• The tiny French AI company, Mistral, continues to amaze – they’ve released Mixtral 8x7B, a high-quality sparse mixture of experts model (SMoE) with open weights which outperforms GPT-3.5, as well as Llama 2 70B on most benchmarks with 6x faster inference, without the guardrails that those models have. - Mistral
• The EU has finally agreed on the AI act, passing “the first rules for AI in the world” – it bans social scoring systems, restricts facial recognition technology, forces foundational model providers to summarize their training data, and imposes reporting requirements for “high-risk” AI. - EU Council
• Robots going for a walk – new approach uses a Transformer model (same architecture that powers LLMs) to enable robots to walk around new environments, overcome obstacles, and adapt to context without being trained for any specific route or environment. - GitHub
• Stability keeps pushing the envelope in terms of tiny, but useful models with StableLM Zephyr 3B, a 3 billion parameter LLM providing accurate and responsive output on a variety of devices without requiring high-end hardware. - Stability
• Traditionally, materials are discovered and then you find a use for them, MatterGen turns it around – it’s a model that creates materials with specific, desired magnetic, electronic, and mechanical properties, and generates materials given target chemical systems. - Microsoft
• AI has killed lock-in by file format – “suddenly switching costs aren’t just lower than before. They’ve disappeared entirely,” so software companies should forget about this tactic and focus on building human-focused software that users love. - Fillout
• The Purple Llama project is Meta’s collaborative effort to create an open ecosystem for responsible genAI development – it provides cybersecurity safety evaluations for LLMs with metrics, tools, and ways to reduce AI-generated code vulnerabilities, with the Llama Guard model to take care of potentially risky outputs. - Meta
• Google embarrassed themselves in front of the industry with a faked promotional video for the Gemini model – and Greg Technology shows that GPT-4 is perfectly capable of doing what Google had to pretend that Gemini can do. - Greg Technology, Ars Technica
