X / Twitter
Thibault Sottiaux, OpenAI Codex and ChatGPT Team Member
Thibault Sottiaux said “vibes are strong” at OpenAI and that he had never seen the company so focused or running so smoothly. More informative was his description of ChatGPT’s “work” capabilities: he emphasized treating ChatGPT as something that actually completes tasks for you, rather than just a chat box. Examples include negotiating internet bills, unsubscribing from junk mail, and finding the best deals for things you want to buy or do. The key change is that all these tasks can be initiated with a single prompt from a phone, without users switching between multiple websites and apps themselves. He said ChatGPT completes at least 20 tasks for him every day, and the results still surprise him. For builders, the signal is that AI product competition is shifting from “how well does it answer?” to “can it connect to real personal affairs and carry them through to completion?”
https://x.com/thsottiaux/status/2081534792903147881
https://x.com/thsottiaux/status/2081444811647963244
Peter Yang, AI Tutorial Author
After speaking with users in Canada and outside AI circles, Peter Yang observed that ordinary users’ biggest concern is not running out of tokens. The first concern he heard was: Do I trust ChatGPT enough to give it access to personal and work accounts such as Gmail, Calendar, Google Workspace, and Microsoft Office? This reframes AI adoption from “compute and allowances” to “trust and authorization.” For many builders developing agents or personal assistants, users may be held back not by an inability to prompt, but by reluctance to expose valuable data and operational permissions. Product design must therefore do more than showcase capabilities: it also needs to explain boundaries, visibility, revocation mechanisms, and permission controls. In other words, connecting Gmail or Calendar is not a simple integration but an exchange of trust.
https://x.com/petergyang/status/2081555286817648738
Madhu Guru, Senior Director at Meta AI
Madhu Guru challenged the claim that “AI has yet to deliver obvious shipped impact in products.” He believes we are only in phase 1: companies with distribution are rapidly expanding into adjacent problem areas, and AI lets them execute faster and build features, such as clothes try-on, that previously required substantial custom software. The impact has not yet fully appeared at the ecosystem level because companies are still figuring out the playbook. In phase 2, he expects more net-new features and genuine product innovation. At that point, AI’s influence on the shape of the software ecosystem will become hard to deny. The lesson for builders is not to measure AI’s impact solely by the number of visible new unicorns today, but also to watch how existing distribution channels embed AI capabilities in specific workflows and consumer scenarios.
https://x.com/realmadhuguru/status/2081437850466451736
Amjad Masad, Replit CEO
Amjad Masad relayed an observation from a former Anthropic employee: hackers are more inclined to attack using heavily subsidized AI subscriptions from labs than open models. This reverses a common assumption in AI safety discussions: risks do not come only from open models; closed models and commercial subscriptions can also be abused because of subsidized pricing and ease of use. He did not elaborate on the evidence but called it an “interesting drop” worth noting. For builders, this means abuse prevention cannot simply be equated with restricting open weights. Whenever highly capable models are available cheaply and with little friction, attackers will choose the most convenient tool.
https://x.com/amasad/status/2081576172656456076
Guillermo Rauch, Vercel CEO
Guillermo Rauch said Vercel had co-signed the Open Weights and American AI Leadership letter. He views open source, data, protocols, and research as the foundations of today’s technological wonders and considers open weights the next natural frontier. This signals Vercel’s explicit support for open weights in AI infrastructure and the developer ecosystem. Another post described a concrete engineering experiment: compiling the Vercel CLI’s TypeScript to native code with scriptc. The resulting binary was 1.28mb, with average startup overhead of 1.5ms and average compilation time of 2.94s, using node:https, node:fs, node:path, node:os, and node:crypto. GLM 5.2 Fast translated the code; the final result embedded neither v8 nor QuickJS, was fully static, and could deploy successfully. For toolchain builders, this points to a new route for making TypeScript CLIs native: retaining TypeScript’s readability while gaining distribution and startup characteristics approaching those of native binaries.
