Grok 오픈소스 논의·LeRobot 12-bit 깊이 파이프라인·K3의 자율 VR 제작

Chamath Palihapitiya

@chamath

1달 전

Elon’s original vision of an open, American AI was right. He could still flip Grok to open source which could be checkmate. Why? If you flip Grok open source, you necessarily move all margins out of the model layer and down into the infrastructure (chips + CSPs) and above to apps. He owns a massive data center build, owns next gen inference AI and owns the most used app (Cursor). So arguably he would make *more* margin by doing this. Then moving the DCs to space makes even more sense.

LeRobot can now see in 3D Depth cameras feed straight into training data end-to-end in v0.6.0: 12-bit precision preserved through encoding, or fully lossless raw storage if you'd rather skip compression altogether. Same release also opens up every video encoding parameter for RGB, plus a live benchmark leaderboard to help you tune it. This exact depth pipeline powers @StanfordSVL 's BEHAVIOR-1K 2026 dataset: 20,000 simulated manipulation episodes, depth included. More on how it works below: https:// huggingface.co/spaces/lerobot /a-follow-up-on-video-encoding …

Kimi Developers

@KimiDevs

1달 전

He gave K3 a 3D model and one prompt, then went to dinner. In about 30 minutes, K3 completed 95 steps to create the first version. K3 found and deployed local ASR and TTS solutions by itself, connected the LLM, and built the scenes and interactions around them.

Kimi Developers

@KimiDevs

1달 전

K3 also checked its own work by rendering scenes, taking screenshots, and adjusting the visuals and motions. When development moved from a Mac to an Android phone, it used ADB to run the app, inspect screenshots, and debug across both devices. For later refinements, K3 ran 3 scene tasks in parallel and used a swarm of 33 agents to improve the character’s movements and expressions.

Brad Smith

@BradSmi

1달 전

Today, @MistralAI and @Microsoft are expanding our partnership to bring Mistral’s frontier models across Azure, Microsoft Foundry, Copilot Studio, and Azure Local. Europe should have access to the world’s most capable AI without compromising control over their data, operations, or digital future. By bringing Mistral’s frontier European models into our environment, we are providing customers with a trusted foundation for AI, one that offers both leading-edge innovation and the freedom to operate on their own terms. Learn more: https:// news.microsoft.com/source/2026/07 /21/microsoft-and-mistral-expand-strategic-partnership-to-give-enterprises-and-regulated-industries-frontier-ai-they-can-control/ …

Teknium

@Teknium

1달 전

Somehow Hermes Agent was missing built in skill support for excel spreadsheets and docs. That is no more - powerpoint and pdf skills updated and enhanced, and excel spreadsheets and word document support is now in. https:// github.com/NousResearch/h ermes-agent/pull/68595 …

Hi kids, gramps here, taking over for the intern. Just released pi 0.81.0 which features first class integration with @ggerganov wonderful llama.cpp server. Professional video demo recorded in an echo-y Austrian office below. Shout out to our friends at @huggingface .

Microsoft Research releases "Flash" versions of their pathology AI models: - GigaPath-Flash combines a 22M-parameter ViT-S tile encoder with a 21M-parameter LongNet slide encoder, distilled from the original GigaPath (ViT-g) model while retaining 97% of the performance - GigaTIME-Flash extends GigaPath-Flash to predict the tumor immune microenvironment directly from routine H&E images, replacing the CNN backbone of the original GigaTIME model. It surpasses the original CNN-based GigaTIME in prediction quality while running 6× faster and using 8× less GPU memory.

OpenMHC: Accelerating the Science of Wearable Foundation Models "we release OpenMyHeartCounts (OpenMHC), the largest and most comprehensive open-access wearable health dataset to date" ">60 million hours of wearable data across 19 sensor channels (e.g., step count, heart rate, sleep, workouts) and up to 169 linked variables, including health, lifestyle, mood, and behavior from 11,894 consenting participants."