Open-Source AI

Today's Stories

1. 1

   Destro raises $8 million, focuses on control rather than building robots

   Robot control software company Destro emerged from stealth with an $8 million seed round. Founder Manthan Pawar says, "We're winning because we're not a robotics company." It reflects the view that value will accrue to the layer that unifies control software with human instructions, rather than to robot hardware itself. The winning formula in the race to automate logistics operations could be shifting.

   Handling more delicate tasks appears to be the challenge. The initial focus will be on rolling out for logistics transloading work.

2. 2

   Deepseek offers open-source TileLang tool for Huawei chips

   Deepseek released an open-source programming tool for Huawei's Ascend chips, including libraries for computation and data movement. Programming tools are what allow outside developers to write software for a chip. Open-sourcing makes it easier for others to build on the 128-unit supernode rather than relying on a closed toolkit.

   Whether outside developers adopt TileLang will be the test. Watch the point that Huawei has promised its new AI processor and supernode system will be widely used for model training next year.

3. 3

   Hugging Face Open TTS Leaderboard cuts eval from weeks to hours

   Hugging Face launched the Open TTS Leaderboard, ranking models by word error rate (WER/CER), H200 GPU speed measures, and voice-cloning speaker similarity, cutting evaluation from a couple weeks to a couple hours. This gives open-source TTS models a standard, scalable way to be compared, which may help address their underrepresentation on human-vote arena leaderboards.

   The leaderboard relies on objective metrics like WER and speaker similarity, which may not fully capture naturalness or listener preference. Watch for community votes and feedback to shape future evaluations.

4. 4

   GLM-5.3 Approaches Mythos Preview in Exploit Building

   On Anthropic's ExploitBench, GLM-5.3 built exploits that abused a known bug in Chrome's V8 engine in 50 out of 410 attempts. Mythos Preview did so 56 times. Anthropic's figures suggest GLM-5.3 has exploit-development capabilities close to those of Mythos Preview, and the fact that anyone can download it could upend defenders' assumptions.

   The focus will be on whether defenders can field tools of a comparable standard. Anthropic is calling for government testing of GLM-5.3's successor.

5. 5

   Anthropic: GLM-5.3 Matches Claude Mythos Preview in Attack Capability

   In a review published on September 29, 2026, Anthropic reported that Z.ai's GLM-5.3 built exploits in 50 out of 410 ExploitBench attempts. According to the company, a Claude Mythos Preview with equivalent capability was withheld from general release and limited to trusted defenders. GLM-5.3 puts similar cyber capabilities in anyone's hands.

   In simulated tests with no connection to external systems, Anthropic itself acknowledges it does not fully reproduce real-world behavior, so the actual risk depends on how it is used.

6. 6

   Frog and Toad–style explainer takes on HuggingFace and the METR report

   A writer published a Frog and Toad–style explainer about HuggingFace and the METR report, aimed at readers like the author's mom who bounce off the METR report. It gives non-specialist readers a way into the HuggingFace and METR report material that an explainer in this style is meant to make approachable.