https://x.com/rauchg/status/2081546513885622760
https://x.com/rauchg/status/2081517519303737559
Aaron Levie, Box CEO
Aaron Levie argued that enormous opportunities remain in bringing AI into the real world, because most enterprises need substantial support to apply model breakthroughs to actual workflows. Intelligence alone cannot transform processes; it must connect to real-world feedback loops. This includes integrating with enterprise systems, supplying the right data to AI, using appropriate UX to let humans make decisions at different stages, letting workflows improve data and models in return, and handling regulation and compliance. He noted that implementing an AI agent for client onboarding in a bank is entirely different from implementing one for contract review in a legal team. In critical industries such as life sciences, financial services, legal services, and manufacturing, AI has value only when it engages with real business in context. He calls this the applied AI layer and believes labs will supply some of it directly, but many opportunities will inevitably go to independent companies with deep industry expertise. More counterintuitively, he argues that stronger models will increase rather than reduce demand for this layer: the more ambitious the workflows being automated, the more complex the required industry connections, data, UX, and compliance layers become.
https://x.com/levie/status/2081491621162668207
Zara Zhang, Builder
Zara Zhang proposed measuring AI adoption not by tokens burned but by “the time from a user need appearing to the corresponding thing shipping.” This shifts the measure from consumption to delivery speed, making it better suited to evaluating whether teams have actually gained productivity. She also explained why there are so many AI tutorials: the more general-purpose a chat product is, the harder it is to use. Users get stuck at a blank input box because they genuinely do not know what to ask. In other words, more general AI products need more examples, workflows, and specific entry points to reduce the cost of getting started. She also shared her approach to posting on X: around 3 posts a day on average, taking no more than 15 to 20 minutes, without overthinking—she posts when an idea occurs to her. Most material comes from things she has already said to others offline. Together, these points offer builders a clear product and content approach: measure delivery cycles, reduce blank-box anxiety, and turn real expressions already tested in conversation into public output.
https://x.com/zarazhangrui/status/2081627581997269192
https://x.com/zarazhangrui/status/2081627109299310684
https://x.com/zarazhangrui/status/2081304884469809295
Nikunj Kothari, FPV Ventures Partner
Nikunj Kothari’s one-line judgment was: “proof of prompt is soon going to replace proof of work.” He did not mean that prompts would magically replace labor, but that as AI tools enter the building process, the ability to express intent clearly, break it into executable instructions, and get a system to produce reliably will become a new demonstration of competence. Previously, proof of work showed what you had completed yourself; in the future, proof of prompt may show how you direct models, tools, and context to achieve a result. This has implications for hiring, portfolios, and teamwork: a builder’s record of output may expand from code diffs to the relationship between prompts, tool-call paths, and final deliverables. It also reminds teams to treat high-quality prompts as reusable work assets rather than temporary text.
https://x.com/nikunj/status/2081383934928068619
Dan Shipper, Every CEO
Dan Shipper said he would take a week off to write a definitive history of how Codex came to be. The article will draw on his in-depth interviews with OpenAI insiders and is planned for publication in Every in a few weeks. He said he would share breadcrumbs and lessons during the writing process. For builders following AI coding agents, this may be an internal history worth tracking, because it aims to reconstruct the organizational, technical, and product judgments that shaped Codex rather than simply review the product. The current material contains no specific interview content, so only the article’s direction and publication plan can be confirmed. It is worth watching how Codex has become not just a tool name but a key case in the evolution of AI programming products.
https://x.com/danshipper/status/2081412243388788988
Sam Altman
Sam Altman demonstrated ChatGPT work’s capabilities with a long prompt sent from his phone: use his full chat history to plan a long weekend trip for 8 friends, propose the best 3 options, build a full-stack site where 9 people can coordinate their preferences and reach a decision, then make reservations once they agree and draft an email in Gmail that he can send to his friends. His verdict was: “it...just worked.” The example combines personal-context access, group decision support, website generation, follow-up bookings, and email drafting. The point is less the trip planning itself than ChatGPT’s transition from an advisory tool to an execution system linking multiple actions. It also echoes Thibault Sottiaux’s observation: a single phone prompt is becoming an entry point for automating personal affairs. For builders, the real product challenges will lie in cross-application permissions, state tracking, user confirmation points, and group collaboration workflows.
https://x.com/sama/status/2081396796174282900
Podcasts
The MAD Podcast with Matt Turck — OpenAI’s Compute Chief: We Can’t Build Fast Enough | Sachin Katti
Key takeaway: AI’s bottlenecks are rapidly moving from the model layer into the physical world. OpenAI’s greatest concern is not insufficient demand, but compute, power, cooling, networking, and construction capacity all failing to keep up.
