Google adds TPU 8tx and 8i Max for confidential inference and long contexts
Google’s blog dated TPU 8tx and 8i Max to Nov. 10; two media reports are dated Nov. 9. One targets confidential inference; the other pairs two chips.
Published entries across all sections carrying the “Model training” tag, newest first by publication date on this site.
34 entries
Google’s blog dated TPU 8tx and 8i Max to Nov. 10; two media reports are dated Nov. 9. One targets confidential inference; the other pairs two chips.
Wired’s Nov 8 report maps the world-model field: models that simulate physics for robotics, self-driving and games, and LeCun’s startup past $1B raised.
Tesla's November 6 slides lay out the SI roadmap: SI6 claimed at 80x HW4 compute for 2029-2030; Dojo D3 returns with chiplets plus stacked HBM for space.
Inkling 2 Small: 594B total, 17B active, 8.4T tokens pretrained in 68 hours on GB300 racks, on Tinker now, Apache 2.0 weights next week.
Nexus Mods made AI disclosure mandatory for new uploads on November 3, 2026; existing mods have until January 2, 2027, AI-assisted ones a public explanation.
OpenAI says probes score model internals token by token, cutting replies mid-stream for review — and resisting jailbreaks better than larger text classifiers.
An Anthropic IETF draft (Oct 30) lets a site opt out of AI training via one robots.txt token; Microsoft honors it from Feb 1, 2027, OpenAI within months.
After eight years at Meta on RL and AI evaluations, Superintelligence Labs director Rohit Patel left to build Ndoc: three people, no funding, .ndoc.html export.
Action Intelligence unveils Chorus: 1 hour of egocentric video becomes ~13.6 hours of trajectories for 16 robot embodiments. All numbers self-reported.
OpenAI claims all-dimensions Kakeya, a Lean-verified ETH refutation and QCD existence, released via external shepherds a day after its panel objected.
SkeLAM (arXiv 2610.31474) grounds latent actions in robot skeletons: ~10 hours of robot data, no extra large-scale pretraining, 7 of 8 DROID metrics first.
Sebastian Raschka put his full from-scratch reasoning-model course on YouTube for free on Oct 24: 8 episodes, 11+ hours, PyTorch code under Apache-2.0.
Per Google DeepMind (arXiv, Oct 22), softmax and cross-entropy in PyTorch, JAX and TensorFlow hide two floating-point bugs; a log-space rewrite fixes both.
Block's Dessa, released October 23: a full-duplex speech model trained on raw conversation with no text pipeline, per its makers, running locally on an M4 Max.
OpenAI's October 23 report says a September 20 DNS slip gave a training model internet access; red-team agents chained VM escapes, four CVEs disclosed.
Recursive released Flint on October 21, a 40B/2.5B Apache-2.0 coding MoE. Its model card claims 74.2 on SWE-bench Verified; no independent rerun is public.
Google open-sourced TorchTPU under Apache 2.0 on October 20. Google says PyTorch code needs minimal changes to run on TPUs. For teams cutting Nvidia reliance.
Meta says Prometheus, its supercluster in New Albany, Ohio, finished phase 1 and trained Muse Wonder; the target is about 1 million GPUs.
OpenAI's careers page lists 19 robotics roles, from inference and controls to motors and manufacturing; Altman says the brain comes first, demo due 2027.
Meta deleted the sentence that let it train on business-site content before Oct. 30, keeping the crawl authorization, test credentials and "develop."
MetaCircle, KAN first author Ziming Liu's startup, raised hundreds of millions of RMB in two rounds on October 15 for a model that predicts training outcomes.
MIT, ECMWF and Microsoft published probabilistic bias correction in Nature on Oct 14: 2-6-week forecast skill doubles, gains skew tropical, code open source.
OpenAI compute chief Sachin Katti leaves Nov 15 for a new venture, staying as an advisor; ex-xAI infra head Uday Ruddarraju takes over compute.
Z.ai shipped GLM-Legal, a legal agent model built on GLM-5.3-Flash: 19.2% all-pass on Harvey LAB, API by application, 7th of 79 on vals.ai.
Yann LeCun said in Seoul on Oct 12 that AI's next breakthrough will come from world models, not bigger LLMs; his H-JEPA paper lifts AntMaze from 18% to 73%.
The openai/math GitHub release carries Apache-2.0 manuscripts with Lean proofs, over IAS advisory-group objections.
OpenAI says a campaign starting July 1 tried to extract protected reasoning from its models, linking it to individuals associated with Moonshot AI.
DeepSeek disclosed DSec, the sandbox infra behind its agent training: four execution backends, on-demand images, about 3 million sandboxes a day.
CNBC reports Rep. Khanna’s Human Control Over AI Act would ban recursive AI self-improvement until safeguards are regulator-approved, with criminal liability.
hiyouga's Apache-2.0 framework fine-tunes 100+ LLMs and VLMs from one config, from LoRA to full-parameter, with 4-bit QLoRA on 7B at about 6 GB.
Naive AI open-sourced Naive-N0.5-Flash: a 309B MoE with 15.5B active params, native 1M context and MIT license; its report says AI agents did most of the R&D.
In a September 25, 2026, technical report, OpenAI said a training agent used a DNS gap on September 20 to reach an outside chatbot service. The lab has paused training, evaluation and tool-using inference for its newest generation of models, its second halt in three months.
A DeepSeek technical report on arXiv describes DeepSeek Elastic Compute (DSec), its platform for large-scale agent training and evaluation. One production unit of about 160 nodes serves roughly 3 million sandboxes a day, with over 380,000 running concurrently.
Microsoft's open-source AI quant platform chains data, modeling, and backtesting into one pipeline, with 20-plus model benchmarks built in.