https://yurkevi.ch/blog/ Short notes on AI models, tools and research worth knowing about. https://yurkevi.ch/blog/#2026-08-10-tencent-s-hunyuan3d-buffalo-1-0-does-text-to-3d-editing-and https://yurkevi.ch/blog/#2026-08-10-tencent-s-hunyuan3d-buffalo-1-0-does-text-to-3d-editing-and Mon, 10 Aug 2026 08:53:01 +0200 Tencent published Hunyuan3D-Buffalo 1.0 on August 3: text-to-3D, part-level editing and 3D understanding in one model, trained on 87M samples. Weights are not out. https://yurkevi.ch/blog/#2026-08-04-nyra-labs-releases-crisperwhisper-2-0-a-2b-verbatim-speech https://yurkevi.ch/blog/#2026-08-04-nyra-labs-releases-crisperwhisper-2-0-a-2b-verbatim-speech Tue, 04 Aug 2026 09:21:52 +0200 CrisperWhisper 2.0 scores 87.8 disfluency F1 across 10 languages, ahead of ElevenLabs Scribe v2 and Microsoft, and the weights download and run offline for free. https://yurkevi.ch/blog/#2026-08-04-eyeline-labs-and-netflix-release-id-v2v-for-identity https://yurkevi.ch/blog/#2026-08-04-eyeline-labs-and-netflix-release-id-v2v-for-identity Tue, 04 Aug 2026 09:16:42 +0200 ID-V2V restyles a video's background and lighting from a single edited keyframe while holding faces, expressions and lip sync fixed. Weights and ComfyUI support are out. https://yurkevi.ch/blog/#2026-08-01-openai-s-unreleased-astra-model-produces-ten-results-on https://yurkevi.ch/blog/#2026-08-01-openai-s-unreleased-astra-model-produces-ten-results-on Sat, 01 Aug 2026 18:45:31 +0200 OpenAI published ten proofs on long-open math and TCS problems today, produced by its unreleased Astra model, each shipped with a Lean 4 formalization. https://yurkevi.ch/blog/#2026-07-28-hkust-s-homie-puts-specific-people-and-objects-into https://yurkevi.ch/blog/#2026-07-28-hkust-s-homie-puts-specific-people-and-objects-into Tue, 28 Jul 2026 09:01:09 +0200 HOMIE personalizes video with reference humans, products and logos, builds on the open Wan2.1-14B and Wan2.2-14B models, and ships code and weights. https://yurkevi.ch/blog/#2026-07-28-microsoft-releases-mage-flow-a-4b-image-generation-and https://yurkevi.ch/blog/#2026-07-28-microsoft-releases-mage-flow-a-4b-image-generation-and Tue, 28 Jul 2026 08:55:49 +0200 Mage-Flow generates and edits images at native resolution, handles depth and segmentation, scores 0.90 on GenEval, and ships open under MIT. https://yurkevi.ch/blog/#2026-07-24-anthropic-releases-claude-opus-5-beating-fable-5-on-most https://yurkevi.ch/blog/#2026-07-24-anthropic-releases-claude-opus-5-beating-fable-5-on-most Fri, 24 Jul 2026 22:41:11 +0200 Claude Opus 5 shipped July 24, 2026. It outscores Fable 5 on most benchmarks at half the input price, though Fable stays ahead on SWE-bench Pro. https://yurkevi.ch/blog/#2026-07-20-nvidia-releases-ardy-an-open-real-time-text-to-motion-model https://yurkevi.ch/blog/#2026-07-20-nvidia-releases-ardy-an-open-real-time-text-to-motion-model Mon, 20 Jul 2026 09:24:49 +0200 ARDY generates 3D human and humanoid motion from text in real time; code is Apache 2.0 and weights are open, which could help generate synthetic motion data. https://yurkevi.ch/blog/#2026-07-20-claude-fable-produces-a-counterexample-to-the-jacobian https://yurkevi.ch/blog/#2026-07-20-claude-fable-produces-a-counterexample-to-the-jacobian Mon, 20 Jul 2026 09:05:53 +0200 On July 19, 2026, Levent Alpöge announced a Claude Fable counterexample to the Jacobian conjecture: a map with constant Jacobian determinant -2 and no inverse. https://yurkevi.ch/blog/#2026-07-18-demis-hassabis-proposes-a-finra-style-body-to-regulate https://yurkevi.ch/blog/#2026-07-18-demis-hassabis-proposes-a-finra-style-body-to-regulate Sat, 18 Jul 2026 13:12:49 +0200 In a July 14 manifesto Google DeepMind CEO Demis Hassabis called for a US-led, industry-funded AI standards body that could vet frontier models before release and even pause the industry. https://yurkevi.ch/blog/#2026-07-18-moonshot-ai-releases-kimi-k3-a-2-8-trillion-parameter-open https://yurkevi.ch/blog/#2026-07-18-moonshot-ai-releases-kimi-k3-a-2-8-trillion-parameter-open Sat, 18 Jul 2026 12:54:39 +0200 Kimi K3 is the largest open-weight model so far, scores near Fable 5 on the Intelligence Index, and tops the blind-voted Frontend Code arena.