Stability AI rises from the ashes as Black Forest Labs: FLUX.1 is insane

Week 32 of Coding with Intelligence

📰 News

• Gemma 2 2B release

• GitHub announces GitHub Models

  They’ve built a playground too.

• Announcing the results of the inaugural AlgoPerf: Training Algorithms benchmark competition

  By ML Commons. Shampoo submission by Meta takes #1 spot. Find it on GitHub here

• Black Forest Labs exits stealth

  "Our mission is to develop and advance state-of-the-art generative deep learning models for media and to push the boundaries of creativity, efficiency and diversity."

• LangChain launches LangGraph Studio: local agent IDE

📦 Repos

• Cold Compress

  Cold Compress is a hackable, lightweight, and open-source toolkit for creating and benchmarking cache compression methods built on top of GPT-Fast, a simple, PyTorch-native generation codebase.

• Prompt Poet library release: Prompt Design at Character.AI

  On GitHub https://github.com/character-ai/prompt-poet

📄 Papers

• Beyond Euclid: An Illustrated Guide to Modern Machine Learning with Geometric, Topological, and Algebraic Structures

  Mechanistic interpretability background material

• Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters

  Test-Time inference scaling is becoming more & more of a no-brainer.

• Large Language Monkeys: Scaling Inference Compute with Repeated Sampling

  "Is inference compute a new dimension for scaling LLMs?" - Azalia Mirhoseini (@Azaliamirh)

• MoMa: Efficient Early-Fusion Pre-training with Mixture of Modality-Aware Experts

• Transformer Layers as Painters

📱 Demos

• Turbo Edit on HF based on SDXL Turbo

  Works pretty well!

📚 Resources

• Karpathy: RLHF is barely RL

• OpenAI structured outputs guide by Simon Willison

• DPO from Scratch in a Jupyter Notebook

  By Sebastian Raschka from Lightning AI

• IsoFLOP curves of large language models are extremely flat

  isoFLOP curves=curve for same flops but different size model (in paramater count)

• Beat GPT-4o at Python by searching with 100 dumb LLaMAs

---

Want more? Follow me on X! @ricklamers