Blog
Short notes on AI models, tools and research worth knowing about. Click a title to expand. Subscribe via RSS.
nyra labs releases CrisperWhisper 2.0, a 2B verbatim speech recognition model 4 August 2026
nyra labs released CrisperWhisper 2.0 in late July and posted the paper on July 21. It is a 2B-parameter verbatim ASR model that keeps fillers, repetitions and false starts instead of cleaning them up. On their 10-language disfluency benchmark it scores 87.8 F1, ahead of ElevenLabs Scribe v2 at 79.2 and Microsoft MAI-Transcribe-1.5 at 77.5. Mean word-boundary error is 29.6 ms on TIMIT.
The weights are on Hugging Face, free to download and run offline. The stronger Pro variant scores 93.5 F1 but is not part of the free download.
Eyeline Labs and Netflix release ID-V2V for identity-preserving video restyling 4 August 2026
Eyeline Labs and Netflix released ID-V2V in July. It takes a source video and one restyled keyframe, then propagates the new background, lighting and style across the clip while keeping faces, expressions, gaze and lip sync fixed. The paper reports 0.701 AdaFace identity similarity against 0.402 to 0.529 for the baselines.
It builds on VACE and Wan 2.1, and the weights are on Hugging Face. Kijai added ComfyUI support on July 29.
OpenAI's unreleased Astra model produces ten results on open math problems 1 August 2026
OpenAI published ten results on open problems in mathematics and theoretical computer science today. An unreleased internal model called Astra produced them, and each problem had seen no progress on the main result for at least a decade. The list includes the first explicit non-sofic group, a disproof of Connes's rigidity conjecture, and an improved asymptotic upper bound for high-dimensional sphere packing. Every result ships with a Lean 4 formalization and a walkthrough of the model's reasoning. OpenAI puts the total token cost at about $2,000 at Sol API rates. No outside mathematician has worked through the arguments yet.
HKUST's HOMIE puts specific people and objects into generated video 28 July 2026
HKUST researchers published HOMIE last week, a method for putting specific people and objects into generated video while keeping their identity. It handles products, logos and multi-view references, and renders text through OCR maps at up to 1280x720. It builds on the open Wan2.1-14B and Wan2.2-14B video models and uses Qwen3-VL-2B-Thinking to read the references, per the paper. Code and weights are on GitHub.
Microsoft releases Mage-Flow, a 4B image generation and editing model 28 July 2026
Microsoft released Mage-Flow last week. It is a 4-billion-parameter native-resolution model for image generation and editing, and it also handles depth, segmentation, Canny, and sketch conditioning. It scores 0.90 on GenEval, first among open models, ahead of FLUX.2 and Qwen-Image at 0.87. The Turbo variant makes a 1024x1024 image in 0.59 seconds on one A100. The weights are on Hugging Face under an MIT license, with code on GitHub.
Anthropic releases Claude Opus 5, beating Fable 5 on most benchmarks at half the price 24 July 2026
Anthropic released Claude Opus 5 today. It beats Fable 5 on most benchmarks at half the input price, $5 per million tokens against Fable's $10. On Frontier-Bench v0.1 it scores 43.3% to Fable 5's 33.7%. On ARC-AGI-3 it hits 30.2%, three times the next-best model. Fable stays ahead on a few, including SWE-bench Pro, where Opus 5 trails at 79.2% versus 80%. Decrypt reports it becomes the default model for most Claude subscribers.
Nvidia releases ARDY, an open real-time text-to-motion model 20 July 2026
Nvidia released ARDY on July 10. It is an autoregressive diffusion model that generates 3D human and humanoid motion from text prompts and kinematic constraints in real time, streaming frames as it computes them. The code is Apache 2.0 and the weights are open under the NVIDIA Open Model License, which permits commercial use. The work is a SIGGRAPH 2026 paper. Open code plus open weights make it a candidate for generating synthetic motion data.
Claude Fable produces a counterexample to the Jacobian conjecture 20 July 2026
On July 19, 2026, Harvard mathematician Levent Alpöge announced that the Jacobian conjecture is false, crediting Claude Fable with the counterexample. It is a polynomial map ℂ³ → ℂ³ with constant Jacobian determinant -2 that sends three distinct points to (-1/4, 0, 0), so it has no inverse. The conjecture had been open since Keller posed it in 1939. The announcement was a one-line tweet posted during the World Cup final. The map is elementary to check by hand.
Demis Hassabis proposes a FINRA-style body to regulate frontier AI 18 July 2026
Demis Hassabis published a manifesto on July 14 calling for an independent AI standards body modeled on FINRA. Hassabis runs Google DeepMind and shared the 2024 Nobel Prize in Chemistry for AlphaFold. Frontier labs would share models with the body up to 30 days before release, tested for cyber, bio, and deception risks. He told Axios he wants it US-led, industry-funded, and running before year-end. The body could later require models to pass before US deployment and even coordinate a slowdown across the labs. Hassabis says AGI is probably only a few years away. Nadella and Altman back the idea.
Moonshot AI releases Kimi K3, a 2.8-trillion-parameter open-weight model 18 July 2026
Moonshot AI released Kimi K3 on July 16. It has 2.8 trillion parameters and is the largest open-weight model so far. On the blind-voted Frontend Code arena it is first at 1,679, ahead of Claude Fable 5 and GPT-5.6 Sol. On the broader Artificial Analysis Intelligence Index it scores 57, just behind Fable 5 at 60 and Sol at 59. It is on kimi.com, subscription and API at $3 per million input tokens and $15 per million output. The weights open on July 27.