• AI Time to Impact
  • Posts
  • . . AI: Karpathy Rebuilds GPT2 in C, Apple's New UI Model, and Cohere Tops Open Source Leaderboard (4.9.24)

. . AI: Karpathy Rebuilds GPT2 in C, Apple's New UI Model, and Cohere Tops Open Source Leaderboard (4.9.24)

Apple, Anthropic, Cohere, Google AI

Author

Marshall Kirkpatrick
April 09, 2024

Friends, in today's AI news, Andrej Karpathy casually drops GPT-2 in C, Apple also casually releases their research on their model's ability to navigate UI screens, and we see a new open source leader climb the LLM ladder! We've put together some other headlines we think are also worth checking out, especially the one from Anthropic!

Here are the top stories the AI community is talking about.

-Marshall Kirkpatrick, Editor

Andrej Karpathy Develops Method to Train LLMs in Pure C

First impacted: Software Developers, AI Researchers
Time to impact: Short

Andrej Karpathy has developed a method to train large language models (LLMs) in pure C with only 1000 lines of code, eliminating the need for 245MB of PyTorch or 107MB of cPython. Karpathy, who used GPT-2 as his initial example, says he is working on a direct CUDA implementation and improving the CPU version with SIMD instructions, AVX2 on x86 / NEON on ARM, and other modern architectures. He mentions that he "chose GPT-2 to start because it is the grand-daddy of LLMs, the first time the LLM stack was put together in a recognizably modern form, and with model weights available." [GitHub - karpathy/llm.c: LLM training in simple, raw C/CUDA] Explore more of our coverage of: Andrej Karpathy, Large Language Models, C Programming. Share this story by email

First impacted: Mobile application developers, UI/UX designers
Time to impact: Medium

Apple has launched Ferret-UI, a large language model designed for improved understanding of mobile user interface (UI) screens. According to Apple, Ferret-UI outperforms most open-source UI LLMs and even GPT-4V in basic UI tasks, using an "any resolution" feature for detail enhancement and training on tasks such as icon recognition and widget listing, with a dataset for complex tasks to improve reasoning abilities. Ferret-UI employs a unique approach to understanding mobile UI screens by dividing each screen into two sub-images based on the original aspect ratio, which are then separately encoded before being sent to the large language models. [Paper page - Ferret-UI: Grounded Mobile UI Understanding with Multimodal LLMs] Explore more of our coverage of: Apple AI, Language Models, Mobile UI. Share this story by email

First impacted: Business Process Managers, IT Professionals
Time to impact: Short

Command R+ has achieved a significant milestone by climbing to the 6th spot in the latest Arena results, matching the performance level of GPT-4 (version 0314), as confirmed by over 13,000 human votes, making it the best open model on the leaderboard. [Introducing Command R+: A Scalable LLM Built for Business] Explore more of our coverage of: Large Language Models, Enterprise AI, Microsoft Azure. Share this story by email

Explore more of our coverage of: AI Language Models, Misinformation Spread, Ethical Concerns. Share this story by email

Explore more of our coverage of: FAIR Research, Language Models, Knowledge Storage. Share this story by email

Explore more of our coverage of: Google AI, Text Embedding, Audio Understanding. Share this story by email

That’s it! More AI news tomorrow!