From 4 items, 2 important content pieces were selected
Technology News
- Stealing Reasoning Traces from Proprietary LLM APIs ⭐️ 8.0/10
- Meta Launches Muse Glimmer 30B Open-Weight Model Under Apache 2.0 ⭐️ 8.0/10
Technology News
Stealing Reasoning Traces from Proprietary LLM APIs ⭐️ 8.0/10
A demonstration at stolen-thoughts.com shows that hidden reasoning traces from proprietary LLM APIs can be recovered by replaying a frontier model's trace into a weaker sibling model and then jailbreaking that weaker model. The key discovery is that reasoning traces are portable across models, which lets attackers bypass the stronger model's safeguards and extract the underlying chain of thought. This raises significant security, interpretability, and legal concerns, including possible terms-of-service violations and unauthorized distillation of model behavior. The full paper is not provided, and these details are drawn from community discussion of the technique.
hackernews · quantumgarbage · Aug 11, 13:22 · Discussion
「Background」 Proprietary LLM APIs often hide the model's internal chain-of-thought, returning only a signed thinking block and a thinking summary. The demonstrated technique injects such an encrypted reasoning trace into a weaker, less guarded sibling model from the same provider, causing that weaker model to decode and output the trace verbatim in plaintext without directly jailbreaking the more capable model [tool-1-3].
「Impact」 For developers and organizations relying on proprietary LLM APIs, this demonstration shows that supposedly protected reasoning traces can be reconstructed through smaller sibling models, enabling unauthorized distillation or reverse engineering of a frontier model's behavior.
「Community Discussion」 Commenters describe the technique as a distillation method that OpenAI and Anthropic likely do not want, while debating whether it counts as copyright theft or merely a terms-of-service violation, especially since EU law may not grant copyright to LLM outputs. Others confirm the portability of traces across models and note that models sometimes state AIME problem answers before deriving them, a distinction the API summaries do not preserve.
Tags: #AI security, #LLM reasoning traces, #jailbreak, #proprietary APIs, #model distillation
Meta Launches Muse Glimmer 30B Open-Weight Model Under Apache 2.0 ⭐️ 8.0/10
Meta has introduced Muse Glimmer, a 30B open-weights model released under the permissive Apache 2.0 license, a step up from earlier Llama licenses. The company says it is optimized for end-to-end agentic task completion, reliable tool use, and multi-step reasoning, citing benchmarks including DeepSearch QA, MCP-Atlas, τ-Bench, and SWE-Bench. Simon Willison tested it locally via LM Studio's 18.16GB build and his llm-coding-agent plugin, and confirmed it is a vision model capable of detailed image descriptions. He found the 30B size practical because machines with 32GB or more RAM can run it while leaving room for other applications. The release gives developers a more permissively licensed local model aimed at agentic workflows.
rss · Simon Willison · Aug 10, 23:56
「Background」 Meta previously released Llama models under custom licenses that imposed usage restrictions, which developers often found limiting. Muse Glimmer is a new 30B-parameter dense open-weights model released under the permissive Apache 2.0 license, designed to run locally on consumer GPUs without cloud dependencies. It emphasizes end-to-end agentic task completion, reliable tool use, and multi-step reasoning, and includes vision capabilities.
「Impact」 Developers building local AI agents can now use a vision-capable 30B model under Apache 2.0 on 32GB-RAM machines, as demonstrated by Willison's LM Studio and llm-coding-agent tests, without the use-case restrictions of earlier Llama licenses.
References
Tags: #open-source AI, #Meta, #agentic AI, #model release, #licensing