Mistral AI
·
4 months ago
Mistral released Small 4, a single model that combines text reasoning, image processing, and coding capabilities previously split across separate specialized models. The model has 119 billion total parameters with 6 billion active per token, a 256,000 token context window, and achieves 40% lower latency and 3 times higher throughput than its predecessor. Users can now deploy one versatile model instead of switching between separate systems for chat, reasoning, and coding tasks.
Mistral AI
·
4 months ago
● 2 sources
Mistral AI joined NVIDIA's Nemotron Coalition as a founding member to co-develop open-source frontier AI models, combining Mistral's model architecture with NVIDIA's compute resources and tools. The coalition's first initiative is a base model trained on NVIDIA DGX Cloud that will underpin the NVIDIA Nemotron 4 family, which will be open-sourced for post-training and specialization. This partnership enables developers and organizations to customize models for specific domains and use cases while maintaining local control and compliance.
Google Research
·
4 months ago
Google researchers evaluated six large language models on expert-level questions about high-temperature superconductivity, a specialized physics domain with competing theories. NotebookLM and a custom retrieval-augmented generation system that drew from curated databases of 1,726 scientific papers outperformed web-based models, with the study assessing answers across six metrics including accuracy, balance, and evidence support. The results indicate that LLMs restricted to quality-controlled source materials can reach proficiency in complex specialized fields, while those relying on unfiltered web data tend to conflate established theories with speculative ones.
Mistral AI
·
4 months ago
Mistral released Leanstral, an open-source AI agent designed to generate and formally verify code in the Lean 4 proof assistant, addressing the bottleneck of human review in high-stakes software development. Leanstral achieves a score of 26.3 on the new FLTEval benchmark with two inference passes at a cost of $36, compared to Claude Sonnet's score of 23.7 at $549. The model is available free through Mistral Vibe, a free API endpoint, and under Apache 2.0 license for local use, enabling developers to use formally verified code generation without manual proof checking.
Import AI
·
4 months ago
Researchers benchmarked how well large language model agents can autonomously fine-tune other LLMs for new tasks, finding they achieve meaningful but subhuman performance on post-training optimization. Claude Opus 4.6 scored 23.2% on PostTrainBench tasks using a single H100 GPU over 10 hours, compared to 51.1% for human teams, with performance improving from 9.9% six months prior. As AI agents improve at post-training automation, the capability to independently identify models and optimize them for specific objectives will likely accelerate AI development cycles.
Last Week in AI
·
4 months ago
● 2 sources
This podcast episode covers weekly AI news from March 2026, including Perplexity's local Mac-based AI agent, Anthropic's code review tool at $15-$25 pricing, and Nvidia's release of its 120B-parameter Nemotron 3 Super open-weights model. Anthropic sued the Pentagon over a supply chain risk designation requiring removal within 180 days, while xAI experienced major co-founder departures and Yann LeCun's aMI raised $1.3 billion. The episode discusses multiple research papers on activation steering resistance, chain-of-thought control limitations, inference scaling in cyber tasks, and improvements in RNN memory caching and CUDA kernel optimization.
Last Week in AI
·
4 months ago
● 2 sources
Anthropic filed lawsuits against the Trump administration over a Pentagon dispute involving the removal of its AI technology from military systems. The Pentagon issued an internal memo directing military commanders to remove Anthropic's technology from key systems. The legal action challenges the administration's decision to exclude the company from military AI deployments.
Together AI
·
4 months ago
● 2 sources
Together AI will showcase its AI infrastructure and model capabilities at NVIDIA GTC 2026 from March 16-19, featuring integrations with NVIDIA tools like Dynamo 1.0 and new voice AI models. The company is announcing support for NVIDIA Nemotron 3 Super, a 120B-parameter mixture-of-experts model with a 1M-token context window, alongside access to NVIDIA Parakeet TDT 0.6B V3 for speech recognition. Together AI will host multiple technical sessions and booth demonstrations highlighting production inference optimization and agentic AI systems.
OpenAI Blog
·
4 months ago
Codex Security uses AI-driven constraint reasoning and validation instead of traditional static application security testing to identify vulnerabilities. The approach reduces false positives compared to conventional SAST tools by focusing on real exploitable issues rather than generating reports on potential code patterns. This shifts security analysis from flagging suspicious code patterns toward confirming actual vulnerability conditions.
