter ’27 application deadline was Nov 2, with one week left to apply when posted @ycombinator. Its advertised founder package includes $500K in funding, $12M+ in tokens, credits and deals, founder-led partners, access to its founder community, and a Demo Day before thousands of investors @ycombinator.

Casado added a caveat to a quoted analysis claiming that 88% of VC-backed founders who exited their companies received $0; the quoted post says the analysis covered 157,669 startups and 248,454 founders from 1987 to 2025. @Carles_Reina He suggested that relatively few VC-backed companies shut down even when they are not working, and guessed anecdotally that this may account for more than half of the cases included. @martin_casado

Shopify product-design leader Marvin Schwaibold says AI has changed his team’s design process: he uses tools such as Codex to build working prototypes, and sees designers becoming more like editors and full-stack builders, with taste and judgment increasingly important as AI makes building easier. Shopify is rethinking key products for the AI era. @ycombinator

Paul Graham argues that UK universities demand more equity in spin-outs than US universities because they are less experienced and do not realize their terms are above market, rather than because they are tougher negotiators. This is his assessment of university spin-out financing terms, not independently substantiated evidence in the post. @paulg

@gdb announced three scientific discoveries with AI assistance—one each in harmonic analysis, complexity theory, and quantum field theory—and said AI is beginning to improve our understanding of the universe; the post links to an OpenAI announcement. @gdb Martin Casado amplified the announcement, calling it the start of a “great age of discovery.” @martin_casado

OpenAI’s computer-use team says its agents can operate almost any app through clicking and typing, including software without an API, and is close to beating expert humans on speed; Ari Weinstein works on consumer products and James Sun on enterprise products, and both were startup founders before joining OpenAI @a16z. Weinstein and Sun argue that cheaper software creation could enable well-crafted apps for niche groups or even one person, while human preference for direct manipulation means UI will persist and lower costs will raise expectations for software quality @a16z.

OpenAI’s computer-use agents can click and type through apps without APIs; the discussion frames this as a “universal API” for legacy ERP and finance systems, while noting that enterprise adoption requires controls for agents operating without constant human oversight. @a16z Ari Weinstein says the agents are already faster than the average person at computer use and are nearing superhuman speed; a16z’s post puts their back-office cost at $6–8 per hour, versus about $10 for offshore outsourced talent and $30–45 for US labor. @a16z