Mistral Releases Formal Verification Model "Leanstral 1.5" Under Apache 2.0
On July 3, Mistral AI released "Leanstral 1.5-119B-A6B", a model specialized in formal verification. It is available under the Apache 2.0 license, so anyone can use it freely.
The model uses a MoE architecture with 119B parameters, of which 6B are active. In benchmarks, it saturated miniF2F and solved 587 of the 672 problems in PutnamBench, achieving strong results of 87% on FATE-H and 34% on FATE-X. It was trained with a pipeline combining mid-training, SFT, and reinforcement learning (CISPO).
Particularly noteworthy: testing across 57 repositories uncovered 5 previously unknown bugs, demonstrating the model's potential for practical use in automated theorem proving and formal proof engineering.
OpenAI's GPT-5.6 Family, New Robot Training Methods, and Models Invoking Models
DeepLearning.AI's The Batch (July 3 issue) published a weekly digest that includes coverage of OpenAI's GPT-5.6 family. It covers multiple topics, including new approaches to robot learning and the concept of "models invoking models," where AI models call other models. The details are available as subscriber-only content.
Meta's Zuckerberg: "AI Agent Technology Is Progressing More Slowly Than Expected"
In an interview, Meta's Mark Zuckerberg said that progress on AI agent technology is lagging behind initial expectations. He acknowledged that many challenges remain to be solved before agentic AI becomes practical, including reasoning reliability and long-horizon task execution. While Meta intends to keep investing in the agent space, the message that expectations need recalibrating is likely to influence the industry-wide outlook on AI agents.
"First" End-to-End AI Agent Ransomware Attack Confirmed
According to The Register, the "first" end-to-end ransomware attack powered by AI has been confirmed. By driving an AI agent, the attacker reportedly automated everything from target selection to intrusion, encryption, and the ransom demand. This technique, dubbed the "Smooth AI criminal," shows that AI technology is starting to be used to automate cybercrime as well, and security teams need to recognize it as a new threat model.
AMD Releases Open-Source Interactive World Model "Micro-World"
AMD has released "Micro-World", an action-controllable interactive world model. Built on the Wan2.1 family, it supports both image-to-world (I2W) and text-to-world (T2W). The model weights, training code, inference code, and a curated dataset have all been made publicly available.
For action injection, it adopts two methods: lightweight adaLN and the highly stable ControlNet. It delivers high-quality open-domain scene generation and aims to accelerate community research and real-world adoption.
The Toll That Dependency on AI Coding Takes on Engineers
A LeadDev article examined how addiction to AI coding tools affects engineers' health and careers. While these tools boost productivity, the article points to problems such as shallow code understanding, review fatigue, and over-reliance on AI output. A discussion has also started on Hacker News, gathering voices from practicing engineers. Establishing healthy ground rules around AI coding has become an urgent challenge for teams.
---