Cognition launches SWE-2 coding model, rivaling Fable 5.1 and GPT-Astra

A developer is training a 4B model for general coding plus 3D scene generation, sharing an early preview video of results. They say small models have performed poorly at 3D in their own tests until this progress.

Two 600M-parameter TTS models fine-tuned from Qwen3-TTS using RLVR, OPD, and multi-teacher OPD, trained for under $200. They support inline tags, direction control, and voice cloning.

Researchers analyzed donated chat logs from 80 ChatGPT users and found 2,050 memory entries across 65 users; 96% were created unilaterally by the chatbot. 52% contained psychological insights and 28% held GDPR-defined personal data.

Arize AI's Aparna Dhinakaran argues the role has shifted from writing features to running "factories of agents," where the core job is building the loop: evals that review non-deterministic work, telemetry agents consume, and agents that act on it.