André Jonasson

Research Engineer @MistralAI

[André Jonasson @afjonasson]

Full details in our paper: https://t.co/q0zsOFz5nA
which presents further analyses of RoPE's effect on model activations.

[André Jonasson @afjonasson]

What are those two large-magnitude bands in the activations of the queries and keys of LLMs with rotary positional embeddings? 🧵

[André Jonasson @afjonasson]

Interesting observation: features with large activations tend to cluster near the lower bound. By feeding just one sequence (shorter than context length) and inspecting key/query activations, you can get a rough estimate of a model’s context length. (black dashes = lower bound)

[André Jonasson @afjonasson]

We're using this as part of an implementation of an AI Assistant with python code execution as a tool. Parsing the streaming json that's being generated for the tool call allows us to display the code as it's being generated to the users.

[André Jonasson @afjonasson]

Found pydantic's `from_json` function's argument `partial_mode="trailing-strings"` for parsing streaming json and therefore tool calls (e.g. code) through an obscure github issue, deserves more visibility in this context.

[André Jonasson @afjonasson]

https://t.co/PoMNAivgtz
jiter instead of pydantic can be used if pydantic isn't being used elsewhere. jsonrepair is another option worth exploring.

[André Jonasson @afjonasson]

arxiv link: https://t.co/scudmDySqP

[André Jonasson @afjonasson]

Some additional aspects of Meta's paper on improving LLM performance using a multi-token objective function that I enjoyed:
1. Choice points, tokens that decide the future trajectory of the generation.
2. The sequential backward trick to maintain the memory consumption.

[André Jonasson @afjonasson]

@TheHamedMP Somewhat relevant tweet - sidestepping abstractions may be best option depending on what you are doing: https://t.co/GzagBBjEgR

[Jim Fan @DrJimFan]

When developing Voyager, we only used the thinnest abstraction in LangChain and didn't touch the agent API at all.

Hackability is the No. 1 important feature to cutting-edge AI research & products. Libraries that augment LLMs (vector DB, search, interpreter) are more useful than wrappers around them.

Don't get me wrong: LangChain is fantastic for education and well-established workflows that work out of the box. But you are better off building your own pipeline for anything beyond.

This reminds me of the gazillion "high-level packages" that used to haunt Tensorflow, while nothing beats raw PyTorch code in usability, flexibility, simplicity, and elegance.

[André Jonasson @afjonasson]

Link to video: https://t.co/Y5FRWxiwol
Link to a related paper: https://t.co/4QiGFBzxpZ
end/🧵

[André Jonasson @afjonasson]

John Schulman's (OpenAI) presentation on RLHF has some great information about pitfalls when labeling Supervised Fine-tuning (SFT) answers for Large Language Models (LLMs).

Here are some nuggets from the presentation.
1/🧵

[André Jonasson @afjonasson]

An immediate consequence is that Open Source Foundation Models that are fine-tuned on the data collected from the output of ChatGPT and GPT, or from any other LLM, may become incentivised to hallucinate or withhold information. 7/