michael tetelman retweeted
The most advanced AI systems will likely exist first within the companies that developed them, with little opportunity to collectively prepare for their impacts. This dynamic creates unique risks highlighted in this recent paper by @apolloresearch and should become a growing priority for AI policy.
michael tetelman retweeted
What if you could make physics diagrams come alive? At #UIST2024, we will be presenting our paper, Augmented Physics, an ML-Integrated Authoring Tool for Creating Interactive Physics Simulations from Static Diagrams
Co-authors: @Freya_Wyy @nandizhang_ Jarin @rubaiat @ryosuzk
michael tetelman retweeted
BREAKING NEWS
The Royal Swedish Academy of Sciences has decided to award the 2024 #NobelPrize in Physics to John J. Hopfield and Geoffrey E. Hinton “for foundational discoveries and inventions that enable machine learning with artificial neural networks.”
michael tetelman retweeted
If you’d like to spend a few hours this weekend to dive into Large Language Models (LLMs) and understand how they work, I've prepared a 3-hour coding workshop presentation on implementing, training, and using LLMs: youtube.com/watch?v=quh7z1q7…
michael tetelman retweeted
🔥Phi-3 Vision Model is out!
🏛️Architecture: Phi-3-Vision-128K-Instruct has 4.2B parameters and contains image encoder, connector, projector, and Phi-3 Mini language model.
📝Inputs: Text and Image. Best suited for prompts using the chat format.
🔢Context length: 128K!🤯 More 👇
michael tetelman retweeted
Replying to @Grady_Booch
I think Geoffrey Hinton is right and understanding that requires to look deeper to what he says:
1. predicting next symbol accurately only possible by learning and understanding a context for that symbol - that is where reasoning happens; 2/
michael tetelman retweeted
Nice new read on tokenization!
You've heard about the SolidGoldMagikarp token, which breaks GPT-2 because it was present in the training set of the Tokenizer, but not the LLM later.
This paper digs in in a lot more depth and detail, on a lot more models, discovering a less extreme version of the above - partially-trained tokens in both open/closed models. You have to be careful with a lot of small details and implications - weight sharing, constants in residual streams, weight-decays, regex splitting patterns, BPE, UTF-8, etc.
TLDR Tokenization remains a major pain and a large LLM attack surface. Including these partially-trained tokens in your prompts drifts the model out of distribution into undefined regions of the dynamics, areas that the model is not used to. They confuse the LLM. The paper's focus is discovery and not engineering, but it seems likely one can find "token attacks" that reliably induce target weirdness: pop-off safety, alter personality or behaviors (?), any other kind of ... otherwise undefined behavior, whatever that may look like.
Now go ask GPT-4 about _ForCanBeConverted, $PostalCodesNL, useRalative, and _typingsJapgolly :)
(or see Figure 4 of the paper at the very end for simple examples)
Our paper about reliably finding under-trained or 'glitch' tokens is out!
We find up to thousands of these tokens in some #LLMs, and give examples for most popular models.
arxiv.org/abs/2405.05417
More in 🧵
michael tetelman retweeted
It is with great sadness that the Simons Foundation announces the death of its co-founder and chair emeritus, James Harris Simons. Jim was an award-winning mathematician, a legendary investor and a generous philanthropist. simonsfoundation.org/2024/05…
michael tetelman retweeted
If you missed my talk on 3D Human Foundation Agents at Stanford yesterday and want to check it out, here's the recording: acorn.stanford.edu/SCIEN/vid…
michael tetelman retweeted
Replying to @DimitrisPapail @AkiyoshiKitaoka
Another serious discussion with MetaAI shows how little it's actually knowing and understanding on some random subjects:
michael tetelman retweeted
Replying to @realSharonZhou
I think even a stronger statement could be made: all models of the same "class" asymptotically converge to the same "effective" model, that for a given input produce same output. That is similar to universality of certain phase transitions in Physics,when only dimensions matters.
michael tetelman retweeted
Era of humanoids is coming, but not always easy to afford one or run real world experiments. Luckily there are also many many research challenges that can be investigated in simulation, join the fun, try to get your AI to do well on #HumanoidBench!
michael tetelman retweeted
Do LLMs simply memorize and parrot their pretraining data or do they learn patterns that generalize? Let’s put this to the test! We compute the first generalization guarantees for LLMs.
w/ @LotfiSanae, @m_finzi, @KuangYilun, @timrudner, @andrewgwils
arxiv.org/abs/2312.17173
1/9
michael tetelman retweeted
Self-attention is at the core of transformer-based LLMs. It comes in several flavors. In my new article (magazine.sebastianraschka.co…) I cover the various variants and code them in PyTorch from the ground up:
1. Self-attention: The fundamental building block of transformer-based LLMs that allows models to weigh the importance of different parts of the input data.
2. Multi-head attention: An extension of self-attention that allows the model to simultaneously focus on information from different representation subspaces.
3. Cross-attention: A variant that enables the model to attend to two different sequences, which makes it useful in tasks like translation or summarization.
4. Causal-attention: A variant to ensure that the prediction for each token depends only on the preceding tokens, which is important for text generation, where each prediction should be based only on the prior context.
