TLDR Index
TLDR AI

TLDR AI 2023-04-27

6 sections16 entriesinline urlcomplete parse

Parse notes

  • Subscription/account footer truncated before parsefooter_truncated

01GENERAL

02HEADLINES & LAUNCHES

03RESEARCH & INNOVATION

LANGUAGE MODELS FOR PLANNING

Language models are great at producing plausible text and in certain cases reasonable plans. However, in general, they fail to generate long horizon plans. Without changing the model at all, even via fine-tuning, with special prompts we can dramatically improve the success of both plan generation and plan execution. Specifically, language models generate a plan language (fed a few-shot example) and that plan is executed by an actual planner then the resulting plans are translated back to natural language. The improvement is massive.

arxiv.org24 min read

MACHINE TRANSLATION NEEDS ATTENTION

The Structured State Spaces for Sequences (S4) model, successful in various tasks, is applied to Machine Translation (MT), but lags behind the Transformer by 4 BLEU points and struggles with long sentences. This gap is attributed to S4's inability to summarize full source sentences in a single hidden state, which can be remedied by introducing an attention mechanism.

arxiv.org12 min read

HOW TO TRAIN LLMS TO BE COMPETENT AND ETHICAL

One negative side effect of training agents to maximize reward is that they often become power seeking, meaning they try to increase their resources and access to information. This is a novel finding and somewhat hard to measure, this paper introduces a benchmark to help understand the power seeking behavior of these language models. Additionally, they give insights on how to train models to both act with high utility and according to moral guidelines. The world needs more high quality benchmarks and this is an excellent step in that direction.

arxiv.org36 min read

04ENGINEERING & RESOURCES

DEEPFLOYD IF CODE RELEASED

The next image model from Stability AI has a code release. The model weights, blog, inference, and tooling will be coming in a few days. It is a multi-stage model with a frozen T5 text encoder and two super resolution models. One big win here is the ability to render extremely high quality text. The claim is that this model will likely be for research only at first, then open after some community feedback.

github.comGitHub repository

POSEVOCAB: LEARNING POSE EMBEDDINGS FOR HUMAN AVATAR MODELING

PoseVocab, a novel pose encoding method for human avatar modeling, constructs key poses and latent embeddings to effectively encode dynamic human appearances and achieve pose generalization and temporal consistency. The method outperforms state-of-the-art baselines in synthesis quality through the use of feature lines, a compact 3D representation, and a hierarchical query strategy for interpolating pose embeddings.

lizhe00.github.io5 min read

05MISCELLANEOUS

A LOOK AT LAION

The debate over how to regulate LAION, a database with over 5 billion images commonly used to train AI models, is becoming more contentious.

archive.ph7 min read