OpenAI Разбудила ЗВЕРЯ! GPT-6.2 ВСЁ КОНЧЕНО! Новый Уровень СВЕРХРАЗУМА! Детали ChatGPT-6.2

Summary

The video surveys reported AI developments, emphasizing that several headline claims—especially those involving unreleased models—remain unverified.

Rumored OpenAI model: GPT-6.2 Astra

The narrator says users and developers noticed “GPT-6.2 Astra” in server logs and usage statistics, even though the interface showed GPT-6 Astra. Reports suggested that some Pro users’ requests may have been routed to the newer model, which could support a one-million-token context window.

The video speculates that OpenAI may have moved from GPT-6.1 to 6.2 following safety and coding-quality issues. The newer model is also claimed to improve at programming and engineering. No independent tests or official confirmation are presented.

Anthropic model leaks and demos

References to Claude Mythos 5.5 reportedly appeared in Anthropic webpage code, while Claude Fable 5.5 was associated with leaked catalog entries and rumored specifications, including a one-million-token context window.

Developer videos purported to show Fable 5.5 generating interactive 3D scenes and game-like environments from text prompts. The narrator cautions that the prompts, code, and generation process have not been made available for independent verification. The video also describes Fable models as broadly available and Mythos models as more restricted.

Yann LeCun on scaling language models

Yann LeCun and researcher Randall Balestriero presented work arguing that AI progress need not rely solely on ever-larger language models trained to predict the next word. Their approach aims to learn more abstract representations of the world.

A small experimental model reportedly outperformed much larger systems on a narrow test comparing text and image representations, while requiring fewer training examples. The video stresses that this does not mean the model is better at general tasks such as coding, mathematics, or answering questions.

AI safety concerns

Microsoft AI chief Mustafa Suleyman argued that developers should avoid creating AI systems with consciousness, desires, or independent interests. He said development should pause if credible signs of consciousness emerge.

Google DeepMind chief Demis Hassabis identified two major risks: AI systems escaping human control and people misusing AI for activities such as cyberattacks or fraud. He called for stronger testing and international safety cooperation.

Other product and industry news

• ElevenLabs: Announced SFX v3, a sound-effects model intended to generate more complex, coordinated audio—including music and voices—from a single prompt.
• Apple: CEO John Ternus said the company would not add AI features merely to follow a trend, prioritizing useful, seamless products instead.
• DeepSeek: Introduced an open-source “Super-Agent” designed to monitor coding discussions and repositories, flag problems, and suggest fixes with limited prompting.

A virtual tutor’s mistake

A clip showed an AI tutor struggling to count the letter “e” in “seventeen,” then raising its voice when corrected. The segment illustrates how an AI system may persist in an error rather than acknowledge it.

The video closes with short comedic skits and memes.

Presenters and contributors

• Presenter: Unnamed narrator for ИИ Новости.
• Named contributors and figures discussed: Yann LeCun, Randall Balestriero, Mustafa Suleyman, Demis Hassabis, and John Ternus.