AI News: OpenAI Made a Massive Move Against NVIDIA

OpenAI showed first results from its in-house inference chip, claiming up to 104x performance on open-weight models and a wider advantage on its internal frontier models. It is for inference only, not training.

HFHugging Face (Nvidia acquisition)NvidiaUnverified

The Information reports Nvidia agreed to buy Hugging Face, unconfirmed by either company. The creator sees it as Nvidia betting on open-weight models, and cites Vercel gateway token data showing open models growing.

AMApple M5 UltraAppleAvailable in EU

Apple announced M6 and M5 Ultra chips aimed at AI compute, with 4.5x peak GPU compute over M3 Ultra and up to 512 GB unified memory. The 256 GB, 1 TB configuration costs almost $11k.

G5GLM 5.3 FlashZhipu AIAvailable in EU

Open-weight model that launched in stealth as Ox Alpha. It scored 63.4 on the creator's preferred coding benchmark, beating Opus 4.8, and is cheaper than similarly intelligent models. On Beautybench it scored 5.4, ranking 25th.

Q3Qwen 3.8 Flash NextAlibabaAvailable in EU

Open-weight 125B-parameter model, larger than Qwen3.8 27B. It scores about 56 on Artificial Analysis and ranked 19th on Beautybench at about a penny and 26k tokens.

CSChatGPT SitesOpenAIUnverified

Sponsored segment: in the Codex app, Sites builds and hosts a website with authentication and a built-in database. The creator used it for a social media dashboard, with site analytics and WebMCP support for agent actions.

GOGemini Omni 1.1 FlashGoogleUnverified

Updated Google video model at the same price (10 cents for 720p), with 360p to 4K output, start/end frames and 10 seconds of prior context. It tops the Arena text-to-video leaderboard. The creator's Flow test showed glitches such as a robot's head popping off, though scene consistency was good.

G3Gemini 3.5 TranscribeGoogleUnverified

Google's transcription model, described as a Whisper alternative that is more accurate than others. It is available in the Gemini app on macOS, Android, enterprise platforms and the API.

AMAI Mode travel bookingGoogleUnverified

AI Mode in Search can now track flight prices, show them in price and points, and book hotels directly.

GNGemini NotebookGoogleAvailable in EU

Gemini Notebook, formerly NotebookLM, can now add ebooks purchased from Google Play Books to a notebook. The creator notes he mostly buys on Kindle.

ANAntigravityGoogleUnverified

Antigravity added remote control, letting you connect through a web browser to sessions running on other machines. Other AI IDEs already had this.

CWChatGPT Work writing styleOpenAIUnverified

ChatGPT Work can now sign into websites on web and mobile without seeing your passwords. Plus and Pro users can also trigger tasks from changes in Slack, Gmail and GitHub.

GSGPT-5.6 SolOpenAILimited in EU

OpenAI's most powerful model is 20% cheaper for the next three months.

CCClaude Cowork built-in browserAnthropicUnverified

Claude is adding a built-in browser to Cowork, rolling out over coming weeks to Pro, Max and Teams on the desktop app. The creator says it mirrors Codex's browser.

CIClaude in ChromeAnthropicUnverified

Chrome extension now generally available that summarizes articles, fills in details and clicks around, though the creator calls it slow.

CMClaude memory in CoworkAnthropicAvailable in EU

Claude's memory now works across chat and Cowork. Sensitive topics are not stored by default, and users can control what is remembered.

PCPerplexity Computer (local on Mac)PerplexityUnverified

A local agent harness that runs small models such as Nemotron and Qwen on the user's machine. It calls cloud services only when needed and with the user's permission. It targets powerful hardware like the DGX Spark.

SUSunoSunoAvailable in EU

In a segment on Dr. Dre saying he uses AI to produce songs, the creator shares his view on AI music. He enjoys Suno as a collaborator, for example adding drums to his own guitar riffs, but dislikes passing off prompted songs as one's own.

S1S1Skild AIUnverified

A robot foundation model that learns 10-minute tasks never seen before from a single video prompt, with no fine-tuning. Demonstrators wear wrist cameras and the two-armed robot reproduces the task.