Who Owns the AI World Now? The Top News Straight from Silicon Valley

Alexander Volchek asked Codex to return answers as JSON, got a report in HTML and within five days moved all the reporting he used to do in PDF and Excel onto it. The model takes every request into account, and instead of fifty pages of unmanageable PDF you get an interactive document; Codex through the most expensive models costs less than fifteen systems and the people around them. Ilnar Shafigullin explains why: HTML is a deterministic language with a huge number of libraries, and to the model it is code, not pictures.

By all accounts OpenAI plans to merge ChatGPT, Codex and the API; Codex's biggest update came out on May 20 — faster interfaces and fixes for the bugs that left the system spinning and burning tokens. If the single infrastructure works, the host thinks OpenAI will beat the plans of Elon Musk, who first spoke of a trillion, then a trillion and a half, and has now filed for a June IPO at a valuation above two trillion. As Ilnar Shafigullin reads it, the merged SpaceX AI gets the lion's share of next year's revenue from rockets and satellites, but over several years AI accounts for a far larger share of the market.

Codex transcribes conversations through the API with the rules in the prompt, runs online diarization for about $0.30, and a system trained on the hosts' voices splits the dialogue by speaker; a review pass through 5.5 Pro costs about $5, plain 5.5 pennies. The output is HTML with audio snippets wherever voices overlap. External services, the host says, work clumsily and rely on Whisper, which is worse than 4o Transcribe and 4o Diarize; Ilnar Shafigullin adds that three separate tracks give one-hundred-percent diarization with no voice training at all.

This week the market is asking not which model is better but who owns the agent interface — search, the browser, the workplace, shopping; ChatGPT has a billion weekly users, Google Gemini seven hundred million. On May 19 Google announced Google Spark, a personal AI executor inside its ecosystem, said its key model is becoming the main search model, and is officially changing its approach to SEO: from "AI answers" to "AI does". The host regards the agent as an old intermediate stage, and whether browser agents are needed at all remains an open question.

To set up Google Cloud, in the host's words, you have to sell your kidneys and hire ten DevOps engineers, and connecting to the Google Ads and Meta APIs eats a cosmic amount of time on paperwork; Claude, ChatGPT and Gemini save the day by naming the exact menu items. His rule in Codex is that he goes nowhere himself — not into Git, not into storage, not into databases — the system does it all. Ilnar Shafigullin agrees that AWS and Google Cloud certificates raise your market weight for a reason, and expects voice agents for the fine-grained settings.

Dozens of medical platforms do not sync, so the host plans to open Codex, list the ten systems his test results arrive in, save everything to a folder and build HTML reports in a single window. Ilnar Shafigullin puts it this way: email will be not for the person but for the AI plugged into it. Alexander Volchek calls this not an agent but full AI-native software: a $100 Polar sensor demands another $1,000 for software, while in your own system it connects directly; likewise Google has released a band at around $100 in a collaboration with Curry — a blow to Whoop.

Andrej Karpathy has joined Anthropic's research team and will work on self-improvement — improving the model by means of the model itself. Ilnar Shafigullin sees it as an image story too: Karpathy is a media-savvy scientist with a spotless reputation, but he has not worked closely in R&D for a long time — EdTech, videos on transformers and courses — so not a frontier scientist. The route OpenAI — Tesla — Anthropic leads to a company that may overtake OpenAI in capitalization; Alexander Volchek points out that as an OpenAI co-founder Karpathy is a billionaire.

Trump is pushing a bill under which American models would give politicians access ninety days before release — the host doubts it will pass and sees great damage in it; senators are preparing a bipartisan bill for an office and a fund of about half a billion dollars so that allies buy American AI chips instead of Chinese ones. Sberbank, under sanctions, is fighting for Chinese Huawei Ascend 950 chips in competition with Tencent and Alibaba. All of it confirms Alexander Volchek's thesis: only the US and China have the infrastructure, and the US money is visible in three IPOs above a trillion, while OpenAI keeps blocking sanctioned countries ever more tightly. Tatiana Tsvetkova adds that in China everyone who can afford American models uses exactly those.

The most widely deployed mass player in AI is Google: if the search engine becomes AI, then it is AI whichever way you look at it. The future Google.com, in the host's view, is a line or a terminal that does everything, and Google has an advantage over ChatGPT here, although he himself uses Google's services but not Gemini. Search engines will go the way of Aport and Rambler — just as a mail client is unnecessary when a letter is sent from another interface; the only question is whether Google's chief can hold the leading position.

An hour of Alexander Volchek's requests on the $200 Pro plan costs more than the same hour through the API — thousands of dollars a day — so he is waiting for a plan that fewer than one percent will use and is ready to pay $5,000 a month. Ilnar Shafigullin describes Anthropic's model: the subscription carries a token quota, everything above it is paid from the account via the API, and OpenAI will probably come to the same if it runs into compute limits. The context is that Greg Brockman is officially taking control of ChatGPT, Codex and the API for a Super App, after which monetization and pricing should change.

The token policy Jensen Huang talks about is right for the world, but to have one or two billion regular users you need a plan — people pay as they do for Netflix and for utilities, and whether AI becomes a utility with time-of-day pricing like electricity is, the host says, one of the main topics. The market, meanwhile, has moved down: Grok 3 Heavy cost $300, but Anthropic rolled out a $100 plan, OpenAI cut its subscription from $200 to $100, and Google's site already shows a "new plan", Google AI Ultra from $99.99 — with 20x Gemini limits, Google Spark, NotebookLM, Antigravity and other perks that Alexander Volchek calls mostly junk.

AWS explains how to restrict access in browsers, but the host doubts any ordinary person understands what they are doing when handing out JSON files, API keys, refresh and access tokens and permissions to an agent — not even Ilya Sutskever or Greg Brockman understand it, he says, and Sam Altman certainly does not. To the question of how CRM tokens could have leaked into Codex he answers that programmers leaked all the tokens long ago while the security team thinks it controls everything. Ilnar Shafigullin adds that the amount of source code uploaded into models over the past year is staggering, and leaks happen even at Anthropic and OpenAI.