Although I’ve been repeatedly advised it’s not a good social strategy, a glorious way to start a research paper is with specific, righteous criticism of your anonymous colleagues:For read-ability, I have dropped the citations and section references from these quotes without marking the ellipses.a
Transformers are deep feed-forward artificial neural networks with a (self)attention mechanism. They have been tremendously successful in natural language processing tasks and other domains. Since their inception 5 years ago, many variants have been suggested. Descriptions are usually graphical, verbal, partial, or incremental. Despite their popularity, it seems no pseudocode has ever been published for any variant. Contrast this to other fields of computer science, even to “cousin” discipline reinforcement learning.
So begin Phuong & Hutter in a great, rant-filled paper that “covers what Transformers are, how they are trained, what they’re used for, their key architectural components, tokenization, and a preview of practical considerations, and the most prominent models.” As an exercise, in this post I’m dig into the first item by writing down an even more compact definition of a transformer than theirs, in the form of a mathematical function rather than pseudocode, while avoiding the ambiguities rampant in the rest of the literature. I will consider only what a single forward-pass of a transformer does, considered as a map from token sequences to probability distributions over the token vocabulary. I do not try to explain the transformer, nor do I address other important aspects like motivation, training, and computational.
(This post also draws on a nice introduction by Turner. If you are interested in understanding and interpretation, you might check out — in descending order of sophistication — Elhage et al.… [continue reading]
[Summary: Constantly evolving tests for what counts as worryingly powerful AI is mostly a consequence of how hard it is to design tests that will identify the real-world power of future automated systems. I argue that Alan Turing in 1950 could not reliably distinguish a typical human from an appropriately-fine-tuned GPT-4, yet all our current automated systems cannot produce growth above historic trends.A draft of this a
What does the phenomena of “moving the goalposts” for what counts as AI tell us about AI?
It’s often said that people repeatedly revising their definition of AI, often in response to previous AI tests being passed, is evidence that people are denying/afraid of reality, and want to put their head in the sand or whatever. There’s some truth to that, but that’s a comment about humans and I think it’s overstated.
Closer to what I want to talk about is the idea AI is continuously redefined to mean “whatever humans can do that hasn’t been automated yet”, often taken to be evidence that AI is not a “natural” kind out there in the world, but rather just a category relative to current tech. There’s also truth to this, but not exactly what I’m interested in.
To me, it is startling that (I claim) we have systems today that would likely pass the Turing test if administered by Alan Turing, but that have negligible impact on a global scale. More specifically, consider fine-tuning GPT-4 to mimic a typical human who lacks encyclopedic knowledge of the contents of the internet. Suppose that it’s mimicking a human with average intelligence whose occupation has no overlap with Alan Turing’s expertise.… [continue reading]
Here’s a collection of reviews of the arguments that artificial general intelligence represents an existential risk to humanity. They vary greatly in length and style. I may update this from time to time.
Here, from my perspective, are some different true things that could be said, to contradict various false things that various different people seem to believe, about why AGI would be survivable on anything remotely resembling the current pathway, or any other pathway we can easily jump to.
This report examines what I see as the core argument for concern about existential risk from misaligned artificial intelligence. I proceed in two stages. First, I lay out a backdrop picture that informs such concern. On this picture, intelligent agency is an extremely powerful force, and creating agents much more intelligent than us is playing with fire -- especially given that if their objectives are problematic, such agents would plausibly have instrumental incentives to seek power over humans. Second, I formulate and evaluate a more specific six-premise argument that creating agents of this kind will lead to existential catastrophe by 2070. On this argument, by 2070: (1) it will become possible and financially feasible to build relevantly powerful and agentic AI systems; (2) there will be strong incentives to do so; (3) it will be much harder to build aligned (and relevantly powerful/agentic) AI systems than to build misaligned (and relevantly powerful/agentic) AI systems that are still superficially attractive to deploy; (4) some such misaligned systems will seek power over humans in high-impact ways; (5) this problem will scale to the full disempowerment of humanity; and (6) such disempowerment will constitute an existential catastrophe.
