OpenAI Codex and ChatGPT product lead Thibault Sottiaux
Thibault Sottiaux announced that a ChatGPT subscription can now also be used for Devin. His brief announcement describes a new connection between the subscriptions. It does not specify eligible plans, usage allowances or activation steps, which still need to be checked in the product documentation.
Meta senior director of AI Madhu Guru
Madhu Guru argues that open-weight models accelerate the decline in the price of intelligence but did not start it. He identifies three drivers over recent years: distilling the strongest models into smaller ones with cheaper inference, infrastructure efficiency, and competition among providers. He sees a recurring pattern across model companies: a new generation of mid-sized models reaches the intelligence of the previous generation of large models. Users can consequently obtain similar capability for less money. In his account, open weights reinforce an existing trend rather than explain the entire decline in prices.
Anthropic Claude Code builder Thariq
Thariq revisited a side project he built before joining Anthropic. He spent about two weeks on it with Opus 4; it used the Agent SDK, required a continuously running process, and did not work particularly well. He says a single prompt to Opus 5.5 later ported it to Claude Managed Agents and made it much more reliable. This describes his experience with his own project, without a quantitative comparison. In another post he showed it on his browser’s new-tab page and said he had merged several PRs from other contributors. The updates focus on the reliability of managed execution and continued development of the project.
OpenAI Codex product team member Nan Yu
Nan Yu shared his enthusiasm for prompt completion. He describes a progression in programming interaction: writing code by hand, using Tab to complete code, and then prompting agents. Now the prompts themselves can be completed too. His post concerns his own experience with the feature and does not present measured productivity gains.
AI tutorial and interview creator Peter Yang
Peter Yang calls for more AI startups to work on managing and curing diseases rather than adding more personal assistants for email. This is a proposal about startup priorities, not an announcement of a medical product or treatment results. His other practical tip does concern assistants: when using Grok Bot as a chief of staff, do not register its email under your own name. Saying “Let me copy in Peter to find a time for us” becomes awkward when Peter is also your name. He suggests a distinct name for the bot so recipients can distinguish the person from the assistant. The posts address both the problems AI should tackle and how to reduce identity confusion in everyday collaboration.
Vercel CEO Guillermo Rauch
Guillermo Rauch says agents already buy infrastructure products and services frequently through the Vercel CLI marketplace, and the team is extending this capability to domains. He presents it as another step toward agents taking an idea all the way to an online business. In another post he shares Vercel’s own traffic and deployment figures: bots generated 58.18% of network traffic over the last 30 days, compared with 32% in January 2024. He says more than 60% of deployments are now agentic, up from about 3% in January 2026. Agents account for up to 83% of pageviews on Vercel’s own documentation sites, a share that rises as the team optimizes content for them. These numbers refer to network traffic, deployments and documentation pageviews respectively, not a single metric. Rauch predicts that direct human traffic will become a small fraction of the internet in coming years, with more of the web built by and for agents.
Box CEO Aaron Levie
Aaron Levie predicts that agents spawning other agents, continuous background work and agent swarms will consume 1,000 times more tokens than people prompting agents one at a time. He argues that chat-based AI limited to the pace of human prompts will look outdated within a year or two. In his view, the vast majority of tokens will eventually support continuous work in the background and in business workflows. This is a forecast of usage patterns and demand, not a measured industry total. On that basis, he sees adoption as still early and expects substantially more compute and infrastructure to be needed.
FirstMark investor and MAD Podcast host Matt Turck
Matt Turck highlights Synthesia’s ability to generate a professional-level video for a company or product in its own style from a single prompt. He describes it as the vision Synthesia has pursued from its early days and says it became available to everyone that day. The post does not provide plan restrictions or independent quality tests; “professional-level” is his assessment of the output.
AI engineering community organizer Swyx
Swyx warns that tickets for AI Engineer in New York are expected to sell out over the weekend, with the event starting Monday. He describes it as the city’s largest technical conference and the team’s first with a finance main stage. He connects that focus to his own path through investment banking, a hedge fund, big tech and a New York startup. The post does not specify a venue or ticket price, so attendance details remain in the linked event information.
FPV Ventures partner Nikunj Kothari
Nikunj Kothari advises founders to ask their marketing teams to exclude VCs from paid-promotion outreach on X. He says he receives three DMs a day pitching a “collab” or “insane opportunity” that is actually a paid partnership. He believes this shows poor judgment and can make a company appear to be buying attention ahead of a large raise. In a follow-up, he adds that the same pitch should not also be chased by email. His feedback as an investor asks founders to consider whether the outreach fits the recipient rather than simply pursue more exposure.
Builder Zara Zhang
Zara Zhang observed that Astra seems to use the Claude logo often when designing websites. She shared the behavior in a lighthearted post. The readable text provides neither a test count nor specific triggering conditions, so it is treated as a personal observation.
No Priors — Beam: The Great American Open Model with ReflectionAI Co-Founder and CEO Misha Laskin
**Key takeaway: Enterprise adoption of open models depends on inference infrastructure, business evaluations and reliable delivery as well as model capability. Customizing the system can come before fine-tuning the model.**
ReflectionAI co-founder and CEO Misha Laskin previously researched reinforcement learning at Google DeepMind and holds a PhD in physics. He says the company grew from about 30 people to about 300 and released its first open model, Beam. The original plan was to apply reinforcement learning to existing open models. The team later concluded that pre-training and reinforcement learning were tightly coupled and moved to training complete models itself.
The training investment reflects that approach. According to Misha, Beam has 500 billion total parameters and 23 billion active parameters, with pre-training using 6,000 GB300s. Reinforcement learning used more than 10,000 GB300s for four weeks. He claims three to four times the reasoning efficiency of models in the same capability class, emphasizing faster task completion and lower customer costs. These are figures and performance assessments reported by the team.
The commercial value of open models depends on delivering the whole system. He compares buying tokens to renting a stack that includes the model, inference software, cluster and hardware. Enterprises seeking greater control need those supporting components too. ReflectionAI therefore provides deployment tools and services to large enterprises and the public sector, using valuable tasks to drive inference demand rather than simply selling model access.
Enterprise customization may follow a different path from AI-native companies. Misha expects most enterprises to first customize agent systems around open models, connecting existing tools, evaluations and workflows, and consider fine-tuning only as they mature. Financial compliance, cyber defense and legal workloads require identifying valuable tasks, building evaluations and generating suitable data. Switching models does not automatically produce business results.
On safety, he acknowledges legitimate concerns but advocates broader scrutiny of open models to address vulnerabilities a small number of labs cannot exhaustively examine. He invokes the saying, “With enough eyeballs, all bugs become shallow.” This explains his support for open research, rather than proving that openness is inherently safe. He also expects AI to accelerate scientific experiments and research iteration while retaining a role for human intuition. In his view, engineers who can deploy models into real work and understand evaluations and harnesses still have substantial work ahead.