AI news for builders and product teamsUpdated Nov 17, 2026, 14:00 UTC
AI news, straight from the source.
Releases, research, and ideas for developers and product teams. Updated every hour.
Across the feeds
Newest first
The EUROPA consortium raised pretraining throughput for a 951.9B-parameter Mixture-of-Experts model from 90 to about 1,300 tokens/s per GPU on 384 NVIDIA GH200 GPUs of JUPITER, the EuroHPC exascale supercomputer at Forschungszentrum Jülich. The gains came from PyTorch and TorchTitan optimizations, most of which were contributed back upstream.

Gondola, a platform that helps creatives get credit and get hired, has doubled its revenue while its team builds more with OpenAI's Codex, which raised its lead engineer's output by 50%, according to OpenAI.

Microsoft released a public preview of the logicapp extension for the Azure CLI, which lets developers create, run, inspect and test Logic Apps Standard projects from the terminal instead of relying on the VS Code extension. It adds 19 command groups and 64 commands alongside the existing 13 cloud-management commands.

Australia's government, led by Prime Minister Anthony Albanese, has announced a law that would impose mandatory safety rules on frontier AI companies for the first time, including independent audits, ongoing safety checks, published safety frameworks and immediate incident reporting, enforced by a new oversight body with steep penalties.

Google Cloud open-sourced TPU Sync, an Apache 2.0 C++ library that moves KV cache blocks and model weights between TPU HBM, host DRAM, shared storage and the network. Google reports median KV transfer latency for Qwen3 235B dropping from several seconds to about 230 ms in prefill/decode disaggregation.

At the AI Engineer Code Summit, TypeSafe AI CEO Diogo Almeida urged engineers to embed Jev, his non-generative "system one" model, directly inside application code and automate broadly. Glean, UiPath and Workday showed production and prototype uses, from zero-shot classification to sub-second document ranking tournaments.

MIT researchers have published a method in PNAS that generates synthetic training scenarios to build climate emulators that predict unfamiliar situations using far less data and only one full earth system model run. Traditional emulators trained on realistic scenarios failed the same adaptation tests.

TechCrunch has published a speculative shortlist of possible successors to Nvidia CEO Jensen Huang, after a Kalshi market put his odds of surviving the week at 30% and a New York Times report raised questions about his personal conduct.

Hugging Face published a guide with fine-tuning scripts and Docker containers for adapting the TRELLIS.2 3D generation model to a custom object category using LoRA, runnable locally or on Hugging Face Jobs. A footwear experiment trained roughly 9.3 million adapter parameters on 179 objects.

Apple researchers derive a closed-form bound on how much expected reward alignment can gain under a fixed KL budget, governed by a Jeffreys divergence term rather than the square-root KL scaling used in earlier analyses. Empirically, best-of-N nearly reaches this limit while PPO and GRPO fall well short.

OpenAI published a conversation with Sam Nickless, a partner and chief executive officer at Australian law firm Gilbert + Tobin, about how the firm is rethinking the business of law with AI.

The Wall Street Journal reports that Nvidia is about to fire Jensen, according to Gary Marcus, who says the stock is already down in after-hours trading. Marcus warns that an exit could unsettle the circular AI financing Jensen has overseen.

Nvidia CEO Jensen Huang admitted an extramarital affair and now faces reports of a second relationship with an executive who reports to him, a board reportedly preparing to oust him, and a possible congressional subpoena over an alleged Chinese spy.

AWS published a method for catching silent data corruption (SDC) in PyTorch DDP training on SageMaker HyperPod: three in-loop checks that identify a suspect GPU by comparing data-parallel replicas, then cordon and replace its node. The checks were validated with software-injected faults on a cluster of 4 ml.p5.48xlarge instances, 32 H100 GPUs.

The New York Times reported that Nvidia CEO Jensen Huang is suspected of a second extramarital affair, with Kimberly Powell, Nvidia's vice president of healthcare and a direct report of his. Late Wednesday the Wall Street Journal reported Nvidia's board was preparing to oust Huang.

Databricks says it has worked on detecting credential-driven attacks on its platform, though the announcement gives no technical details, versions, dates or results.

Databricks published a profile-style Q&A with Alasdair Anderson, a Lead Solutions Architect, opening with a question about what a typical week looks like in that role.

Nvidia's board is reportedly preparing to decide on CEO Jensen Huang's future at a meeting as soon as Wednesday night, following New York Times reporting of a second alleged extramarital affair, per The Wall Street Journal.

Google has released Gemini Decision 1, a decision-model API that answers yes/no, multiple-choice and rubric-scoring questions and unlike Jev also accepts images, audio, video and PDFs. It is priced at 10 cents per million input tokens with free output, matching OpenAI's GPT-6 Luna.

United Airlines will not allow personal AI assistants to book tickets on its flights, citing privacy risks, unintended purchases and federal rules airlines must meet with customers. The airline says it has not yet seen much demand for agent-based booking.