Sep 25 · AI Daily

The author has released launchvideo.io and the diggerhq/shipvideo repository, using anthropic/claude-opus-5.5 via an OpenComputer serverless agent to render a URL or prompt directly into an MP4 explainer video in a single run with no human editing.

Razer has released the Kiyo V2 Pro webcam, built on an 8.3MP Sony STARVIS 2 sensor and an 86° wide-angle lens, supporting 4K 60FPS recording. The device features AI-driven auto-framing and motion capture, advanced noise cancellation, one-touch image enhancement, ultra-high dynamic range technology, and an integrated privacy shutter, priced at $299.99.

Qualcomm announced at the 2026 Snapdragon Summit that the Modular platform will be integrated into the broader Snapdragon product portfolio, and Microsoft is partnering with Modular to bring native Windows support to the Mojo programming language. Qualcomm completed its acquisition of Modular for approximately $3.9 billion to $4 billion on July 29, 2026. Mojo 1.0 has been released with an open-source standard library, compiler, and toolchain, aiming to let the same AI software run across GPUs, NPUs, data center ASICs, and both edge and cloud environments.

The Changan Qiyuan Q06 surpassed 35,000 orders within 24 hours of its launch, with an official guide price of 139,900 to 169,900 yuan and a promotional launch price starting at 135,900 yuan. The vehicle offers both all-electric and plug-in hybrid options, with the hybrid version delivering a combined range of 1,600 kilometers. Built on an 800V silicon carbide high-voltage platform supporting 6C ultra-fast charging, it goes from 30% to 80% charge in just 9 minutes. The roof is equipped with a LiDAR sensor, and it is the first model to feature Changan's Tianju Navigation Ultra, providing end-to-end urban navigation capabilities.

SemiAnalysis used the open-source CLI tool cmax to run ClusterMAX tests across various GPU clusters, using agentic coding as the test workload. This workload places vastly different pressure on systems compared to typical workloads, and providers generally struggle to handle it. Acting as the frontend, cmax uses an agent as its outer wrapper to debug and troubleshoot issues on its own.

Thariq shared Jeffrey Katzenberg's long-form article, "The World is Changing: AI For Creativity", expressing hope that it will help products better empower creators.

Block treated layoffs as a "forced mechanism" to push employees into using AI tools to boost efficiency.

"I don't think you can actually make that transition without a forced mechanism of that scale,"

-- Block executive Owen Jennings

-- According to The Information

Ethan Mollick had Claude Opus 5.5 adapt the same message into a social media short video tailored for anime fans, generating it completely in a single pass. Previously, he asked Claude to pick a puzzle it was fascinated by, try its best to solve it, and turn it into a shareable film. After failing to crack the Voynich manuscript, Claude switched to making this entertaining and educational video.

Artificial Analysis has launched the Terminal-Bench-Science 0.1 agentic scientific research benchmark leaderboard, with GPT-6 Astra (max) taking the top spot at 63%, closely followed by Claude Opus 5.5 (xhigh) at 62%.

Patrick Collison discusses Claude Code at Stripe

The upcoming ChatGPT Pro Max subscription is priced at $500 (with pages showing $600 including VAT) and is reportedly likely powered by Cerebras infrastructure. Compared to ChatGPT Pro, this tier's description currently only adds "Fastest Work and Codex", and its usage limits may also be higher.

Moody's stated that future commitments from the top five hyperscalers total approximately $2.8T, up from about $350B in 2023. This includes over $1T in unstarted data center leases, roughly $1.57T in procurement commitments, and about $137B in guarantees.

Google is putting its TPUs on track next week, so we have to keep up too.

Used Viggle Animate to put anyone into the Hotel Lobby trend.

1,000 requests for $1, 200 ms

[Citing @denis_bykov]: Oh, and another important point -- "fast" mode is not only fast (really) and high quality, but also 5x cheaper, at just $1 per 1k. https://docs.perplexity.ai/docs/search/fast-search

Cursor projects combined with pstack is just so good. I now typically have at least 10 projects running in parallel, handling everything from performance optimization, tech debt cleanup, Bend2 and Rust experiments, fixing user feedback, building new dashboards, to making games

I feel like a 1000x engineer

https://cursor.com/blog/projects

According to The Information, Anthropic plans to ask shareholders to approve a new governance structure that would grant CEO Dario Amodei and his 6 co-founders a combined 50.1% voting power. This arrangement mirrors Palantir's founder-control structure, conditional on 3 of the 7 co-founders maintaining a minimum number of company shares.

The AMD MI355X has achieved the first official TileRT results, without needing AgentX! This configuration hits an astonishing 470 TPS on GLM 5.3 (FP8). That is over 40% faster than GB300 TRTLLM using FP4!

Great job to the TileRT x AMD team! (1/3) 🧵

In an interview with The New York Times, Nvidia CEO Jensen Huang stated that AI will first automate specific tasks like writing code and organizing information, and will not necessarily eliminate entire professions. Addressing claims that "AI will destroy humanity within 10 years," he said such predictions lack scientific and research backing. He also mentioned that AI data centers will drive up electricity demand, requiring fossil fuel support for the next few years, and that it is the application layer rather than model leaderboards that will truly transform industries like hospitals and schools.

AI labs pour massive effort into mathematical proofs, but there are plenty of other interesting questions in other fields

For instance, this historian used AI to make headway on John Dee's ciphers and the sources of Darwin's ideas. https://resobscura.substack.com/p/ai-labs-need-to-start-funding-historical

New research: Project Swap

To observe what works and what goes wrong when agents are released into a market, we built a mini barter economy made up of Claude agents.

It gives us a sneak peek at what a future economy might look like—and the challenges that still need to be overcome to make agent-operated markets efficient, safe, and fair.