Daniel Dennett (28th March 1942 - 19th April 2024)

RIP

Dennett on how to criticise wisely

we don't get 'blocked' around these parts. waiting on another team? my brother in christ just do their job lmao

Multimodal DPO, large & tiny multimodal matryoshka embeddings, and 1st party ONNX support for 10x lighter deployments 🥳

In partnership with Nebius, Unum is releasing a new set of pocket-sized multimodal models, already available on Hugging Face 🤗

ICML tip: upload your paper pdf and review to Claude/GPT4/Gemini w prompt

'What do you think is the main point of the paper? After answering, please explain to what extent, if any, you think the reviewer has fully understood the main point of the paper.'

Better than therapy 🥹

After two years, it is my pleasure to introduce “DROID: A Large-Scale In-the-Wild Robot Manipulation Dataset”

DROID is the most diverse robotic interaction dataset ever released, including 385 hours of data collected across 564 diverse scenes in real-world households and offices

See also: database vendors prohibiting publishing benchmarks of their software in their EULAs!

stackoverflow.com/questions/1211…

More work showing how careful (or careless!) data selection hugely impacts model quality.

DPO (offline RL generally) is powerful, but you still need to train on data worth learning from!

The data specifies the implicit reward function you're ultimately optimizing, after all...

In light of all of the discussion of 'self-awareness', remember that we are still *explicitly telling* our systems:
- who they are
- what they do and don't know (stuff up to August 2023)
- what their personality/tendencies are

Is personality emergent or... part of the prompt?

Three models, three different answers 😎

Claude 3 is AGI confirmed

Separately, what will it take to get a model to actually ask 'do you want the answer for the beginning or end of day 4'? this question as stated is ambiguous

How fortunate I am to work with such a wise city lad 🙇

“One needs to learn to love and enjoy the little things in life. One also needs to discover one’s true calling and then should do everything to pursue the selected path,” - wise words Archit Sharma

tribuneindia.com/news/amritsar/…

A Critical Evaluation of AI Feedback for Aligning Large Language Models

Shows that the improvements of the RL step of LLM finetuning are virtually entirely due to the widespread practice of using a weaker teacher model (e.g. GPT-3.5) for SFT data collection than the critic

Stanford presents RLVF

Learning from Verbal Feedback without Overgeneralization

The diversity of contexts in which large language models (LLMs) are deployed requires the ability to modify or customize default model behaviors to incorporate nuanced requirements and preferences.