Email Details

OpenAI Academy πŸ“š, Alibaba open source models 🌐, Evaluating Long-Context ModelsπŸ’»

Alibaba has released over 100 open-source AI models, enhancing its technology to compete with rivals. The new Qwen 2.5 models have upgraded math 

TLDR

TLDR AI 2024-09-24

πŸš€

Headlines & Launches

OpenAI Academy (8 minute read)

OpenAI is starting a program for low and middle income countries to expand access to AI knowledge. It also has a professional translation of MMLU (a standard reasoning benchmark) in 15 different languages.
China's Alibaba launches over 100 new open-source AI models, releases text-to-video generation tool (3 minute read)

Alibaba has released over 100 open-source AI models, enhancing its technology to compete with rivals. The new Qwen 2.5 models, upgraded in math and coding, span applications from automobiles to gaming. Alibaba has also launched a new proprietary model, Qwen-Max 2.5, and a text-to-video tool to strengthen its AI and cloud services offerings.
Apple Intelligence in More Countries (4 minute read)

Apple's iOS 18.1 will introduce key AI features, including an enhanced Siri, generative AI tools in Photos, and ChatGPT integration. iOS 18.2 will expand these features with localized support in various English-speaking countries and add Image Playground and Genmoji. Future updates, such as iOS 18.4, will further enhance Siri's personalization and introduce new language support.
🧠

Research & Innovation

The Practitioner's Guide to the Maximal Update Parameterization (31 minute read)

Maximal Update Parameterization, or muP, is a way to initialize your model so that you can transfer hyperparameters at any scale. This blog from Eleuther and Cerebras includes a minimal nanoGPT example and lots of instruction on how the process works.
Teaching Diffusion Models to Count (25 minute read)

Getting a diffusion model to generate a certain number of an object is currently challenging. This work introduces a counting token that allows a model to successfully generate just a few or very many of a certain object. It isn't perfect as it builds on base stable diffusion, but it does much better than current alternatives.
Evaluating Long-Context Models (19 minute read)

Researchers have developed a standardized evaluation protocol to compare methods for extending language models to handle long document contexts.
πŸ§‘β€πŸ’»

Engineering & Resources

New research by Orca: 62% of organizations have deployed an AI package with at least one CVE (Sponsor)

The results are in - and they're not pretty. Orca Security has published its new State of AI Security Report, based on scans of billions of cloud assets. Organizations and cloud providers are prioritizing AI development velocity over security considerations, leading to preventable risks. See the full findings
3D Vision Transformer (GitHub Repo)

This repository hosts an implementation of a 3D Vision Transformer designed for efficient field boundary delineation using time series satellite imagery. The model leverages spatio-temporal correlations to improve accuracy and robustness, particularly in challenging conditions such as partial cloud cover.
Speeding Up Long-Context Processing with CritiPrefill (GitHub Repo)

CritiPrefill is a method designed to accelerate the prefilling phase of long-context processing in large language models. By identifying and skipping non-essential computations, this approach speeds up the process by up to 3x on certain models.
Improving LLM Reasoning with MAgICoRe (20 minute read)

MAgICoRe is a new strategy to improve reasoning in large language models by addressing challenges in refinement processes. It categorizes problems by difficulty, using simpler strategies for easy tasks and multi-agent iterative refinement for harder ones.
🎁

Miscellaneous

Document Similarity Search with ColiPali (17 minute read)

A great blog post that explores the popular multimodal RAG system and how it can be used to solve practical problems.
Developing New Materials with AI (6 minute read)

By analyzing X-ray crystallography data, the model could help researchers develop new materials for many applications, including batteries and magnets.
When will AI outthink humans? (13 minute read)

This article explores when AI might surpass humans in cognitive volume, introducing "thought-hours" as a metric to quantify AI's cognitive output compared to human labor. Using assumptions around reading speeds and productivity, a thought-hour equates to 10,000 tokens. Current trends suggest AI could outthink humans within a decade, given the rapid growth in AI capabilities and cost efficiencies.
⚑

Quick Links

Trustworthy AI apps start with a trustworthy sign in experience (Sponsor)

You can't sell AI to businesses without modern authentication: SSO + MFA + all the bells and whistles. Luckily, you don't need to build it from scratch: Clerk is free for <10k MAUs
Microsoft updates its AI suite with more agents and Copilots (1 minute read)

Microsoft is expanding its generative AI suite to include automated agents, adding features within its Copilot assistants and unveiling a new tool to help multiple workers collaboratively interact with artificial intelligence.
Sam Altman leaves OpenAI board's safety and security committee (2 minute read)

OpenAI said that CEO Sam Altman is leaving the board's safety and security committee, which will now be fully composed of independent board members.
Practitioners Guide to Triton (Jupyter Notebook)

A great lecture from the CUDA Mode group about getting started with the CUDA framework, Triton.

Love TLDR? Tell your friends and get rewards!

Share your referral link below with friends to get free TLDR swag!
Track your referrals here.

Want to advertise in TLDR? πŸ“°

If your company is interested in reaching an audience of AI professionals and decision makers, you may want to advertise with us.

If you have any comments or feedback, just respond to this email!

Thanks for reading,
Andrew Tan & Andrew Carr


If you don't want to receive future editions of TLDR AI, please unsubscribe from TLDR AI or manage all of your TLDR newsletter subscriptions.

Β© 2024 Email Dashboard. All rights reserved.