TLDR Index
TLDR AI

TLDR AI 2024-10-09

6 sections22 entrieslinks blockcomplete parse

Parse notes

  • Subscription/account footer truncated before parsefooter_truncated

01GENERAL

Sponsored

THEM LEGALLY

Criminal hacker group USDoD has allegedly leaked 2.7BN records of personal information — including names, addresses, dates of birth, Social Security numbers, and phone numbers.

get.incogni.ioSponsor

02HEADLINES & LAUNCHES

ANTHROPIC BATCHES API

Anthropic has announced a new batch processing API that enables developers to send batches of up to 10,000 queries per batch. Each batch is processed in less than 24 hours and costs 50% less than standard API calls. This makes processing non-time-sensitive tasks more efficient and cost-effective.

anthropic.com2 min read

UPDATE ON REFLECTION-70B

A deep dive post mortem on what happened with the much hyped Reflection-70B model. It had bugs in its benchmark code. Its team has fixed them and performance is still very impressive, but not quite to the level they claimed.

glaive.ai8 min read

03RESEARCH & INNOVATION

WHAT MATTERS FOR MODEL MERGING AT SCALE

Model merging is a way to combine two models and get the performance of each. However, it doesn't always scale properly with model size. This paper explores what is needed to make it work well with a really large model.

arxiv.org21 min read

NORMALIZED REPRESENTATIONS ON THE HYPERSPHERE

Normalizing the internal representation of language models is the target of a lot of resource activity. This work shows that if you put every internal vector on a hypersphere, it dramatically decreases convergence time for reasonably sized models.

arxiv.org38 min read

CODEMMLU

Language models are getting so good at writing code that many benchmarks are saturating. This paper suggests a benchmark that is harder and aims to determine how models perform on reasoning and code generation tasks.

arxiv.org18 min read

04ENGINEERING & RESOURCES

COLOR LOGIT PRINTING IN THE TERMINAL

The intensify package is a nice way to show logprobs or other real number modulated text based on the color of the terminal.

github.comGitHub repository

MEASURING VIDEO QUALITY IN AI

JEDi is a new metric based on Joint Embedding Predictive Architecture (JEPA). It improves evaluation accuracy while needing fewer samples and better aligning with human assessments, offering a stronger alternative to FVD.

oooolga.github.io4 min read

PLACE RECOGNITION FOR ROBOTICS

PRFusion and PRFusion++ are multimodal models designed to improve place recognition in robotics and computer vision.

arxiv.org17 min read

05MISCELLANEOUS

FOUR-LEGGED ROBOT LEARNS TO CLIMB LADDERS

ETH Zurich used reinforcement learning to enable the ANYMal robot to climb ladders with 90% success rate. Enhancing quadrupedal robots' capabilities in industrial settings, the system improved climbing speed by 232x compared to existing methods. This research highlights the integration of robot morphology and advanced control policies.

techcrunch.com2 min read

THREE SUBTLE EXAMPLES OF DATA LEAKAGE

This article explores the pitfalls of data leakage in machine learning, highlighting two scenarios where improper data handling led to misleading model performance. One company made the mistake of filtering data based on an upper price limit before modeling, while another faced issues by not adhering to a strict chronological split. Key takeaways include the importance of detecting leakage and understanding its impact on model accuracy.

links.tldrnewsletter.com2 min read

THE REAL DATA WALL IS BILLIONS OF YEARS OF EVOLUTION

AI development faces potential challenges from a "data wall" as language models approach training on all available text. This article argues against relying on human analogies for overcoming data limitations, highlighting the vast data and evolutionary processes that contribute to human intelligence. While human learning strategies might not directly apply to AI, this doesn't rule out other modalities or algorithmic progress for advancing AI capabilities.

dynomight.substack.com10 min read