Semanlink - Emergence AND Transformers

Tags:

- Intelligence has emerged: why? how?
- Let's study this with *controlled experiments* and *toy models*
- Clean and clear insights that peer slightly behind the magic curtain

emergence; rise of scale. Phase transition: fct(compute time for training: nb of params, dataset size)

(12:25) Transformers: the jump when compared to traditional NN: instead on oparating on single input X, operates on a set of inputs. (cf. training to say whether an image contains 2 elements of the same kind). **The aptitude to compare elements in the input sequence enables analogies** - the essence of reasoning.(Classical NN treats one high dimensional vector, tranformers hanlde sets of vectors. Attention layer replaces learned filters W of classical NN by the other elements in the sequence. Attention module compares part of input against other part of input. "Relative machines rather than absolute". Sets are a very powerful level of abstraction)

emergence
> The truth is that nobody has a clue what is going on"
>
> "**Something unknown is doing we don't know what**" (A. Eddington)

being inspired by physics' methodology: controled experiments and toy mathematical models

2023-03-28 About

Matthew Honnibal sur Twitter : "Some of the things ChatGPT can do are emergent behaviours... Other things it can do have been specifically taught to it. Is there some speculative list somewhere about what tasks were supervised?" / Twitter

Tags:

2023-01-14 About

Some remarks on Large Language Models

Tags:

> There turned out to be a phase shift somewhere between 60B parameters and 175B parameters, that made language models super impressive.

> **The performance of current days language models are not obtained by language modeling**
>
>    - [Traditional] LMs are not [grounded](tag:grounded_language_learning)
> 
> **3 conceptual steps between GPT-3 and chatGPT: Instructions, code, RLHF.** The last one is, I think, the least interesting despite getting the most attention
>
> Instruction tuning: For example, the human annotators would write something like "please summarize this text", followed by some text they got, followed by a summary they produced of this text. -> Some symbols ("summarize", "translate", "formal") are used in a consistent way together with the concept/task they denote. And they always appear in the beginning of the text. -> the act of producing a summary grounded to the human concept of "summary"
>
> code: programming language code data, and specifically data that contains both natural language instructions or descriptions (in the form of code comments) and the corresponding programming language code. This produced another very direct form of grounding. the human language describes concepts (or intents), which are then realized in the form of the corresponding programs.
>
> "[RL with Human Feedback](tag:reinforcement_learning_from_human_feedback)". This is a fancy way of saying that the model now observes two humans in a conversation, one playing the role of a user, and another playing the role of "the AI", demonstrating how the AI should respond in different situations. This clearly helps the model learn how dialogs work, and how to keep track of information across dialog states (something that is very hard to learn from just "found" data).

2023-01-03 About

Unveiling Transformers with LEGO - YouTube

Tags:

2022-06-30 About