Nan Yu
Nan Yu argues that, apart from rare cases, group chat makes a poor multiplayer AI experience and tends to produce generic output. This is his product judgment; the post does not detail failures or an alternative design.
Thariq
After a week of use, Thariq says good work does not require unlimited tokens and that the $200 plans are excellent. He also predicts that intelligence will never be more expensive than it is now; the former is personal experience and the latter a forecast.
Dan Shipper
Every CEO Dan Shipper praises argon while attributing part of its appeal to Antigravity feeling stuck in 2025, without the accumulation of complex features. He also links Every’s vibe check of argon for further reading. These are his impressions and a reading recommendation, without a reproducible comparison.
Nikunj Kothari
Nikunj Kothari previews Slice for Thought, an in-person San Francisco series starting with OpenRouter head of insights Peter Walker at noon on November 4. Each session lasts 60 minutes: Kothari interviews the guest for 45, then the guest asks him about venture for 15. Wednesday pizza lunches admit 10–12 listeners who can ask questions, with off-the-record Chatham House Rules and registration linked in the post. Separately, he jokes that vaguely posting startup revenue precedes a fundraising kickoff by about 11 days, citing n=17 and Grok Bot use. That small-sample observation does not establish a general fundraising rule; the event date was prospective in the original post and has passed by this delayed publication.
Thibault Sottiaux
OpenAI’s Thibault Sottiaux says the team can finally discuss some of its Git work and credits the people involved, without detailing the technology in the post. Separately, he says a dot will choose models for the task, drawing on different models for different parts of the work. The text establishes that routing claim but does not specify supported models, selection rules or availability.
Guillermo Rauch
Vercel CEO Guillermo Rauch argues that stateless sandboxes are the wrong benchmark because agents need their workspace, including identity, skills, repositories and memories. He says the Vercel Sandbox team is meeting latency and throughput objectives while prioritizing reliability and security. He also reports clusters in 19 global regions, weekly feature releases and customer-demand growth in double-digit percentages per week. Those coverage and growth figures are his statements. His point is to benchmark the real workspace; the post provides neither numerical service targets nor independent tests.
Garry Tan
Y Combinator CEO Garry Tan points to direct manipulation of weights as an interesting use of smaller models, with a linked example. In another update, he says Capy Captains increased his issue-fixing throughput by about 5X overnight. He attributes the change to coordination across captains and subthreads. This is his own workflow report, without a task set, timing definition or comparison conditions.
Peter Yang
Peter Yang argues that a writing and editing model can be held back by its harness. He asked ChatGPT and Gemini to edit the same post, finding most ChatGPT suggestions poor and Gemini’s suggestions strong. But he could not get working browser use in Antigravity, so he pasted Gemini’s edits back into ChatGPT to apply them in his browser. He hopes Antigravity improves quickly. The example separates editorial quality from interface-operating capability and describes his particular experience.
Aditya Agarwal
Aditya Agarwal announces that WisprFlow co-founder and CEO Tanay Kothari will return to SPC on November 12. He describes Kothari as a four-time founder and IOI medalist and reports a recent $280M Series B at a $2B valuation. Places are limited, with an application link in another post; the funding figures remain attributed to the announcement.
Amjad Masad
Replit CEO Amjad Masad argues that AI agents should run in dedicated, organization-controlled environments with access limited to the resources each task requires. His emphasis is on organizational control and task-specific access. The post supplies no particular implementation or full security audit establishing a protection guarantee.
Latent Space — Building Legal Agents When You Can't See the Data - Jacob Lauritzen, CTO of Legora
Key takeaway: Even when real client traces are unavailable, evaluation, citation verification and access boundaries must be part of the legal-agent product.
Legora CTO Jacob Lauritzen describes a product that grew from questions over Swedish law and search into drafting, file editing, bulk review and workflow orchestration. Legal work often involves confidential client material; serving a law firm does not give a software supplier permission to inspect all its data. He says, “We can’t look at any traces, basically.” Product improvement therefore cannot assume engineers can access production documents.
Evaluation separates tasks and practice areas. Engineers and lawyers construct realistic project environments, such as a 10,000-file data room, then assess risk discovery, analysis and report drafting independently. A model release can improve one task and regress on another, so an overall ranking is insufficient. Online signals are narrowed to task categories and known failure types. Removing names alone may not protect confidential material; reproductions depend on clients willingly providing suitable anonymized examples.
Tabular Review puts each contract in a row and questions such as termination periods or intellectual property in columns. The system first runs ten rows so users can refine questions and inspect results before expanding to the full set. Cells support review marks, flags and comments, with citations back to original files. An agent citing table results can preserve the path down to the underlying evidence; clickable citations do not mean every user actually checks them.
Workflow orchestration uses composable building blocks. Intake brings Slack requests into Lists; lawyers and agents progress the work in Legora and return results to Slack. Legal engineers configure workflows as power users, while forward deployed engineers code complex integrations and feed missing capabilities back to the core product. They generally still do not touch client document content; exceptional access may require formal onboarding like a client employee.
Model selection also considers latency, cost, governance and regional availability. The team is considering post-training for bounded retrieval tasks rather than rebuilding the general model behind its main agent. Cross-section edits, consistent definitions and chunked retrieval in long contracts remain difficult. Lauritzen’s idea of an agent acting as an API is exploratory, with execution cost, latency and uncertainty unresolved. His views on law-firm economics and jobs likewise do not establish that junior lawyers have already been replaced.