Sep 5, 2024

Show HN: An open-source implementation of AlphaFold3

Ligo's open-source implementation of AlphaFold3 aims to advance biomolecular structure prediction, initially releasing single-chain prediction capabilities with plans to expand to ligand, multimer, and nucleic acid predictions. Sign up for beta testing here.

AlphaProteo generates novel proteins for biology and health research

AlphaProteo, a new AI system by Google DeepMind, designs novel proteins with high binding affinities, potentially revolutionizing drug design and disease understanding.

Show HN: Graphiti – LLM-Powered Temporal Knowledge Graphs

Graphiti dynamically constructs temporal Knowledge Graphs from both structured and unstructured data, enabling complex relationship tracking and historical context maintenance over time.

Launch HN: Maitai (YC S24) – Self-Optimizing LLM Platform

Maitai introduces a self-optimizing LLM platform that enhances request routing, autocorrects errors, and fine-tunes application-specific models for incremental improvements, demonstrated through a demo video and an interactive game at maitaistreasure.com.

The Memory Wall: Past, Present, and Future of DRAM

DRAM scaling has significantly slowed, with density increases now only doubling every decade compared to the historical rate of doubling every 18 months, highlighting a stark departure from the rapid advancements seen in the past.

Reflection 70B, the top open-source model

Reflection 70B, an open-source model, introduces a novel Reflection-Tuning technique for self-correction in LLMs, setting a new standard in AI development.

LongLLaVA: Scaling Multi-modal LLMs to 1000 Images Efficiently via Hybrid Architecture

LongLLaVA, a hybrid Multi-modal Large Language Model (MLLM), introduces a novel architecture combining Mamba and Transformer blocks to enhance long-context understanding in images and videos.

[R] Fixed Point Diffusion Models

The Fixed Point Diffusion Model (FPDM) introduces an implicit fixed point solving layer into diffusion models, enhancing image generation by making the process more efficient and effective.

[R] Is exploration the key to unlocking better recommender systems?

Google DeepMind's research suggests that exploration in recommendation systems can significantly enhance long-term user experience by diversifying content.

[R] What if self-attention isn’t the end-all be-all?

Masked Mixers propose an alternative to self-attention mechanisms in transformers, addressing information loss issues.

[R] DiffUHaul: A Training-Free Method for Object Dragging in Images

DiffUHaul enables seamless relocation of objects within images, addressing the challenge of spatial reasoning in image editing without the need for training.

The New Test Images – Image Compression Benchmark

The new test images for image compression research replace outdated ones like lena and pepper, offering high-resolution and high-precision variants to challenge and evaluate algorithms more effectively.

[P] What's the best performance metrics for segmentation tasks and how to improve performance of highly skewed dataset?

For brain tumor segmentation tasks with highly skewed classes, the choice of performance metrics significantly impacts the model's ability to distinguish between the majority background and the minority tumor classes.

CUDA-Free Inference for LLMs

Achieving FP16 inference for LLMs like Meta's Llama3-8B and IBM's Granite-8B solely with OpenAI’s Triton Language resulted in 0.76-0.82x performance compared to CUDA-based methods on Nvidia GPUs.