TLDR AI 2024-11-11

Read the whitepaper

MODAL BUYS TIDBYT

The first acquisition by the elastic scaling GPU company. Modal bought Tidbyt, an NYC-based hardware company, for the in-house talent of the team around infrastructure and containerization.

OPENAI FACES AI IMPROVEMENT SLOWDOWN

OpenAI's next major model, code-named "Orion," reportedly shows less improvement over previous models, suggesting a slowdown in AI advancement. To tackle this, OpenAI has formed a foundations team focused on enhancing models with alternative methods, such as synthetic data training and post-training adjustments, as the availability of new data has decreased.

WORLD'S LARGEST OPEN-SOURCE MODEL

Near Protocol has announced plans to create a massive 1.4 trillion parameter open-source AI model. It aims to surpass existing models like Meta's Llama.

FRONTIER MATH BENCHMARK

Epoch AI has released a new challenging mathematics benchmark. Most frontier models are unable to solve more than 2% of the problems.

BITNET A4.8: 4-BIT ACTIVATIONS FOR 1-BIT LLMS

One of the key challenges with 1.58bit LLMs was the lack of hardware acceleration support. This work proposes 4.8bit activations to take advantage of the INT4/FP4 kernels in new hardware. It comes at no runtime cost.

SUPERCHARGING CLIP WITH LLMS

LLM2CLIP combines CLIP's visual and textual alignment with the advanced language understanding of LLMs.

TORCH COMPATIBLE MUON OPTIMIZER

Muon was the optimizer used for the GPT-2 training record. It is a momentum adapted SGD style method. This repository contains an implementation that can be dropped in for AdamW.

MOCHI VIDEO MODEL WITH OPTIMIZED INFERENCE

Mochi is the best open source text-to-video model. On launch, it took 8 H100s to run. Now, thanks to the community, it can be run on a single 48GB L40 with no loss in quality.

TRAINABLE PYTORCH REPRODUCTION OF ALPHAFOLD3

Protenix is a working and trainable reproduction of AlphaFold3, a protein folding project from DeepMind. It was written by Bytedance's 'AI for Science' team.

LLAMAPREVIEW

LlamaPReview is an AI assistant for GitHub that offers one-click installation and auto-reviews of pull requests with context-aware analysis. It supports multiple programming languages and integrates with GitHub Actions, providing insightful feedback directly on PRs. Currently available for free, it enhances code quality by identifying issues and suggesting optimizations.

SMOLLM2

Hugging Face's SmolLM2 is a compact language model family with sizes ranging from 135M to 1.7B parameters trained on 11 trillion tokens. These models efficiently run on-device and support various tasks, with weights available under the Apache 2 license. Quantized models like the 1.7GB and 138MB versions provide diverse flexibility for different computational needs.

EMBEDDINGS ARE UNDERRATED

Machine learning embeddings can transform technical writing by enabling mathematical comparisons of arbitrary text, improving features like recommendation systems with semantic similarities. They place text within a multi-dimensional space, providing intuitive semantic relationships, useful for tasks like identifying related content. Docs site owners who offer embeddings for their content could foster innovative applications from their communities.

SAMSUNG DEBUTS AI-POWERED 'NEXT-GENERATION BIXBY,' BUT YOU CAN'T USE IT YET

Samsung has launched a "next-generation Bixby" with enhanced AI capabilities on the Galaxy W25 and W25 Flip in China.

EVEN MICROSOFT NOTEPAD IS GETTING AI TEXT EDITING NOW

Microsoft is integrating AI-powered text editing in Notepad, enabling users to rephrase, adjust tone, and modify text length through a feature called Rewrite.

AI FOR REAL-TIME FUSION PLASMA BEHAVIOR PREDICTION AND MANIPULATION

A new multimodal machine learning methodology enhances super-resolution data for better analysis of complex fusion plasma phenomena, such as Edge Localized Modes (ELM), aiding in the stabilization of future fusion reactors.