… [continue reading]
[This topic is way outside my expertise. Just thinking out loud.
Here is Google’s new language model PaLM having a think:
Alex Tabarrok writes
It seems obvious that the computer is reasoning. It certainly isn’t simply remembering. It is reasoning and at a pretty high level! To say that the computer doesn’t “understand” seems little better than a statement of religious faith or speciesism…
It’s true that AI is just a set of electronic neurons none of which “understand” but my neurons don’t understand anything either. It’s the system that understands. The Chinese room understands in any objective evaluation and the fact that it fails on some subjective impression of what it is or isn’t like to be an AI or a person is a failure of imagination not an argument…
These arguments aren’t new but Searle’s thought experiment was first posed at a time when the output from AI looked stilted, limited, mechanical. It was easy to imagine that there was a difference in kind. Now the output from AI looks fluid, general, human. It’s harder to imagine there is a difference in kind.
Tabarrok uses an illustration of Searle’s Chinese room featuring a giant look-up table:
But as Scott Aaronson has emphasized [PDF], a machine that simply maps inputs to outputs by consulting a giant look-up table should not be considered “thinking” (although it could be considered to “know”). First, such a look-up table would be beyond astronomically large for any interesting AI task and hence physically infeasible to implement in the real universe. But more importantly, the fact that something is being looked up rather than computed undermines the idea that the system understands or is reasoning.… [continue reading]
Having heard Geoffrey Hinton’s somewhat dismissive account of the contribution by physicists to machine learning in his online MOOC, it was interesting to listen to one of those physicists, Naftali Tishby, here at PI:
The surprising success of learning with deep neural networks poses two fundamental challenges: understanding why these networks work so well and what this success tells us about the nature of intelligence and our biological brain. Our recent Information Theory of Deep Learning shows that large deep networks achieve the optimal tradeoff between training size and accuracy, and that this optimality is achieved through the noise in the learning process.
In this talk, I will focus on the statistical physics aspects of our theory and the interaction between the stochastic dynamics of the training algorithm (Stochastic Gradient Descent) and the phase structure of the Information Bottleneck problem. Specifically, I will describe the connections between the phase transition and the final location and representation of the hidden layers, and the role of these phase transitions in determining the weights of the network.
Based partly on joint works with Ravid Shwartz-Ziv, Noga Zaslavsky, and Shlomi Agmon.
(See also Steve Hsu’s discussion of a similar talk Tishby gave in Berlin, plus other notes on history.)
I was familiar with the general concept of over-fitting, but I hadn’t realized you could talk about it quantitatively by looking at the mutual information between the output of a network and all the information in the training data that isn’t the target label.… [continue reading]
A lot of people sound worried that new and improving techniques for creating very convincing videos of anyone saying and doing anything will lead to widespread misinformation and even a break down of trust in society.
I’m not very worried. Two hundred years ago, essentially all communication, other than in-person conversation, was done through written word, which is easy to fake and impersonate. In particular, legal contracts were (and are) typeset, and so are trivially fakeable. But although there were (and are) cases of fraud and deception through foraged documents, society has straightforward mechanisms for correctly attributing such communication to individuals. Note, for instance, that multi-billion-dollar contracts between companies are written in text, and we have never felt it necessary to record difficult-to-fake videos of the CEOs reciting them.
The 20th century was roughly a technological Goldilocks period where technology existed to capture images and video but not to fake them. Images, of course, have been fakeable at modest cost for many years. Even in 1992, Michael Crichton’s Rising Sun used high-tech fraudulent security footage as a realistic plot point in then-current day. Although we may see some transition costs as folks are tricked into believing fraudulent videos because the ease of faking them has not yet entered the conventional wisdom, eventually people will learn that video can’t be trusted much more than the written word.Which is to say, most of the time you can trust both text and video because most people aren’t trying to defraud you, but extra confirmatory steps are taken for important cases.a This will not be catastrophic because our trust networks are not critically dependent on faithful videos and images.… [continue reading]