Independent coverage / AI
Independent coverage of AI: news, research, chips, and policy, plus AI video and image tools, how-to guides, reaction roundups, and opinion.
Chained voice agents decide from a transcript, not the audio. How chained pipelines, speech-to-speech models and full-duplex GPT-Live differ, what a transcript leaves out, why chains still give more control, and what GPT-Live costs at $0.05 per minute.
A new paper replaces Gemma 3 270M's 168M-parameter embedding table with frozen 512-bit codes. Trained on the same 2.05B tokens, it ties or beats the learned table on LAMBADA, HellaSwag and PIQA, and a ternary version is a 46 MB file running 615 to 740 tokens per second on a laptop CPU.
OpenAI plans to retire custom GPTs on Dec 11, 2026 and move them to plugins. Here is what transfers, what does not, the five migration steps, and a tested prompt for turning old GPT instructions into a plugin brief, as of October 2026.
Claude Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100k, a tenth of Haiku 4.5. Here is how it scores, what the new effort setting does, and where Anthropic says Sonnet 5.5 or Opus 5.5 is the better pick.
RTX Spark puts an NVIDIA Blackwell RTX GPU, a Grace CPU and up to 128GB of unified memory in one chip for Windows laptops. Here are the specs, the laptops, what NVIDIA says each memory size runs locally, Nemotron 3.5 Bolt, and retailer prices from $2,599.99.
Emirates NBD says it will deploy OpenAI's GPT models on in-country inference in the UAE, a first for a MENA bank. Here is what the bank plans, what OpenAI's inference residency keeps in the UAE, what may still run outside, and what the deal leaves out.
Three IDs, aurora-red-alpha, aurora-green-alpha and aurora-blue-alpha, appeared in OpenAI's API model list on October 17 and 18, then turned up in Arena image tests. OpenAI has not announced them. Here is who spotted what, the early tester impressions, and what remains unconfirmed.
A UIUC study on 300 SWE-rebench issues finds that at equal token budgets, showing a coding agent a few whole files beats prompt compressors and chunk retrievers: 43.7% resolved with GPT-5.6 and 46.7% with Claude Opus 5. Retrievers miss the file to edit; compressors alter the text patches must quote.
Nvidia's Nemotron 3.5 Bolt 74B-A9B packs a model compressed from Super VL into 24GB with 2-bit experts for RTX Spark and DGX Spark. Here is what it needs to run, how its speed compares in Nvidia's own tests, and when Bolt E4B arrives.
Greptile raised a $50M Series B led by Menlo Ventures at a $750M valuation. It says it reviews over 1% of GitHub PRs daily. How TREX runs your code, how Model Inversion picks the reviewer, and what Base, Plus and Apex reviews cost.
One email. The most important model releases, research findings, and industry moves, curated from LLMs Blog.
