TLDR Index
TLDR AI

TLDR AI 2024-12-05

6 sections20 entrieslinks blockcomplete parse

Parse notes

  • Subscription/account footer truncated before parsefooter_truncated

01GENERAL

Sponsored

SOME COMPANIES 🤯

👋 Say hi to Bland, the world's most realistic AI phone-calling

bland.aiSponsor

agent

Chances are, you've already spoken to it without even noticing... It can:

bland.ai

02HEADLINES & LAUNCHES

OPENAI 12 DAYS OF SHIPMAS

OpenAI will be having 12 live streams over the next 12 days to ship new product and model features.

threadreaderapp.com1 min read

META'S NUCLEAR ENERGY PLANS

Meta announced plans to collaborate with nuclear energy developers through a new request for proposals. The initiative aims to add 1-4 gigawatts of nuclear capacity in the U.S. to support Meta's AI innovation and sustainability efforts.

sustainability.atmeta.com4 min read

AWS REINVENT TOP ANNOUNCEMENTS

AWS announced various additions and tools including bringing prompt routing and caching to its Bedrock LLM service at AWS re:Invent 2024.

aws.amazon.com4 min read

03RESEARCH & INNOVATION

GENIE 2: A LARGE-SCALE FOUNDATION WORLD MODEL

Genie 2 is a large scale latent diffusion model for world generation. It works with character control as input and is also classifier free. The outputs are stunning and maintain control and consistency over time.

deepmind.google9 min read

MOTION PROMPTING

Motion Prompting is a method that trains a video generation model with novel input types. Instead of just text or an image, it uses text, first image frame, and a tracking field of pixels. This then allows for novel control at inference time where it can provide a new pixel field (e.g., for an object moving in a new direction) and generate a video based on that. The method is extremely compelling, but not open sourced.

motion-prompting.github.io6 min read

04ENGINEERING & RESOURCES

FREE PROCESS REWARDS WITHOUT PROCESS LABELS

Process reward models (PRMs) offer finer-grained feedback by evaluating reasoning trajectories step-by-step, unlike outcome reward models (ORMs), which assess entire responses. However, training PRMs requires detailed intermediate annotations, posing significant challenges. This paper shows that an implicit PRM can be derived at no additional cost by training an ORM on response-level labels using log-likelihood ratios of policy and reference models, enabling optimization without requiring specific loss objectives.

github.comGitHub repository

DYNAMIC 4-BIT QUANTIZATION

The Unsloth team aims to shrink a 20GB language model into 5GB without sacrificing accuracy. There are many algorithms that attempt this, but there are a number of challenges around outliers and compressibility. Llama is a famously hard model to quantize. The method selectively avoids quantizing certain parameters, which greatly improves overall accuracy.

unsloth.ai18 min read

TACKLING REPETITIVE IMAGE GENERATION

AccDiffusion v2 improves diffusion models for generating higher-resolution images without training, addressing challenges like object repetition and local distortion.

github.comGitHub repository

05MISCELLANEOUS

OPTIMIZING AI INFERENCE AT CHARACTER AI

Character AI has a strong inference pipeline. In this post, the team walks through their int8 quantization and flash attention 3 implementation. It's a good read for folks interested in serving large language models at scale.

research.character.ai18 min read

CERTAIN NAMES MAKE CHATGPT GRIND TO A HALT, AND WE KNOW WHY

OpenAI's ChatGPT employs hard-coded filters to prevent the AI from generating false statements about specific individuals, leading to broken conversations when these names are mentioned. This approach stems from past incidents like defamation lawsuits against OpenAI, prompting the company to restrict outputs related to contentious names. The hard-coded filters can pose challenges as they limit ChatGPT's functionality and leave the system vulnerable to adversarial attacks.

arstechnica.com5 min read

FINETUNING LLM JUDGES FOR EVALUATION

Evaluating LLMs has become challenging due to their complex and open-ended outputs. Traditional evaluation relies on human judgment, which is detailed but inefficient. Thus, automatic metrics and model-based evaluations like LLM-as-a-Judge are crucial for scalable assessment. Innovations such as specialized fine tuned judges (e.g., Prometheus) and synthetic data generation are enhancing LLM evaluation's precision and adaptability across diverse tasks and domains.

cameronrwolfe.substack.com52 min read