The archive
New models, what they can do, and how to build with them.
Google DeepMind's new world model generates interactive environments with ambient sound and dialogue, and uses Gemini models to plan multi-step events. It's a trusted-tester preview with no benchmarks attached.
5 min read
OpenAI says its newest models post large gains on two new evaluations covering COBOL, Fortran, OCaml and Visual Basic, including 97.9% on a COBOL migration test. The eval design matters as much as the numbers.
5 min read
OpenAI's new GPT-6 developer guide lays out three tiers—Astra, Sol and Luna—with a 100x price spread, plus steering, async tools and delegation for tasks that run for hours.
6 min read
A CSAIL-led tool pairs Microsoft's TRELLIS 3D generator with GPT-4 so novices can spot and fix structural flaws before printing. Nearly 80% of generated test models were flawed.
5 min read
Google DeepMind's new frontier model claims a 1M-token output limit and state-of-the-art scores on software engineering and security benchmarks, but access starts with a small group of vetted defenders.
5 min read
AWS says its internal NarrateAI assistant serves 4,000+ executives using adaptive routing, cross-account quota failover and paragraph-level evaluation that runs while the model streams.
5 min read
Google DeepMind says Gemini 3.8 Live with Live Avatar pairs low-latency video generation with speech, adding lip-synced visual personas across 97 languages for Gemini Enterprise customers.
5 min read
A clear walk through how large language models are built and trained, from next-token prediction to instruction tuning, and why they seem to reason but still fail.
9 min read
Google DeepMind's two new text-to-speech models add prompt-built custom voices, 30-second voice replication and line-by-line performance direction — with pricing and latency undisclosed.
5 min read
OpenAI says its two new GPT-6 tiers halve API prices versus GPT-5.6 promotional rates and improve prompt caching. The benchmark claims come with fine print worth reading.
6 min read
Anthropic says Opus 5.5 costs 40% less to run than Opus 5 and tops its internal alignment audit — while warning that benchmark margins no longer track real-world differences.
5 min read
How retrieval-augmented generation connects a language model to your own documents, covering chunking, retrieval, reranking, grounding, common failures, and evaluation.
8 min read