Anthropic is now in the big three: OpenAI, Google DeepMind, Anthropic

Week 39 of Coding with Intelligence

📰 News

• Anthropic raises $4B from Amazon

• Elegant and powerful new result demonstrates shortcomings of current LLMs

  I think this result also demonstrates the importance of evaluating prompting strategies even with current Best-in-Class LLMs. E.g. if you're building a Q&A system it makes sense to first let the LLM generate related information as directly using all the information available in the model doesn't always work. arXiv: https://arxiv.org/abs/2309.12288

📦 Repos

• Magentic: LLMs as Python functions

  YALA: Yet Another LangChain Alternative

• (Offline) Evaluation and Unit Testing for LLMs

• LLM Guard: scan LLM inputs & outputs

📄 Papers

• Boolformer: Symbolic Regression of Logic Functions with Transformers

• Monarch Mixer: alternative to Transformer architecture for LLMs from Stanford research lab

• GPT-4V System Card paper

• Distilling step-by-step: Outperforming larger language models with less training data and smaller model sizes

  Also available as a fine-tuning alternative on Vertex AI

• Chain-of-Verification Reduces Hallucination in Large Language Models

• Contrastive Decoding Improves Reasoning in Large Language Models

📱 Demos

• Code Arena by Glaive

• Spotify uses voice cloning to translate podcasts to different languages using original speaker's voice

  Article about Spotify's move https://www.digitaltrends.com/movies/spotify-using-ai-to-clone-and-translate-podcasters-voice/

• In-browser Low Latency voice ordering system

  Latency is not to be underestimated for voice AI UX. This demo shows how natural it can start to feel once latency is low enough.

🛠️ Products

• ElevenLabs improves long-form audio support

• Lightrail: AI powered Spotlight for Developers

  Author wrote a thoughtful peace on LLM UX too: https://vishnumenon.com/lightrail/2023/09/17/AI-Interfaces.html

📚 Resources

• ICML Efficient Systems for Foundation Models talk videos

  Tim Dettmers on 4-bit inference, Philippe Tillet from OpenAi talking about Triton, Cerebras LLM training, LLM tooling panel.

• Replit eval-python: Python eval as a service

  There's also a free endpoint: https://eval-python.replit.app/eval curl -X POST -H 'Content-Type: application/json' -H 'Authorization: Bearer mycooltoken' -d '{"code": "print(\"hello, world!\")"}' https://eval-python.replit.app/eval

• RAG is more than just embedding search

• Collection of papers on LLM merging

• A Hackers' Guide to Language Models by Jeremy Howard from FastAI

• Our Humble Attempt at “How Much Data Is Needed to Fine-Tune”

• Prompt engineering for Claude's long context window

---

Want more? Follow me on Twitter! @ricklamers