Adam Muhtar

I do NLP stuff. Views are my own. ML @imperialcollege

[Bank of England Research @BoE_Research]

Today’s BU post explores how latest generation of large language models can be applied to prudential supervision. https://t.co/G3wDGnB67Q

[Brad Chattergoon @bradchattergoon]

True. I didn’t realize Linear Algebra is *the* core applied math class until well after i took it… which unfortunately was after I already took the other applied math classes that needed linear algebra. Would have been good to see this meme as a first year undergrad.

[Curt Langlotz @curtlanglotz]

"What will happen to GPT-{n} once LLMs contribute much of the language found online? We find that use of model-generated content in training causes irreversible defects in the resulting models, where tails of the original content distribution disappear." https://t.co/zFM6kyRKsp

[BURKOV @burkov]

My daughter just started college: "Dad, we are studying matrices. There are rules of how to multiply them, I get them, but I don't get why we need all this."

Me working on an illustration for the Transformer chapter of my new book: "Oh, look at my screen. Here's why:"

[MF FOOM @MF_FOOM]

mf trained a simple model to translate ada-002 embeddings back to text and found something interesting:

sentence embeddings have arithmetic properties

the classic "king - man + woman = queen" still works just like w/ token embeddings — but can go a few steps beyond that as well

[Jason Wei @_jasonwei]

Yann LeCun is obviously a legend but I found this tweet to be quite misinformed.

The whole point of "emergent abilities" such as few-shot prompting and chain-of-thought prompting, is that we clearly *did not* explicitly train or fine-tune them into the model. These abilities require scaling up to a lot of parameters, which is the reason we call them "emergent"---they are quantitive behavior shifts that come from qualitative changes. At least this is how researchers use the term.

All we did was train a language model with a lot of parameters on a big dataset. These abilities were not intentionally added into the model during training---rather, people found them via evaluations after the model was trained. A great pointer to this is Jacob Steinhardt's post: https://t.co/Y47ZuzkL00

What's also significant is that we often aren't able to predict these emergent phenomena ahead of time, and this has safety implications since risks can also be emergent.

Also, I don't know why finetuning is mentioned, because these abilities are clearly in the base model, which is should be apparent from papers before RLHF became popular... (e.g., PaLM-1, GPT-3 paper)

But even if you insist to take a finetuning as an example, ChatGPT RLHF was originally done in English, and it extended to many other languages. This multilingual ability was not "finetuned" into the model, which is a pretty good example of zero-shot generalization.

Please correct me if I'm wrong, but the technical use of the term definitely still holds.

Mustafa's tweet is also a bit misinformed, in my circles people don't use "emergence" to mean autonomy or agency. Again, what AI researchers usually mean is that it is a qualitative change that arises from quantitative changes.

[kirsten lum @kirsten_lum_]

This is (another reason) we should build AI systems WITH (not for) decision makers, and why retrieval augmented generation will probably be one of the most successful techniques.

I’d guess AI that “simply” surfaces past right human decisions will meet less human resistance.

[Eugene Yan @eugeneyan]

QLoRA is epic. Nonetheless, is there any reason NOT to use QLoRA?

Pros: Lower memory footprint, faster training, performs better on evals (due to implicit regularization?)

Cons: ???

[Hyung Won Chung @hwchung27]

Random life hack to fall asleep quickly.

Pick a recursion or backtracking problem that is non-trivial. Run through test cases in your head.

This will quickly max out your working memory. Your brain will beg to be shut off. And you will be in sleep.

TLDR; make your brain OOM

[Andrej Karpathy @karpathy]

E = mc^2 + AI
😂😂😂
t-shirt meme potential

[Andrej Karpathy @karpathy]

Very nice & inspiring, "no-gradient architecture" for high-level skills/learning. LLM here is the "prefrontal cortex" orchestrating the lower-level mineflayer API via code generation++.

Meta-comment is that I remember how hopeless it felt to work on agents in environments like Minecraft around ~2016, feeling stuck on how RL at the time would ever randomly explore their way into performing long-horizon tasks from super sparse rewards. This block has now to a very large extent been lifted - the correct thing was to forget all that, first train LLMs that learn (1) world knowledge, (2) reasoning and (3) tool-use (esp writing code) all from internet text, then point them back at the problem in this kind of a way. TLDR If I had read about this "no-gradient" approach to agents in 2016 my mind would certainly be blown.

Also haha @ source code in the voyager/prompts/*.txt directory :D

[Andrej Karpathy @karpathy]

There's a chance that LoRA finetunes work so well that it dramatically alters the finetuning vs. retrieval + few-shot prompting power dynamic in favor of the former for many applications.

PEFT (Parameter Efficient Finetuning, LoRA included) are emerging techniques that make it very cheap to finetune LLMs because most of the parameters can be kept frozen and in very low precision during training. The cost of pretraining and finetuning decouple.
https://t.co/vKe6xtMjad

+LoRA (the code is very short/readable)
https://t.co/S9pwKrieUs

[Ethan Mollick @emollick]

This might be the first OpenAI result that made me actually laugh out loud.

Mr. Fluffernutter's rise is indeed quite disturbing.

[Ewan St. John Smith @psalmotoxin]

Today marks a special day in the annual rhythm of the @CorpusCambridge Chronophage Clock

[John Horton @johnjhorton]

@generic_void married to non-economist - i'll field this one

[Adam Muhtar @adammuhtar_]

@elanorng https://t.co/uXo2PTzl0W instance is the most natural spot to set up, that's where the migration is headed to

Elon gutting Twitter engineers is unconscionable

[Adam Muhtar @adammuhtar_]

@elanorng Lmk when you've decided to dump tweets for toots 🐘

[Adam Posen @AdamPosen]

I want to profess my admiration for the report today by @bankofengland Dep Gov Jon Cunliffe on the real causes of the market disruption for UK pensions and gilts
Speaking truth to power is a duty for an independent central bank. Well done!

[Jonathan Mummolo @jonmummolo]

court transcripts of lawyers questioning experts about statistical concepts is a highly underrated genre

[Josh Dean @josh_t_dean]

Economists seem to want to rationalize everything except non-economist’s views about economic policy.

Instead of writing off people as being “misguided”, it seems like a useful exercise to ask what mental model of the economy/government/firms/workers justifies this policy.