Desk
Broad artificial intelligence coverage.
Block’s Dessa is a speech-to-speech model trained from scratch only on human conversation, with no text in the loop, and it runs locally on high-end MacBooks. Its demos are awkward in a human way. Block frames emotion-aware voice as a safety feature; the same ability is a lever for persuasion.
Part 2 turns TensorRT's tuning theory into knobs, a search loop and a hard accuracy gate, and Nvidia's own numbers show why one default build was leaving real speed behind.
Part 1 of Nvidia's TensorRT tuner series is really a confession: hand-written compiler heuristics stop scaling once every model behaves differently, and the fix is to measure rather than predict.
OpenAI’s second Habitat post covers the Azure Cosmos DB layer behind ChatGPT: a move from one shared container to many and back, read latency cut from about 120 ms to about 10 ms, and outages that showed managed services become less managed at its scale. It is also building its own storage system.
Since October 16, Google’s privacy policy carves an exception into its old promise not to use Drive, Gmail or Photos content for ads: data users make available to services that support personalized ads. People who linked those apps to Personal Intelligence may now see ads shaped by them. The safeguards are real; the bundled consent is the problem.
Nvidia's Microduck walkthrough shows agents taking a robot from a written prompt to a trained walking policy, and the tidy pipeline makes it easy to forget how much still has to be checked by people.
After publishers and advertisers in the US, EU and UK objected, Meta delayed enforcement of its new business terms from October 30 to November 20 to explain them better. Publishers say the problem is the contract, not the communication, and want robots.txt written into the terms.
Google is building agent infrastructure on the assumption that every agent will eventually be compromised, and we think that is exactly the right starting point.
You.com is betting that the hard part of open-weight AI isn't the model price, it's getting a security reviewer to sign off, and it's selling guarantees instead of discounts.
inDrive signed an MoU with Pony.ai and Nvidia to explore robotaxi testing in Kazakhstan, starting in Astana in the first half of 2027 with safety operators. It is a capital-light bet on cheap markets and snow, and inDrive’s earlier Riyadh pilot shows MoUs don’t always become rides.
Jensen Huang’s Astana visit came with commitments: Firebird says it holds a US export license for 45,000 Grace Blackwell GPUs, Kazakhtelecom is expanding clusters, and Nvidia is investing in Higgsfield AI. Sovereign AI for Kazakh is a sound goal; power, governance and US-policy dependence are the open questions.
OpenAI and NASK, which houses Poland’s AI Safety Institute, signed an agreement to explore joint safety, security and technical research. There are no deliverables yet, but NASK’s CERT Polska already used OpenAI cyber models to help find six MikroTik RouterOS vulnerabilities. The open question is independence.