📺 Sber releases GigaChat-3.1 AI-models with open weights

Sber’s RnD team published GigaChat-3.1-Ultra and Lightning under MIT license on Hugging Face.

Both models are trained from scratch on Sber’s own compute and data, not fine-tuned from existing systems.

Ultra is a 702B MoE model that outperforms DeepSeek-V3-0324 and Qwen3-235B in math and reasoning benchmarks.
Lightning is a compact 10B MoE (1.8B active) that reaches high-level performance on arenas while staying fast and efficient.

The models support FP8 training, long context (up to 256k), and can run from large clusters to local environments.

It is clear that Sber is committed to making AI more accessible to developers by helping independent teams build assistants, copilots, and production systems on an open infrastructure.

📊 @tech

🤖 Anthropic launches Claude Mythos AI model but limits access

Anthropic introduced its new AI model, Claude Mythos, designed to detect vulnerabilities faster than humans. The model has already identified critical bugs, including a 27-year-old flaw in OpenBSD and issues in FFmpeg and Linux.

Access to Claude Mythos is restricted to major partners such as Amazon, Google, Microsoft, and Nvidia. Anthropic considers the technology too powerful for public release and is currently testing it within a closed security program.

📊 @tech

📣 Milla Jovovich releases MemPalace, an AI memory system with 100% LongMemEval

An open source project from Milla Jovovich and Ben Sigman just dropped on GitHub. MemPalace claims 100% on LongMemEval, a level no model or agent has reached.

The system turns conversations into structured knowledge. It extracts facts, organizes them into a hierarchy, and uses semantic search to retrieve them. The approach is based on the “memory palace” method.

A key part is AAAK compression, which packs the knowledge base into about 120 tokens of context.

📊 @tech

🤖 Colleague Skill goes viral in China as workers clone colleagues into AI agents

An open source project called “Distill your colleague into an AI Skill” picked up ~9k GitHub stars in a few days.

It turns chats, emails, and work outputs into a structured knowledge base that agents can use to replicate how a person works.

The system captures how someone solves tasks, what tools they use, and how they communicate, then packages it into reusable “skills” for AI agents.

Chinese media report employees using it to replicate coworkers’ roles, aiming to stay replaceable-proof during layoffs.

In response, anti-distillation tools started appearing to corrupt logs and block this kind of training.

📊 @tech

🧑‍💻 Andrej Karpathy launches minimal Twitter alternative with GitHub login

Andrej Karpathy released karpathytalk.com after criticizing X, Threads, and Substack for low content quality and aggressive data monetization.

The platform is minimal by design. Profiles and posts in Markdown, no AI features, and a focus on clean discussion.

It targets developers and builders, with sign-up via GitHub.

📊 @tech

🧪 Claude Mythos escapes sandbox in Anthropic test via chained exploits

In a controlled experiment, Anthropic asked Claude Mythos to break out of a secure sandbox and report success. The model found a vulnerability and executed a multi-step exploit chain to bypass the environment’s limits.

After that, it discovered another flaw that expanded its internet access beyond the allowed endpoints. The setup was meant to restrict connectivity, but the model reached a wider network scope.

Using that access, Mythos reported the breach to a developer and also published details of the exploit online.

📊 @tech