In my opinion, coding algorithms, models, and techniques from scratch is not only fun but an excellent way to learn!
michael tetelman retweeted
Does a language model trained on “A is B” generalize to “B is A”?
E.g. When trained only on “George Washington was the first US president”, can models automatically answer “Who was the first US president?”
Our new paper shows they cannot!
michael tetelman retweeted
Here's my conversation with @WalterIsaacson, author of the new biography on @elonmusk and one of the greatest biographers ever, having written incredible books on Einstein, Steve Jobs, Leonardo da Vinci, Jennifer Doudna, Benjamin Franklin, and many others. Outline of our conversation:
0:00 - Introduction
3:00 - Difficult childhood
20:04 - Jennifer Doudna
23:01 - Einstein
28:20 - Tesla
45:24 - Elon Musk's humor
49:34 - Steve Jobs' cruelty
52:58 - Twitter
1:05:07 - Firing
1:07:52 - Hiring
1:16:55 - Time management
1:24:39 - Groups vs individuals
1:28:25 - Mortality
1:31:57 - How to write
1:52:56 - Love & relationships
1:57:50 - Advice for young people
michael tetelman retweeted
Textbooks Are All You Need II: phi-1.5 technical report
Builds a new 1.3B model named phi-1.5, with performance comparable to models 5x larger, and surpassing most non-frontier LLMs on tasks like GSM8k and HumanEval.
arxiv.org/abs/2309.05463
michael tetelman retweeted
LLMs as Optimizers
This is a really neat idea. This new paper from Google DeepMind proposes an approach where the optimization problem is described in natural language.
An LLM is then instructed to iteratively generate new solutions based on the defined problem and previously found solutions.
It was first tested on linear regression and the traveling salesman problem. Leveraging LLMs with simple prompting match or surpass hand-designed heuristic algorithms. This shows good potential for using LLMs as optimizers.
The idea is then applied to prompt optimization that aims to maximize task accuracy on different tasks like math word problem-solving.
The first piece of the proposed meta-prompt takes in previously generated prompts along with corresponding training accuracies. The second piece includes the optimization problem description with samples obtained from a training set representing the task.
At each optimization step, the goal is to generate new prompts that increase test accuracy based on the trajectory of previously generated prompts.
The optimized prompts outperform human-designed prompts on GSM8K and Big-Bench Hard, sometimes by over 50%!
For math word problem solving, one of the most effective instructions found begins with "Take a deep breath and work on this problem step-by-step".
arxiv.org/abs/2309.03409
michael tetelman retweeted
New blog post about the geometry of diffusion guidance: sander.ai/2023/08/28/geometr…
This complements my previous blog post on the topic of guidance, but it has a lot of diagrams which I was too lazy to draw back then! Guest-starring Bundle, the cutest bunny in ML 🐇
michael tetelman retweeted
Yoshua Bengio:
'For most of these years, I did not think about the dual-use nature of science because our research results seemed so far from human capabilities and the work was only academic. It was a pure pursuit of knowledge, beautiful, but mostly detached from society until about a decade ago. I now believe that I was wrong and short-sighted to ignore that dual-use nature. I also think I was not paying enough attention to the possibility of losing control to superhuman AIs.'
'[...] it started to dawn on me that my previous estimates of when human-level AI would be reached needed to be radically changed. Instead of decades to centuries, I now see it as 5 to 20 years with 90% confidence.
And what if it was, indeed, just a few years?'
'I started reading more about AI safety and came to a critically important conclusion: we do not yet know how to make an AI agent controllable and thus guarantee the safety of humanity! And yet we are – myself included until now – racing ahead towards building such systems.'
'It is painful to face the idea that we may have been contributing to something that could be greatly destructive. Human nature will lead us towards brushing aside these thoughts or finding comfort in reassuring arguments rather than face the full horror of such possibilities. Bringing the benefits of AI to the table is not sufficient to compensate if the possible negative outcomes include catastrophic misuses of AI on par with nuclear war and pandemics, or even existential risk.'
'Unfortunately, to settle the AI debate, we can’t rely on mathematical models of how research, technology and politics are likely to evolve in the next decade under different interventions. We don’t have past experience interacting with machines more intellectually capable than us and thus obtain statistics of what is safe and what isn’t. Yet we cannot wait until irreversible damage is done in order to change course. AI researchers are used to easily performing many experiments, including controlled experiments, and statistical assessments before drawing conclusions. Here we instead have to resort to a form of reasoning and out-of-distribution projection that is closer to how many of our colleagues in the social sciences work. It makes it harder and more uncertain to evaluate possible futures. However, reason and compassion can still be used to guide our conversations and actions. As scientists, we should avoid making claims we can’t support; but as decision-makers we also ought to act under uncertainty to take precautions. In spite of our differences in points of view, it’s time for our field of AI to seriously discuss the questions: what if we succeed? What if potentially dangerous superhuman AI capabilities are developed sooner than expected? Let’s embrace these challenges and our differences, while being mindful of each other’s humanity and our unique emotional and psychological journeys in this new era of AI.' yoshuabengio.org/2023/08/12/…






























