Dec 9, 2025

ML Times - Dec 9, 2025

Daily

Weekly

• Donating the Model Context Protocol and Establishing the Agentic AI Foundation
  Anthropoc has donated the Model Context Protocol (MCP) to the Agentic AI Foundation, a collaborative initiative under the Linux Foundation, aimed at fostering open standards in agentic AI development, supported by major tech companies like Google and Microsoft.

• Mistral Releases Devstral 2 (72.2% SWE-Bench Verified)
  Devstral 2 is a 123B-parameter coding model achieving 72.2% on SWE-bench Verified, making it a leading open-source alternative with significant cost efficiency compared to competitors like Claude Sonnet.

• Jepsen: NATS 2.12.1
  NATS JetStream 2.12.1 exhibits significant vulnerabilities, including data loss due to file corruption and power failures, with tests revealing that acknowledged messages can be lost even when a majority of nodes remain operational.

• IBM to acquire Confluent
  IBM has agreed to acquire Confluent for $31.00 per share, aiming to create a unified platform that enhances data accessibility for enterprises, thereby accelerating AI integration and operational modernization.

• The Universal Weight Subspace Hypothesis
  Deep neural networks trained on diverse tasks converge to shared low-dimensional parametric subspaces, revealing a systematic structure in their weight distributions across various architectures and tasks, as evidenced by analysis of over 1100 models. Link to article

• Has the cost of building software dropped 90%?
  The cost of building software may have dropped by 90% due to advancements in agentic coding, which significantly reduces labor and coordination overhead, allowing projects that once took a month to be completed in a week.

• Deep dive on Nvidia circular funding
  NVIDIA's Q3 Fiscal 2026 earnings reveal a $31.9 billion net income, yet a concerning $8 billion cash flow gap suggests potential liquidity issues, raising questions about the sustainability of their "virtuous cycle" in AI.

• We collected 10k hours of neuro-language data in our basement
  Conduit collected ~10,000 hours of neuro-language data, the largest dataset of its kind, to train models that decode semantic content from noninvasive neural data, enabling zero-shot predictions from participants' thoughts before they articulate them.

• Transformers know more than they can tell: Learning the Collatz sequence
  Transformers achieve up to 99.7% accuracy in predicting long Collatz steps when using bases 24 and 32, while performance drops significantly to 37% and 25% for bases 11 and 3, respectively, indicating a strong dependency on input encoding.

• AWS Trainium3 Deep Dive – A Potential Challenger Approaching
  AWS's Trainium3 introduces significant advancements in AI accelerator technology, focusing on optimizing performance per total cost of ownership (TCO) while enhancing operational flexibility through a unique switched scale-up architecture that supports frontier Mixture-of-Experts (MoE) models.

• Launch HN: Mentat (YC F24) – Controlling LLMs with Runtime Intervention
  Mentat offers developers deterministic control over LLM behavior, enabling real-time adjustments to reasoning and bias without the need for fine-tuning or prompt engineering, as demonstrated in their API documentation.

• Agentic QA – Open-source middleware to fuzz-test agents for loops
  Agentic QA is an open-source middleware designed to fuzz-test agents for infinite loops and PII leaks, preventing costly errors like the ~$50 loss from a LangChain agent's malfunction.

• Group Representational Position Encoding
  GRAPE (Group RepresentAtional Position Encoding) unifies multiplicative rotations and additive logit biases, enhancing positional encoding through group actions, with Multiplicative GRAPE yielding a norm-preserving map and Additive GRAPE recovering models like ALiBi and FoX. Link to article

• Launch HN: Mentat (YC F24) – Controlling LLMs with Runtime Intervention
  Open WebUI enables seamless integration of web applications through SvelteKit, enhancing user experience with dynamic content loading and efficient rendering.

• [P] Open-source forward-deployed research agent for discovering AI failures in production
  Agent Tinman is an open-source research agent that operates alongside AI systems to continuously identify potential failures in real-time environments, enhancing model reliability through structured experimentation and intervention proposals.