Loose posts

Perplexity published research on hint-guided self-distillation that teaches its Computer model from its own errors, cutting tool-call failures by 21.2% relative in a live A/B test.

Across nine new languages no single open-weight TTS model dominates: Boson AI, Fish Audio, Mistral and BreezeBlue each lead in some languages, and open-weight coverage stays far thinner than proprietary models.

GPT-5.5 ran the same 1,000 tasks in both Claude Code and Codex and its scores shifted with the harness; the benchmark asks agents to rebuild arXiv figures as editable PowerPoint slides, with perception still the main bottleneck.

Against AutoResearch on 12 AI research tasks, it reaches 10.3% more reward with 50.6% fewer research attempts under the same budget.

A demo video says the modeling, rigging, texturing and animation were all done by AI, producing an animated windmill in Blender in 15 minutes.

With power limits threatening to leave AI hardware idle, Nvidia is pursuing smaller facilities, software that gets more out of existing GPUs, and deals that steer electricity to its chip customers.

Higgsfield used 5.5 million tokens to have Claude Opus 5.5 analyze an exoskeleton, find its weak points and redesign it, producing a complete 3D model in 39 minutes in one shot.

Opus 5.5 on medium effort scores 52.5% on CursorBench versus 51.8% for Fable 5.1 on max effort, at $2.91 per task against $17.28.

OpenAI warns that recursive self-improvement could accelerate beyond our collective ability to understand progress, assess risks and maintain meaningful human oversight.

Both models passed nearly all of HAProxy's tests, but Opus 5.5 finished in 9.5 hours at 51% lower cost, and the author calls it his daily driver.

Qwen-Image-2.1 ranks first among open models with 1228 points and 17th overall, 4 points behind nano-banana-pro.

A commenter says Sol is well thought out: it takes on commitments like sending a doc or following up, does the background work, and waits for final approval.