GPT-6.2 a Letdown? Claude Advances, the NVIDIA Crisis and a New Tesla | AI News
Has GPT-6.2 disappointed, and why is Claude Code overtaking Codex — and what do Jensen Huang's departure from NVIDIA, shopping through ChatGPT and a car you can buy without a license change for the market?
Anthropic went public at a valuation of about 1.8 trillion dollars, and ChatGPT got 6.2 Astra and Sol. Ilnar Shafigullin noticed that 6.1 got dumber shortly before 6.2 came out, while the new Astra refuses ordinary tasks for cybersecurity reasons; Alexander Volchek sees severe degradation of the ordinary chat and a slow Codex, and Ultra Fast in Astra speeds work up four times instead of eight. Claude Code, in his experience, does design at the top level, and Anna Volchek notes that the debate over models "getting dumber" repeats with every release and describes how she used up all her limits. NVIDIA's board ruled out Jensen Huang's return, and Anthropic's models wrote software for AMD chips in three days — work that would previously have taken half a year. ChatGPT Wallet through Dots carries a purchase through to the final cart and finds hidden charges itself, Instant Checkout takes OpenAI into marketplaces, and the money for a failed purchase in the US, according to Alexander, is returned by the bank. The Cybercab can be bought without a license: Anna would buy one, Ilnar for now is calmer with a live driver. Instagram is testing a standalone Direct, and Ilnar and Anna see the future of messengers in integration with models — as with Claude Bot in Slack.
What to watch for
Key takeaways
Anthropic went public on the day of recording: preliminary reports spoke of one and a half trillion dollars, now of 1.8 trillion. Alexander Volchek expects the listing to bring the market a new storm, in no clear direction, but there will definitely be more money in artificial intelligence.
Having switched tasks to 6.2 Astra and Sol, Alexander is becoming more and more disappointed in Codex. Having looked into what US tech companies use for programming, he sees that they use Codex too, but Claude leads, and even at Google, people say, they program with Claude Code.
Ilnar Shafigullin learned about the release of 6.2 because 6.1 got much dumber on his usual tasks. Switching to 6.2, he got back a model that did its tasks more or less properly.
On some of the tasks Ilnar calmly did with 6.1, 6.2 Astra refuses, citing its rules. He concludes that the model was tightened even further on cybersecurity — although he has full access to the codebase of his own project.
According to Alexander, the Ultra Fast mode in Codex works eight times faster, but in 6.2 Astra four times, and the ordinary Fast twice rather than two and a half times. Ultra Fast also instantly eats up tokens, and the ordinary ChatGPT chat, in his view, has degraded severely over two or three weeks.
Claude Code did top-class work on module design for Alexander, while the ordinary chat in Claude did disgustingly. Claude Design is closing: if you write "make me a page design" in Claude Code, it will launch internal modules.
Anna Volchek, in a rare case for her, used up her entire limit on every platform, including Pro mode in ChatGPT and Codex, and upgraded her skills for economical use. The day before, Alexander used up a whole Claude Ultra plan in a single day.
NVIDIA's board ruled out Jensen Huang's return, Congress requested explanations about chip deliveries, and the US Treasury Secretary said the company's head should be hired from outside. Alexander fears discord: Huang had many personal agreements, for example on allocating capacity.
Before, companies from China, Iran or Russia could buy infrastructure with chips banned for them in Singapore or the US; according to Alexander, a clarification released this week explicitly prohibits sales through such middlemen.
Anthropic's models fully wrote the software for AMD's new chips in three days — before, this would have taken six months of work by a whole team. If moving to other hardware no longer requires months of engineering work, Alexander believes, choosing AMD becomes much easier.
Ilnar explains that a new graphics card delivers about 30 percent of its potential at launch, and libraries like CUDA are refined over six months to a year. Shrinking this gap is useful for NVIDIA too, but for AMD it is a huge plus: NVIDIA's hegemony rests on the infrastructure around the hardware.
ChatGPT Wallet works in Dots: when Alexander asked it to buy a present it had found, Dots warned that there were also taxes and other charges and sent a final cart, like Apple Pay. Alexander made three purchases this way, including on sites he had not known before.
OpenAI integrated directly with several American companies, Shopify in particular, so that products can be bought right in ChatGPT, and Shopify's shares rose immediately. In Alexander's view, OpenAI wants to get into the marketplace market, while responsibility for the order lies with the seller.
Ilnar asks whether ChatGPT will help with a return if a product from an obscure website is unsatisfactory. Alexander replies that in the US it is enough to open the transaction in the bank and press "refund"; Anna finds out that in Poland the bank first reviews the dispute.
Anna would definitely buy a Cybercab, while Ilnar for now feels calmer with a live driver, although he himself calls this fear irrational. Tesla, he recalls, drives better where it was trained more, and in atypical weather the quality is lower.
