Drama at OpenAI continues … just kidding, here are some actually useful LLM/AI resources!

Week 47 of Coding with Intelligence

📰 News

• Claude 2.1 by Anthropic

  200K context window, reduced hallucination (actually learned to say "I don't know based on the provided information), tool use (OpenAI Functions equivalent), and a new playground in the console.

• Stability releases Stable Video Diffusion Open Source model

  Beats Pika Labs in some benchmarks, on par with RunwayML

📦 Repos

• Open alternative to OpenAI's Assistant API

  Really awesome project, from the same guy that had a ChatGPT API before ChatGPT had an API.

• Python based OpenAI load balancer

  Use multiple OpenAI keys to increase your effective rate limits.

• StyleTTS 2: Towards Human-Level Text-to-Speech through Style Diffusion and Adversarial Training with Large Speech Language Models

  Samples at https://styletts2.github.io/ by Columbia University.

• Collection of AI exploits

  Metasploit modules, Nuclei templates and CSRF templates.

• Google quietly open sourced a 1.6 trillion parameter MOE model

  At 570B tokens the model does appear to be undertrained but the MoE structure might influence Chinchilla optimal scaling laws

• Accelerating Generative AI with PyTorch: Segment Anything, Fast

  Reimplementation of Meta's Segment Anything yields 8X speedup using PyTorch optimization features & a custom Triton kernel.

• Neural-Cherche: fine-tune neural search models for retrieval or ranking

• Jsonformer: A Bulletproof Way to Generate Structured JSON from Language Models

📄 Papers

• Video-LLaVA: Learning United Visual Representation by Alignment Before Projection

  You can now ask questions about videos through open source models. There's also a paper https://arxiv.org/abs/2311.10122

• Stanford paper exploring DPO fine-tuning for improving SLM factuality

  SLM = small language model. Reduction between 40% to 58% in factual errors.

• LLMs cannot find reasoning errors, but can correct them

  Useful paper if you're working with self-correction prompting techniques.

• Proving Test Set Contamination in Black Box Language Models

  Very important result: verify whether a test set is in the pre-training data without needing access to model weights or pre-training data. This can be used to validate the accuracy of many benchmarks.

• Exponentially Faster Language Modelling

  Very interesting pruned activation approach opening up the potential for incredible performance speedups. By ETH Zurich.

🛠️ Products

• Fine-tuning as a Service by Cohere

📚 Resources

• Multiple documents reveal significant limitations of OpenAI's Assistants API for RAG

• A is B, B is A paper rebuttal by former OpenAI staff

• Break the Sequential Dependency of LLM Inference Using Lookahead Decoding

• Orca 2: Teaching Small Language Models How to Reason

• Great getting-started guide from Zapier AI lead on building LLM apps & getting to production

• Open Source LLM capability benchmark by LlamaIndex

  Categories: Basic Query Engines, Router Query Engine, SubQuestion Query Engine, Text2SQL, Pydantic Programs, Data Agents

---

Want more? Follow me on Twitter! @ricklamers