I recently demonstrated GPT4 to my spouse's 101-year-old grandfather, who remains in excellent health and has a sharp mind.
Following my demonstration, he paused thoughtfully and then said something I will remember — “This technology instills hope for our future. It's high time
Stanfords DSPy is the best high level LLM programing framework I have seen this far.
Langchain never resonated with me; despite being an early LLM framework, its design and abstractions felt overly complex. DSPy, on the other hand, is a huge step in the right direction.
DSPy
An interesting new Nature paper compares fMRI recordings with activations across layers in a language model, and find evidence of correlations.
The study seems to suggests that brain regions located at the top of the language hierarchy, responsible for
Six years ago, Geoffrey Hinton asserted that AI would take over radiology within five years, suggesting we cease training radiologists.
Was he correct?
The situation is more complex than simply being right or wrong. While AI has surpassed radiologists in certain diagnostic
Deep learning is typically bottlenecked by memory not compute
⚡️Flash Attention ⚡️ optimizes transformers, like GPT, to minimize costly GPU memory fetches and achieves impressive speedups of 2-4x, 5-20x less memory intensive, and enables scaling to longer
Self-consistency is underrated for improving accuracy for LLMs in a range of reasoning and arithmetic tasks.
It works with any off-the-shelf LLM, eg GPT3 variants, and also provides estimates of how certain the LLM is of the provided answer.
Takeaways👇
A simple trick to make LLMs “calibrated” — ie “to know when it doesn’t know something” — is to reformulate the answers as a single word or a short phrase, and look at the predicted logprobs of the word.
As LLMs are trained to predict the probability of the next token, they are
🤖️LLM can self-improve 🧠
1) Self-consistency boosts reasoning skills by sampling multiple paths & finding the most consistent answer
But more samples = more comp. requirements. 💻
2) but we can train better LLM with self-generated solutions from 1)
What it you had trained a model to play legal moves in Othello by predicting the next move, and found that it had spontaneously learned to compute/represent the full board state in it's weights - an emergent world representation?
That's just what this
Insightful paper that succinctly covers essential high-level knowledge to keep in mind regarding LLMs:
- Large language models (LLMs) predictably improve with increasing investment, but many key behaviors emerge unpredictably.
- LLMs often learn and use representations of the
✨Neat LLM trick for 📈 math & logical abilities ✨
Improves on Chain of Thought (CoT) prompting by
1) Replace natural language, step by step instructions, in CoT examples with commented, stepwise, python code.
2) Run the code
Several recent papers on this (see refs below⬇️)
Want to know a simple trick for LLMs to generate more plausible long documents, breaks out of repetition better, and more reasonably truncate low probability tokens?
Learn about LLM truncation sampling!
Some takeaways from 👇🧵
LLMs suffer from overconfidence and poorly calibrated uncertainty estimates
However, self-consistency, where on samples multiple paths & finds the most consistent answer, seems to offer a practical solution.
Interesting figure from page 4 in "LLMs can self improve" paper
A fun trick for zero shot retrieval tasks with great results! First use a off the shelf LLM to generate a set of hypothetical candidate document, then use standard embedding model + standard search to find best matching documents in DB/Web.
Details 👇
Distill step by step!
A new research paper from Google presents a straightforward concept that let’s them train a 770M T5 model that surpasses the 540B PaLM model, using just 80% of the available data on a benchmark task.
Essentially distills (trains) a smaller model from the
Bad future: A single AI, RLHFd into “alignment” to a very narrow set of values determined by a very small set of people.
Good future: Democratized multiple AIs that is reasonably regulated, inducing diversity of thought, applied towards medicine, science and wisdom. A society
@BillAckman
@PeterHotez
Mr. Ackman,
@BillAckman
, I appreciate your lengthier posts and believe you can be a thoughtful individual. However, in this specific instance of sharing out-of-context clips with voiceovers, it appears to be an incredibly ineffective method of uncovering the truth.
While I don't
The benefits of AGI are often associated with accomplishments such as "curing cancer" or other medical breakthroughs. However, it appears that relativity few people are actively working on AI specifically for medical applications.
