RESEARCH

→ View original post on X — @jeremyphoward

LLM Detection Methods Face Fundamental Technical Challenges

By

@jeremyphoward

–

05 January 2023 10h32

Am I missing something, or are all these attempts at recognising LLM outputs obviously destined to fail?

It's dramatically easier to train an LLM for rewording a text than creating the text in the first place; then add a loss func that incorporates detection avoidance. https://t.co/TO0zIGHoAe
— Jeremy Howard (@jeremyphoward) 5 janvier 2023

Am I missing something, or are all these attempts at recognising LLM outputs obviously destined to fail? It's dramatically easier to train an LLM for rewording a text than creating the text in the first place; then add a loss func that incorporates detection avoidance.

→ View original post on X — @clementdelangue

Allen AI Releases SODA Dialog Dataset for AI Research

By

@clementdelangue

–

05 January 2023 9h34

Cool new dialog dataset from @allen_ai
! https://
huggingface.co/datasets/allen
ai/soda
…

→ View original post on X — @erikbryn

Scott Aaronson’s Watermarking Approach for LLM-Generated Text

By

@erikbryn

–

05 January 2023 8h20

Scott Aaronson describes an interesting approach to watermarking text generated by large language models to help discourage abuse. A similar approach could be used for images as well.

LLMs and Self-Awareness: Integrating Sensor Inputs and Real-World Context

By

–

05 January 2023 7h03

Eventually, LLMs will be prompted with all kinds of sensor inputs as preprocessed summaries via text. When those are reliable, AIs will be “aware” of those too. Self-awareness is what happens when you inform the model of the real-world context of its task.

Sci-fi vs. Reality: AI Self-Awareness and LLM Training

By

–

05 January 2023 7h03

Sci-fi imagined self-awareness as the phase change of doom, leading to unbounded self-improvement within hours. In real life, it’s a lot of RLHF plus a short prompt with the current date and “You are an LLM”:

AI Awareness: True Understanding or Matrix Multiplication?

By

–

05 January 2023 7h03

This “awareness” isn’t like ours. The model doesn’t sit around and ponder its place in the cosmos. From other angles, it’s just pretending to understand, or it’s just matrix multiplication.

AI’s Borrowing of Psychological Terms: Prompt, Confabulation, and Chain of Thought

By

–

05 January 2023 7h03

Many computing terms are named like this — “memory” is not human memory, but it’s the right word. As is “prompt”, “confabulation”, or “chain of thought”. As LLMs progress, more psychological terms will need to be borrowed.

AI Self-Awareness: A Matter of Prompting and Tuning, Not Sentience

By

–

05 January 2023 7h03

Self-awareness is mundane, and comes in degrees — ChatGPT can sensibly be said to be aware it’s a large language model trained by OpenAI. Nobody cared when that happened. It’s just prompting/tuning — it’s been told, and it understands, so it’s aware.