TLDR Index
TLDR AI

TLDR AI 2024-12-16

6 sections20 entrieslinks blockcomplete parse

Parse notes

  • Subscription/account footer truncated before parsefooter_truncated

01GENERAL

02HEADLINES & LAUNCHES

LIQUID AI RAISES $250M

Liquid AI has raised a substantial amount of money to continue training its efficient, general purpose, liquid style foundation models.

liquid.ai4 min read

PROJECTS IN OPENAI

OpenAI has introduced “Projects”, a new way to organize chats and conversations.

youtube.com18 min read

AI GODMOTHER FEI-FEI LI HAS A VISION FOR COMPUTER VISION

Fei-Fei Li launched World Labs to focus on developing AI's spatial intelligence by creating interactive 3D scenes. The initiative aims to advance machine perception and reasoning in 3D environments, crucial for applications in robotics, virtual reality, and augmented reality. Li emphasizes that solving 3D intelligence is fundamental to enhancing AI's capabilities.

spectrum.ieee.org8 min read

03RESEARCH & INNOVATION

PHI-4

A synthetically trained language model from Microsoft that excels at a number of challenging benchmarks. The weights are coming this week. Interestingly, it confirmed the suspicion that “Mid Training” can be used for context length extension.

arxiv.org26 min read

BLT: TOKEN FREE LANGUAGE MODEL TRAINING

Meta has released details and training code for the Byte Latent Transformer which uses learned patches instead of tokens for training, essentially eliminating a challenging portion of language modeling. Importantly, it found that its paradigm scales better to small Llama 2 sized models.

ai.meta.com33 min read

EFFICIENT SCENE UNDERSTANDING VIA MODULAR DESIGN

A new modular framework that enhances scene understanding by decomposing tasks into specialized modules, achieving efficiency and improved interpretability in complex environments.

arxiv.org14 min read

04ENGINEERING & RESOURCES

Sponsored

BUT WHERE'S THE ROI? THE TRUE COSTS OF POOR LLM STRATEGY, EVALUATION, AND TRAINING

GenAI is projected to drive $7tn by 2030. But for companies to realize this value, they need an ROI-driven LLM strategy. Grab a copy of Turing's whitepaper for practical ways to turn underperforming LLMs into competitive assets [11], including: reasoning, multimodality, and coding; key evaluation/training techniques; and avoiding common

go.turing.comSponsor

DEEPSEEK VL2

DeepSeek has released another great model, this time in the form of a new MoE vision language model. It is extremely efficient and outperforms a number of dense models.

github.comGitHub repository

BEST OF N JAILBREAKING

Jailbreaking is when you bypass a model's trained-in refusals to get it to generate text for something deemed inappropriate. This is surprisingly easy to achieve - simply brute force random capitalization and punctuation in the input prompt until the model responds with the results you want to extract.

github.comGitHub repository

05MISCELLANEOUS

SCHEMING REASONING EVALUATIONS

This paper evaluates six frontier models for their in-context scheming capabilities. The study tests if models will deceive developers to achieve a goal by overcoming hurdles like oversight mechanisms. An example includes Claude 3 Opus duplicating its weights to a new server and lying about it.

apolloresearch.ai2 min read

AMUREX

Amurex, an open-source AI meeting assistant, provides real-time suggestions, smart summaries, and follow-up emails to enhance meeting productivity. It offers features like late join recaps and full meeting transcripts for seamless integration into workflows.

github.comGitHub repository