OpenAI Codex learns by watching; Qualcomm eyes a $10 billion chip buy

OpenAI shipped a Record and Replay feature for its Codex desktop app: walk the agent through a workflow once — say, filing an expense or uploading a video with its thumbnail and subtitles — and Codex saves it as a reusable, editable "skill" it can run on its own. It is the most consumer-facing step yet toward computer-use agents that automate by demonstration instead of by code. The feature needs a paid ChatGPT account with Computer Use enabled and is not yet available in the EU, UK, or Switzerland.

Qualcomm is reportedly negotiating to acquire Tenstorrent, the AI-chip startup led by legendary designer Jim Keller, in a deal valued between $8 billion and $10 billion. Tenstorrent builds RISC-V-based processors it claims beat general-purpose GPUs on specific AI workloads, so the purchase would hand Qualcomm a credible data-center play against Nvidia and a path beyond smartphone chips. The talks are unconfirmed and could still fall apart, but they show how aggressively mobile-chip makers are now chasing AI silicon.

Google released the Agentic Resource Discovery (ARD) specification, an open, Apache-licensed protocol that lets AI agents look up which tools and services exist, decide which to use, and verify they are safe to connect to. Microsoft, Salesforce, GitHub, Hugging Face, Nvidia, and Amazon are backing it under a new Linux Foundation working group — a coalition that reads as an attempt to set the agentic web's plumbing before OpenAI and Anthropic do. Reference implementations are already live from Google Cloud and Hugging Face.

At its New York summit, Amazon Web Services rolled out two services aimed at the weak spots of production AI agents. Continuum, in gated preview, automates the full code-vulnerability lifecycle — discovery, ranking by business impact, proof of exploitability, and fixes — starting in a human-in-the-loop "learn mode" before it acts on its own. Context builds a knowledge graph from a company's databases, documents, and chats so agents make grounded decisions instead of confident wrong ones. Both target the gap between flashy demos and reliable deployment.

Aswath Damodaran, the NYU finance professor known as the "dean of valuation," cautioned that a downturn in AI could prove more painful than the 2000 dot-com collapse. His concern is the debt: the late-1990s boom ran almost entirely on equity, so the losses stayed with shareholders, but much of today's data-center buildout is financed with borrowed money — often from less-transparent private credit rather than banks. If returns disappoint, he warns, the distress would not stay contained. He is not calling a top, only questioning whether the spending can be repaid.

Signal president Meredith Whittaker used a Bloomberg interview to push back on the rush to anthropomorphize assistants like ChatGPT and Claude, reminding users they are statistical tools, not sentient companions. Her sharper warning was structural: autonomous agents, on-device scanning, and targeted advertising are converging into what she called a new architecture of surveillance, in which agents gain broad access to personal data and the power to act on it. She says she uses AI to format documents but not to think, so the writing stays her own.

Snap is carving out its in-house generative-video group into a separate startup called Dotmo, citing the high cost of doing the work internally. Dotmo will focus on AI that generates interactive gaming experiences; Snap takes a large equity stake and licenses its technology, while chief technology officer Bobby Murphy becomes lead investor and keeps his Snap role. It is Snap's second spinoff this year after its Specs smart-glasses unit — a sign of how expensive frontier AI has become for mid-sized players to carry alone.