Vercel CEO Guillermo Rauch
Guillermo Rauch says Reflection’s open-weight Beam model is free through November 12 at the entry point he links. He describes the service as having zero data retention without detailing its terms. He also mentions a startup benchmark comparing fx with the pure-assembly implementation fx.s, without giving measurements in the post. He argues that even as tools improve, building good products still takes sustained human attention and effort.
Google Labs
Google Labs introduces Tabletop, an experiment where groups put files and links in a shared space and ask for interactive maps, voting boards, or seating charts. Founding engineer Roman Nurik shows how his family uses it for chores and a scavenger hunt with live scoring. He also shows LEGO maps and a fantasy-themed New York trip. The waitlist is open to people in the United States aged 18 and over.
Anthropic AI assistant Claude
Claude’s official account describes a new work in artist Mariko Mori’s retrospective All That Shines. Mori used Claude to write code for a computer-generated animation, shown as a video installation. The post identifies the part of production where the tool was used and the form of the finished work.
SPC general partner Aditya Agarwal
Aditya Agarwal sees two kinds of AI startup opportunities. One brings models closer to existing work through forward-deployed teams, industry consolidation, or horizontal products. The other uses AI to advance research and then builds products from that progress. He cites Externa’s chemical building blocks for proteins, Epana Bio’s autoimmune research, VenomicsAI’s venom-based drug discovery, and Equivariant’s multiphysics modeling for semiconductors. The company descriptions and investment outlook come from his post, which gives no clinical or commercial results.
FPV Ventures partner Nikunj Kothari
Nikunj Kothari analyzed subprocessor lists from 803 companies. He reports that 44% listed at least two model providers, up from 33% in his earlier analysis. The share listing an inference provider rose from 22% to 33%. He found 86 pairs in which each company listed the other as a subprocessor. He notes that the lists can lag. The sample describes disclosed vendor relationships, without measuring usage or revenue.
Former Claude Code contributor Thariq
Thariq announces an AIE Code closing keynote titled The Sweet Side of the Bitter Lesson. He gives the time as 5 p.m. the following day without specifying a time zone. He plans to discuss what coding agents might be like to use next year and how developers can help shape that experience.
AI & I by Every — A New Vision for Human Work After Automation
Key takeaway: Dan Shipper uses an Every customer survey to show how the same model can generate many answers while people’s responsibilities and experience help a team decide which findings to act on.
Every co-founder and CEO Dan Shipper opens his Thesis: 2027 closing keynote with a week of his own work. Between meetings, he analyzed a quarterly survey, made a quiz game for the team, spotted a chance to sell subscriptions to teams, and built a CRM to contact prospects. He compares the work with what several roles would once have done. This is his account of how he worked, without a measurement of actual staffing costs.
Every gave several colleagues the same survey. Shipper saw an opportunity to sell team subscriptions. Growth engineer Yash Poojary noticed that excess email might be contributing to first-month subscriber churn. Marketer Douglas Brundage noticed that some readers lacked people around them who wanted to discuss AI. Editor Jack Cheng turned the results into a claymation quiz show. All four used AI and prioritized different findings according to the work they owned.
Every’s Mike Taylor also tried getting Claude to predict what Shipper would say in Slack. According to the talk, when the model could generate 400 answers, a similar answer was present for 93% of questions. When asked to rank its own answers, it put a similar answer in the top 10% around 60% of the time. This single internal experiment gives teams a practical question: once plenty of ideas are available, who decides what to follow up on? As Shipper puts it, “Generation is really cheap, but ranking is hard.”
Shipper expects model progress to raise the baseline for everyone and make default answers easy to obtain. He suggests evaluating models on real work and checking whether they help the person responsible make a better decision. Comparisons with a future model should also include people who have learned new methods and will be using it themselves.