X / Twitter
Thibault Sottiaux, OpenAI Codex & ChatGPT
Thibault Sottiaux addressed GPT-5.6 Sol usage concerns among Codex and ChatGPT Work users. He explicitly denied any “nerf,” saying OpenAI had optimized inference and passed the savings to all GPT-5.6 Sol subscriptions, with this alone expected to provide approximately 10% more usage. He explained that increasing the product’s context limit from GPT-5.5’s 272k to 372k had produced higher-than-expected billable consumption, so it had first been reverted to 272k, with a gradual return to 372k planned. The team also investigated reasoning-effort experiments, internally called juice values, and rolled back the changes. Multi-agent use at high and xhigh reasoning effort was slightly above expectations, and auto-review revealed a small efficiency issue; both would be fixed. He also confirmed that GPT-5.6 Sol would remain in already-paid ChatGPT subscriptions, including Go, Plus, Pro, Team, Edu, and other paid plans, at least until OpenAI released a better model.
https://x.com/thsottiaux/status/2076495156757577895
Peter Yang, AI Tutorial Creator
Peter Yang observed widespread GPT-5.6 Sol use in the community that day, roughly guessing that over 90% were using Sol and fewer than 10% Terra or Luna. This was his interpretation of community feedback, not official data, but reflects preferences rapidly concentrating on the strongest available model after release. He also discussed communication when community sentiment turns negative, arguing that companies should not communicate less or become more “corporate” during controversy. Instead, they should sound more human, transparently explain what happened, and seek solutions with the community. He singled out Anthropic, saying its models are excellent too but its distant communication puzzles him. OpenAI’s more direct engagement this time is something other AI companies should learn from, in his view.
https://x.com/petergyang/status/2076512796481880270
Amjad Masad, Replit CEO
Amjad Masad shared a “Vibe Research” experiment fine-tuning Qwen-8b on Replit to play chess. He ran 3 experimental branches in parallel and reported real progress. What surprised him was how much models’ ability to do ML work had improved from previously poor performance. He believes someone with good intuition who can guide experiments may now accomplish interesting ML work with models despite having no previous machine-learning experience. He also demonstrated Replit’s computer use model playing his new chess engine, combining models, tools, and a concrete game environment rather than discussing theory alone. His focus is how AI coding environments make previously specialist experimentation accessible to more builders.
https://x.com/amasad/status/2076227936202662357
Guillermo Rauch, Vercel CEO
Guillermo Rauch’s central advice was to make models a cog in your own machine rather than outsource your brain. He connected Vercel’s AI infrastructure layers: AI SDK as the open model API, another open Agent API, and AI Gateway as open ZDR inference. Startups and enterprises must own their data, evals, model choices, and software layers, he stressed. The point is not to reject external models but to avoid handing judgment, processes, and replaceability to one provider. The enduring assets are an enterprise’s own data feedback loops, evaluation systems, and application-layer control. His view closely relates to agentic coding, model routing, and zero-data-retention inference, all serving the same goal: composable infrastructure rather than an uncontrollable black box.
https://x.com/rauchg/status/2076364176252191222
Aaron Levie, Box CEO
Aaron Levie argued that a central question of 21st-century enterprise architecture is how to maximize a company’s own IP when models possess extensive general intelligence. This IP includes not only documents but decisions, insights, workflow patterns, and best practices. The “bitter lesson” cannot simply dismiss these questions: when everyone can access frontier intelligence, differentiation comes from using it distinctively. Enterprises need workflow evals, routing across intelligence tiers, and captured traces to improve processes continuously. Their information should compound in value as AI improves, rather than be absorbed by the model layer. Consequently, substantial value remains to be created in the applied AI layer between underlying AI and enterprises. Companies solving these problems will be better positioned to win the next generation of enterprise workloads.
https://x.com/levie/status/2076338364635287637
Zara Zhang, Builder
Zara Zhang shared a concrete AI coding workflow: “meeting transcripts as PRDs.” She first discusses a feature’s implementation with colleagues, then sends the transcript to Codex to build a prototype from the discussion. Her summary is: “The meeting is the prompt.” Requirements need not first become a traditional PRD because the conversation already contains substantial intent, constraints, and implementation clues. Meeting records can therefore enter development directly rather than remain post-meeting notes. The example shows Codex’s value in turning unstructured collaboration into runnable prototypes as well as writing code.