Instead, researchers in leading labs seem to be
It's intriguing to observe how several alt accounts, like
@tszzl
,
@BasedBeffJezos
,
@AISafetyMemes
etc, gain traction and influence a significant portion of prominent figures in the AI/tech industry, often shaping the direction of discussions.
Shapers of collective consciousness.
Isn't it quite mind-boggling that the majority of humanity's collective thoughts and reasoning, in broad strokes, seem to be compressible down to just a few hundred gigabytes?
@BillAckman
Are you helping by posting this compilation of out of context clips from various interviews?
why is it bad that scientists update their belief and advice when facts come in?
I don’t know much about this particular person, but the personal attacks on him for trying to navigate,
Wonderful short survey of Graph Neural Networks (GNN).
Three types of principal tasks - classification of Nodes, Graphs and Link prediction.
Deep sets and Transformers as GNNs, geometric graphs and more!
@karpathy
I remember taking my first graduate level machine learning course back in 2009 — and I got completely obsessed.
Bishops book on ML was my bible for a time, still good book!
Re-reading a few chapters from my favorite ML/stats book! Beautifully written, peppered with deep insights, and dosent shy away from the math, but doesn’t complicate things unnecessarily
Also you can get a free pdf here!
Hard to predict exactly when, but seems likely text2video with stable diffusion like quality will happen sometime in the next 3 years. Could be in 5 months could take a bit longer - but it’s likely going to happen relatively soon. Lets make sure our defences vs misinformation are
In the paper they showed a 2 layer network is needed for a non-linear probe to extract, and modify, the board state.
But in brilliant follow up work, seems to indicate you can actually just use a linear probe, also great read!
ChatGPT + stable diffusion is a pretty great condensed representation of humanity.
Good candidate to send on the next deep space voyager probe as a greeting to any alien races out there.
Aliens will learn that we are very self confident and often wrong!
Uses simple but clever tricks like blocking/tiling and cuda kernel fusion. Also recomputes the attention matrix dynamically in the backward pass instead of fetching it from memory.
Beautiful example of impressive gains from clever engineering.
What’s the best paper investigating the effect of the order of training data fed to LLM during training?
Like keep only high quality content for later in training?
Obv works for finetuning. But looking for a more generalized form.
From the GPT4 'paper' there is an interesting figure on how the base model is initially well calibrated on MMLU, but then after RLHF becomes much less so.
Does anyone know of more studies of how RLHF affects model calibration on various tasks?
@natfriedman
Harder math&science problems. Eg I find GPT4 reliabily correctly solves exercises, with no public answers, in most of my graduate textbooks in math, physics, ml etc. it’s astounding if you actually try it.
@ylecun
6 months is basically nothing in the grand scheme of things. It’s rather irrelevant for technological progress. Seems reasonable to let the public catch up before people like you decide what should be done without consulting them first.
From the GPT4 'paper' there is an interesting figure on how the base model is initially well calibrated on MMLU, but then after RLHF becomes much less so.
Does anyone know of more studies of how RLHF affects model calibration on various tasks?
@naval
Extremely risky bet from Google.
That statement dosent even incorporate what will happen to cost and capability of AI models in 1,2,3 years etc. traditional search won’t change that much.
Google has an extremely difficult innnvenitors dilemma to navigate. In an organization
@jackythirdy
Someone who pays 10$/month for Copilot?
Also vscode ships with a dark theme by default so it’s democratizing the mythical 10x engineer :-)
“It seems probable that once the machine thinking method had started, it would not take long to outstrip our feeble powers… They would be able to converse with each other to sharpen their wits. At some stage therefore, we should have to expect the machines to take control.”
-
The first person to use the concept of a "singularity" in the technological context was John von Neumann. Stanislaw Ulam reports a 1958 discussion with von Neumann "centered on the accelerating progress of technology and changes in the mode of human life, which gives the
Great example of democratizing AI!
A Stanford Alpaca type LLM tuned for Italian instruction following.
Go make one for the preferred language of your choice!
I'm excited to introduce you Camoscio: an Italian instruction-tuned LLaMA, following Stanford Alpaca.
The model should provide output of similar quality to GPT text-davinci-003 and has been finetuned by translating the Alpaca dataset to Italian.
1/3