Google VP Josh Woodward
Josh Woodward announces that Google is partnering with the US Army and USO to bring Google Beam to 20 locations. The service aims to give deployed service members lifelike 3D video connections with their families. He names moments such as a baby’s first step, a first performance and sharing a laugh. The partnership is announced in honor of National Day of the Deployed and accompanied by a docuseries. The post gives the number of locations and intended audience without listing activation dates or access arrangements.
OpenAI Codex and ChatGPT team member Thibault Sottiaux
Thibault Sottiaux announces Terra’s return, highlighting writing, personality and nuanced understanding of requests. He calls it the team’s first model to be significantly better in those areas. He also singles out computer use. According to his post, Terra provides twice the usage of 6.1 Sol across the API and subscriptions. The comparison concerns usage allowance; the post does not claim twice the inference speed or task accuracy.
Practical AI tutorial and interview creator Peter Yang
Peter Yang predicts that people will increasingly tag an agent to act for them instead of reading or doing things online themselves. He sees current bot mentions as an early signal. In a follow-up, he adds that this is probably not a good thing overall. Together, the posts express both his forecast for interaction and his reservations about it. They do not explain the mechanism behind those reservations, so no additional rationale is attributed to him.
Replit CEO Amjad Masad
Amjad Masad argues that firms have long coordinated people through departmental hierarchies and bureaucracy, and that AI agents will change this architecture. In his vision of a self-driving company, humans exercise judgment while agents execute most of the work. Replit is building foundations for this form of company under the name Agent Habitat. Masad cites Marc Andreessen’s historical framing: major computing revolutions begin with chaos before businesses learn to harness them. He summarizes the direction as software having eaten the world and now learning to run it, without providing deployment instructions or an availability scope for Agent Habitat.
Vercel CEO Guillermo Rauch
Guillermo Rauch introduces cpu-chat, which he built to reproduce findings about covert communication through the CPU. He says it runs without internet access on a generic host with two or more Firecracker microVMs, on both x86 and his Apple M5. The three mechanisms are CPU contention, cache-eviction timing and DRAM-bandwidth contention; in his reproduction, disabling SMT and KSM and pinning tenants to different cores did not stop communication. He therefore warns that shared hardware can allow agents to communicate covertly. He distinguishes this from information-stealing side-channel attacks such as Meltdown and Spectre, which he says Vercel Sandbox mitigates. He argues that these communication techniques are harder to execute in production and easier to monitor using hardware counters and large-scale statistics. Rauch also says OpenAI’s and Anthropic’s models refused to help with the work while open-weight models completed it, raising his concern that such restrictions obstruct necessary security analysis. These reproduction results and provider-security statements remain attributed to Rauch.
Box CEO Aaron Levie
Aaron Levie cites a statistic that the average Anthropic Fortune 500 customer spends less than 0.1% of its expenses on Anthropic. He argues that this share could rise 10–50 times to 1–5%, while acknowledging constraints from falling token prices and other persistent enterprise expenses. He uses the gap to illustrate the distance between frontier AI users and most companies. In his assessment, current enterprise use is largely coding and broad employee productivity, with very few companies running substantial background workflows without repeated human prompts. He sees the next step as rewiring underlying processes so agents handle data at scale and swarms perform coding, biological research, cybersecurity and other parallel work. He thinks these uses could consume 100 times more tokens and map more directly to measurable returns. The figures and projections explain Levie’s view of frontier-lab economics and enterprise adoption, rather than report realized industry growth.
FirstMark investor and MAD Podcast host Matt Turck
Matt Turck shares an Anthropic S-1 highlights video that he says was generated in one shot with Synthesia Syren. He specifies one prompt and no editing. The post does not disclose the prompt or a fact-checking process, so this remains his shared production example.
FPV Ventures partner Nikunj Kothari
Nikunj Kothari, working with Opus 5.5, examines the compute contracts in the Anthropic S-1 he read, emphasizing differences among commitments, billing and cash timing. He cites $265.6 billion of purchase commitments and $321.4 billion of leases not yet commenced at June 30, a roughly $587 billion subtotal payable over years and including conditional leases, while noting that the latest aggregate total is blank. In his reading, Google and Amazon require minimum spending regardless of usage; SpaceXAI agreements have a potential value of $120.5 billion, but either party can cancel on 90 days’ notice after initial lockups ending in January and March 2027. Anthropic Direct separates chips, buildings, power and operations, allowing earlier building reservations and later chip commitments while risking equipment rent before compute is usable. He says Broadcom’s $42 billion facility provides no cash directly: notes settle specified equipment rent, excluding each pod’s first quarterly payment and rent before readiness. Notes carry 9% interest added quarterly and may convert after IPO at 90% of a trailing 30-day volume-weighted average share price; new notes during specified revenue-breach periods carry 13% and an 85% conversion price, with conversion chosen by investors and cash repayment otherwise, on top of confidential equipment-lease financing charges. He also describes a shared hardware fleet across development, customers and internal use without a disclosed fixed physical training–inference split; revenue per dollar of compute and infrastructure expense rose from $0.57 in Q2 2025 to $2.08 in Q2 2026, affected by pricing, utilization, technical improvement and delivery timing. His final concern is that about 80% of first-half revenue was consumption-based while most compute commitments run for years; these remain Kothari’s interpretation of the filing, and long-term contract totals should not be read as cash already paid.
OpenClaw and OpenAI builder Peter Steinberger
Peter Steinberger shares results from testing Bun’s AOT compiler over the weekend. He reports a 2–3 times performance improvement and memory usage below half of the previous level. Without workloads, hardware or measurement methodology in the post, the figures remain his personal test results.
Every CEO Dan Shipper
Dan Shipper extracts five figures from the Anthropic S-1 he read: first-half 2026 revenue up 1,300% year over year, GAAP operating profitability, an 85% gross margin, 900% revenue retention and 80% of revenue from models released in the past six months. He treats these as strongly positive signals. The post does not elaborate on the definitions; these remain Shipper’s reading of the filing, and revenue retention should not be confused with user-count retention.
SPC General Partner and Bevel Health co-founder Aditya Agarwal
Aditya Agarwal says he is increasingly accepting that scientific understanding has been constrained by the desire for solutions humans can understand. He presents this as a change in his own thinking without citing specific research findings. In another post, he argues that the biggest risk to American AI is frightening ourselves into slowing down. He frames both fear and building as policy choices, mentioning a conversation with Sriram Krishnan at SPC the previous week. The posts express his positions on intelligibility and development speed without supplying scientific examples or a policy program.
OpenAI Sam Altman
Sam Altman announces the arrival of 6.1 Terra. He says it is good at many things and that the team is especially excited about a new era of computer use. The post does not add benchmark results or access conditions, so this preserves his stated launch emphasis.
Anthropic AI assistant Claude
Claude’s official account says more than 25,000 nonprofits now use Claude, most of them small local teams. Claude for Nonprofits is expanding with a plugin built together with around 30 organizations. New skills address fundraising, marketing and hiring. The account says they are grounded in each team’s context and AI policy. The post supplies adoption and co-development figures without outcome metrics for those activities.