The main thing about the Cybercab, according to Alexander, is that it can be bought without a license. Anna sees this as a solution for people who cannot pass the driving test because of fear or anxiety; the car goes to charging and service on its own, but requires a subscription and is tied to its owner.
Alexander is sure that with self-driving cars on a mass scale there will be no traffic jams: the cars will distribute themselves and tell you in advance when to leave. Anna also sees this as a solution for parking: the car can drop you off and pick you up later.
Instagram is testing a standalone messenger, Direct, without a mandatory social network account. Ilnar thinks messengers are not a market for OpenAI and Anthropic, but notes that Anthropic after its IPO depends on the market and irrationality may appear in its behavior too.
Ilnar gives Anthropic's example: Claude Bot has been added to practically all work chats in Slack and carries out assignments. Most likely, he believes, such integrations will develop rather than one common messenger; Anna sees integration with models as the future of messengers.
What this episode is about
The Sunday ToTheMoon episode in the news-of-the-week format is hosted by Alexander Volchek, Ilnar Shafigullin and Anna Volchek. They begin with Anthropic's stock market debut on the day of recording: preliminary reports spoke of one and a half trillion dollars, now of 1.8 trillion, and Alexander expects it to bring the market a new storm.
The main topic of the first part is GPT-6.2 Astra and Sol, released about a week after Fable 5.2 and Mythos 5.2. Alexander is becoming more and more disappointed in Codex for development and sees that Claude leads at US tech companies. Ilnar learned about the new version because 6.1 got much dumber, while 6.2 Astra began refusing ordinary tasks, citing its rules — clearly cybersecurity was tightened even further.
Alexander talks about the severe degradation of the ordinary ChatGPT chat over two or three weeks and about a slow Codex: Ultra Fast in 6.2 Astra speeds work up four times rather than eight and instantly eats up tokens. The ordinary Claude chat also did poorly, but Claude Code did module design at the top level, and Claude Design is closing — its functions are moving into Claude Code.
Anna notes that the conversation about models "getting dumber" repeats every time new versions come out, and says that in one week she lived through all the scenarios at once: some tasks the models began to do worse, others faster and better. In a rare case for her, she used up all her limits on every platform and upgraded her skills for economical use.
Then comes NVIDIA: the board of directors ruled out Jensen Huang's return, Congress requested explanations about chip deliveries, and the US Treasury Secretary proposed hiring the company's head from outside. Alexander fears discord because of Huang's personal agreements, and mentions a clarification banning chip sales through middlemen and OpenAI's agreement with NATO countries.
Alexander calls the most explosive the news that Anthropic's models fully wrote software for AMD's new chips in three days — work that would previously have taken six months for a whole team. Ilnar explains why this matters for NVIDIA too: new graphics cards deliver about 30% of their potential at launch, and NVIDIA's hegemony rests on infrastructure like CUDA.
After a short news item about the free errands mode in Gemini, Alexander talks about three purchases through ChatGPT Wallet: the wallet works in Dots, carries the purchase through to the final cart and warned him itself about taxes and charges. Instant Checkout, an integration with Shopify and other stores, in his view takes OpenAI into the marketplace market. Ilnar worries about returns from obscure websites, Alexander explains how in the US the bank returns the money, and Anna checks how it works in Poland.
Almost a quarter of the episode is about the Tesla Cybercab with no steering wheel and no driver: Anna would buy one, Ilnar for now is calmer with a live driver and recalls the limitations of autonomous driving, and Alexander is confident in full self-driving and that with self-driving cars there will be no traffic jams. They discuss riding without a license, the subscription, the binding to the owner, self-charging and Waymo.
In closing — the standalone Direct messenger that Instagram is testing, and the future of messengers. Ilnar believes it is not a market for OpenAI and Anthropic, Alexander dreams of a system that sees all conversations and works with Dots, and Ilnar and Anna see the future in integrations — like Claude Bot in Slack work chats.
The episode is assembled from news but rests on one observation: the impression of a model depends on the task more than on the version number. Ilnar saw 6.1 get dumber, Alexander saw the ordinary chat degrade and Codex slow down, Anna saw the same models do worse in some scenarios and better in others; and all three admit that this debate repeats with every release. The assessments of Anthropic's IPO, the consequences of Huang's departure, the future of the Cybercab and of messengers are the hosts' opinions on the day of recording, not company forecasts; the rules of ChatGPT Wallet, Instant Checkout and the Cybercab are worth checking against the official terms.
Episode transcript
The episode is in Russian; below is an English reading guide to the transcript (the full EN transcript is a machine translation). Voice matching applied to 184 segments: 0 identified, 0 mixed, 184 probable, and 0 unresolved.
Read transcript on a separate page
Loading…