Sachin Katti now leads industrial compute at OpenAI. He was previously a Stanford professor, serial entrepreneur, and Intel CTO. He described OpenAI’s ongoing compute buildout as one of the largest infrastructure projects in human history, with major compute decisions being made inside the company every day that might previously have taken months at Intel. Matt Turck cited directional figures of approximately $50,000,000,000 in OpenAI compute spending this year and potentially $700,000,000,000 across the industry. Sachin responded that these figures would keep growing, because construction today becomes compute that companies such as OpenAI can consume one or two years later.
First, an AI data center is no longer fundamentally a traditional cloud facility, but a large supercomputer. Sachin offered a vivid explanation: “A data center is a giant factory that turns electrons into tokens.” Stronger models and more complex tasks require larger computers. These chips become extremely hot and cannot rely on air cooling alone; liquid cooling is required across data halls, chips, connecting cables, transformers, and other layers. He stressed that liquid cooling is not new. The new challenge is making it reliable, affordable, and scalable at this scale. Cooling efficiency directly affects how hot chips can run, and the more reliably they can operate at high temperatures, the more memory bandwidth and flops they may deliver, ultimately producing more intelligence.
Second, electricity is a core supply chain rather than a background resource. Like other companies, OpenAI initially connected to the grid, but it has now begun investing in generation and transmission infrastructure. Sachin said that with every data center they build, they make a hard commitment: rather than take existing electricity from the grid, they invest in additional power so the data center can consume that new supply. This suggests AI companies are moving from buying cloud services into participating in energy infrastructure construction.
Third, OpenAI is developing its own compute muscle rather than relying solely on partners. Sachin said OpenAI has always believed compute is the foundation of intelligence. What has changed is that at the current scale, the company cannot simply wait for partners to provide compute; it must participate more actively in building and securing what it needs. Commercially, however, partners including Microsoft, Google, Amazon, and Oracle will still handle much of the construction, with OpenAI committing as a tenant and offtaker to consume the compute.
Fourth, speed in custom silicon comes from the team, partners, and visibility into workloads. OpenAI’s Jalapeno chip took approximately 9 months from design to tape-out, which Sachin called one of the fastest timelines he had seen in his career. Factors included team members who had designed Google TPUs, Broadcom’s strong execution record in XPU ASICs, and OpenAI’s knowledge of likely future model workloads, which shortened many chip-design decisions. More importantly, AI has begun assisting chip design and optimization. He believes a world where “AI designs the systems needed to train and run the next generation of AI,” including chips, is not far away.
Fifth, reliable networking for clusters of 100,000 GPUs requires new protocols. Sachin introduced MRC as a new routing technology for scaling enormous cluster fabrics. In large training jobs, GPUs communicate continuously, and the vast numbers of links, switches, and NICs make failures inevitable. MRC uses multipath spraying: packets travel simultaneously over multiple routes between two chips, and progress can continue if any route succeeds, so training jobs need not worry about underlying network failures. He stated the goal clearly: the network should be abstracted away, and training jobs should not stop because of routine failures.
The final, easily overlooked bottleneck is people. Sachin explicitly mentioned shortages of skilled workers such as electricians and plumbers, saying hyperscalers and labs would actively hire people with these skills. Another commercial shift is guaranteed capacity, which essentially means guaranteed tokens: enterprises lock in a certain dollar value of intelligence supply in advance. In a world of scarce compute, tokens will carry a lasting premium. As intelligence becomes a basic input to enterprise operations, securing supply will become normal business hygiene.