GPT-4.5 Rumour, Apple's New On-Device LLMs, New 3D, Visual, and Autonomous LLMs

Before we start, share this week's news with a friend or a colleague:

Key Takeaways

This week’s key developments include:

• Apple introduced OpenELM, a family of efficient, open-source LLMs designed to run on-device without needing cloud servers.

• Three sets of LLMs were released on Hugging Face by Alibaba, Internist.ai, and BigCoder, all of which showed impressive performance at various benchmarks.

• Anthropic developed a new probing technique using defection detectors, which could boost safety and reliability in AI applications.

• Helm.ai’s new foundation model deals with core issues in autonomous vehicle development.

• A new drug discovery company called Xaira launched with $1 billion and Run.ai is set to be acquired by Nvidia.

Got forwarded this newsletter? Subscribe below 👇

Rumour of the Week: Probably GPT-4.5 in LMSYS?

LMSYS chatbot arena, a tool for real-time comparison of LLMs, has a new mysterious gpt2-chatbot, rumored to be OpenAI's stealth test for GPT 4.5. Twitter rumours suggest it excels in reasoning, math, coding, and ASCII art, surpassing previous models (but is limited to 8 messages per day so there's only so much testing you can do). There's also speculation that OpenAI is privately benchmarking this model before a public release. Sam Altman fueled further speculation:

Check out this mega-thread by Andrew Gao on what's known about the model.