https://x.com/zarazhangrui/status/2076300222884626754
Nikunj Kothari, FPV Ventures Partner
Nikunj Kothari criticized a common SF phenomenon: people say they are tokenmaxxing, with many subagents continually looping, but few directly answer “doing what for whom?” Simplicity and direction remain essential even in the frenzy of AI. More tokens, agents, and automation do not replace a clear problem definition. Before making tokens “go brrr,” builders should ask whether what they do for a living matters, because time is the only thing they cannot recover. He also called outbound sales one of the most humbling skills: it can be learned, but truly excellent practitioners inspire envy. Sales will become increasingly important, he believes, consistent with his emphasis on direction: AI increases execution efficiency but cannot replace choosing markets, understanding customers, and selling.
https://x.com/nikunj/status/2076458876816540144
Claude, AI Assistant
Claude’s official account announced extended Claude Fable 5 access across all paid plans, with Claude Code’s weekly rate limits remaining 50% higher through July 19. Users can still devote up to half their weekly allowance to Fable 5. After reaching that share, they may continue with usage credits or switch to other models within the remaining allowance. The update explains access, weekly allowances, and Claude Code limits together to avoid misunderstanding Fable 5 availability. For heavy users, extending 50% higher weekly limits through July 19 directly changes workflow capacity. Claude also linked further details on the rules.
https://x.com/claudeai/status/2076351399999557669
Podcasts
No Priors — How Nuclear Will Unlock Energy Abundance with Valar Atomics Founder Isaiah Taylor
Key takeaway: Isaiah Taylor sees nuclear energy’s bottleneck not in designing more elaborate, elegant reactors but in iterating, starting, and operating them like manufacturing, with a sufficiently short tick rate.
Isaiah Taylor is Valar Atomics’ founder and CEO, speaking from its Utah nuclear facility. His approach is unconventional: rather than begin with decades of nuclear research experience, he became frustrated enough with the industry’s slow progress to act himself. His family background includes memories of the Manhattan Project, but the trigger was realizing the US had largely stopped building reactors after the 1970s and the industry had not restarted fast enough.
His first judgment links US nuclear stagnation to Three Mile Island, while saying the event itself caused no deaths, injuries, or public radiation dose; the real rupture involved public confidence and industry capability. The US subsequently shifted from strength in large civil infrastructure to advanced manufacturing. Restarting nuclear should therefore follow manufacturing complex equipment rather than reproduce the civil-infrastructure path of the 1960s: manufactured rather than constructed.
His second judgment is that today’s nuclear industry resembles a modeling-and-simulation industry. Many companies focus on the most refined, efficient, complex designs, but Taylor sees a “Toyota Camry problem,” not a Lamborghini problem. He said directly: “We don’t wanna make Lamborghinis. We want to make a very simple, very cheap, very safe reactor that we can make literally tens of thousands of.” Complex designs might gain some performance, but manufacturing a thousand simple, safe, inexpensive reactors would let Valar win on cost.
The third concept is tick rate: the interval from founding to first splitting an atom, then the second and third times. Valar took two years and four months from Delaware incorporation to its first split atom, and approximately seven months from Project Nova to splitting an atom again in the current reactor. Taylor aims to reduce the interval through six months, four months, and one month to minutes. He expects eventually to start a new reactor every few minutes, because nuclear economics depend less on uranium costs than on producing plants quickly and cheaply.
The fourth highlight was the NVIDIA demonstration. Valar connected an NVIDIA Blackwell system directly to its reactor, calling it the first-ever AI chip powered by a nuclear reactor, and used it to host nuclearwebsite.com. The site displays how many uranium atoms were split to deliver the page, and Valar merchandise can only be purchased there. The demonstration makes the compute-energy relationship concrete: an AI chip drawing directly from a nuclear reactor rather than an abstract discussion of data-center power.
Taylor sees AI-driven electricity demand as more than a short-term tailwind. Energy is a commodity whose demand depends on price; lowering prices creates new demand. At 1 cent or even one-tenth of a cent, people would invent new uses. He calls this hyper techno industrialism: as AI converts human inputs into energy consumed by robots and automation, energy becomes a more central constraint on manufacturing, transport, and physical-world expansion. For AI builders, compute limits eventually meet power limits, and Taylor is targeting that underlying input.