dmm | Entries tagged with understanding internals of ai

I think humans tend to have these idée fixes:

www.anthropic.com/news/golden-gate-claude

A pretty interesting example of conversation by a very experienced interlocutor:

www.lesswrong.com/posts/cxuzALcmucCndYv4a/daniel-kokotajlo-s-shortform#hyXAiafwbwfEPiX5Q

Lean theorem prover seems to be the most convenient and practical (good enough for mathematicians to use it in real work, I am seeing plenty of such stories lately; there is a bit of controversy on whether to use Lean 3 or Lean 4).

The best quote according to "leanprover" twitter:

>Our favorite quote from the paper: "Monadic syntax is excellent for expressing stochastic algorithms, and working over finitely supported distributions avoids the need for integrability side conditions during proofs."

Links are in the comments.

A good way to mark this occasion is to try to read a new paper which seems to be a major breakthrough in understanding and harnessing the magic of Transformers:

"Uncovering mesa-optimization algorithms in Transformers"

"we demonstrate that minimizing a generic autoregressive loss gives rise to a subsidiary gradient-based optimization algorithm running inside the forward pass of a Transformer. This phenomenon has been recently termed mesa-optimization"

"Moreover, we find that the resulting mesa-optimization algorithms exhibit in-context few-shot learning capabilities,
independently of model scale. Our results therefore complement previous reports characterizing the
emergence of few-shot learning in large-scale LLMs"

This is a good starting point:

"A Mathematical Framework for Transformer Circuits", Dec 2021
transformer-circuits.pub/2021/framework/index.html

I am reading more and more LessWrong in recent months (mostly, after the Simulator theory by Janus (work done while at Conjecture) has been posted there in September).

I still think the Simulator theory is probably the single most important research breakthrough of 2022.

These days LessWrong is dominated by writing related to AI safety (the topic is made particularly acute by the recent progress in LLMs: ChatGPT and even more capable Bing Chat; no consensus whatsoever, of course, but I do think that GPT-3 release in May 2020 is, in some sense, an equivalent of the nuclear fission discovery on 19 December 1938, and that ChatGPT performance (+ Bing Chat clearly drastically enhanced capabilities even compared to that) is, in the same sense, an equivalent of the first working nuclear reactor on 2 December 1942, if one goes by "AI today is what nuclear energy has been back then" analogy).

So, one thing which might be useful is that there is GreaterWrong alternative viewer (which looks different from LessWrong default viewer and which can be visually tuned in terms of presentation style; also different default front page for the site if one uses GreaterWrong). Which viewer is better might depend on your device (display, browser, etc).

Another thing, Conjecture people tend to produce some of the best, most interesting articles there.

I'll put a few links into the comments.

Эти системы - это не то, что традиционно было принято понимать под AI.

Это системы нового типа - симуляторы. Система "создаёт виртуальную реальность с персонажами", и мы общаемся не с самой системой, а с этими персонажами.

Эти персонажи могут иметь очень разный стиль и очень разные способности: они могут быть умными, глупыми, правдивыми, лживыми, быть похожими на хороших или плохих профессионалов той или иной профессии, и так далее.

Искусство "prompt engineering" как раз и состоит в том, чтобы создать интересных персонажей-собеседников, с теми качествами, которые нам бы хотелось, чтобы они имели. "Prompt engineering" - это очень нетривиальная быстро развиваюшаяся область. Со временем, будут разные завертки, где будет "невидимая пользователю part of the promt", настраивающая систему тем, или иным образом. Но сейчас надо творчески экспериментировать, чтобы получалось по-настоящему интересно.

То, что создано, это не AI в традиционном смысле, а скорее очень хорошо обученная "ткань искусственного мозга, не заполненная изначально никакими личностями и никакими фиксированными свойствами характера"; и искусство состоит в том, чтобы в ней возникали интересные/желанные нам персонажи и динамики.

~~P.S. I am having a local ongoing intermittent internet outage which sucks (replies might be slow).~~

www.lesswrong.com/posts/r2yTwkGt3kbQG2mXi/axrp-episode-19-mechanistic-interpretability-with-neel-nanda

To read my https://twitter.com/home more regularly (that's absolutely the best source of info at the moment).

A small fraction of today's catch:

New work by Janus

A new involved take on AI safety/alignment

(What's the right way to organize all that information?)

Links are in the comments (I think the new work by Janus is more important even for alignment, and is just overall more important of the two topics of this post)...

I have been looking at a recent rather remarkable paper which includes the DeepDream creator among its authors, and I've decided to check whether I missed any of his works; and I turns out there is this paper I really should be aware of. This really resonates with some of the thing I have been exploring this year.

arxiv.org/abs/2007.00970

"We present a novel method for learning the weights of an artificial neural network - a Message Passing Learning Protocol (MPLP). In MPLP, we abstract every operations occurring in ANNs as independent agents. Each agent is responsible for ingesting incoming multidimensional messages from other agents, updating its internal state, and generating multidimensional messages to be passed on to neighbouring agents. We demonstrate the viability of MPLP as opposed to traditional gradient-based approaches on simple feed-forward neural networks, and present a framework capable of generalizing to non-traditional neural network architectures. MPLP is meta learned using end-to-end gradient-based meta-optimisation. We further discuss the observed properties of MPLP and hypothesize its applicability on various fields of deep learning."

This week, Nov 17-18, Thu-Fri, 8am-11:45am Boston time, "Quantum physics and the first-person perspective": www.essentiafoundation.org/quantum-physics-and-the-first-person-perspective/seeing/

JuliaCon 2023, juliacon.org/2023/ the call for proposals is posted, deadline Dec 18: pretalx.com/juliacon2023/cfp

I've spent more quality time focusing of two breakthroughs in understanding the nature and the behavior of machine learning models which came from the "penumbra" of "prosaic alignment" start-ups and which I wrote about in my previous two posts.

"Grokking is (more or less) solved." I took brief notes between Oct 21 and Oct 23: github.com/anhinga/2022-notes/tree/main/Grokking-is-solved

"Generative autoregressive models are similators." I took extensive notes between Oct 5 and Oct 23: github.com/anhinga/2022-notes/tree/main/Generative-autoregressive-models-are-similators

I am continuing to develop thoughts related to these topics, I am going to gradually write more about those topics in the comments.

The most interesting conceptual AI advances seem lately to come from "prosaic alignment" start-ups. These are companies which believe that the current trend of improving Transformer models is likely to lead straight to AGI, and that better understanding of the nature and properties of these model is key to AI safety (and, of course, it's also key to better AI capabilities).

And it is often the case that the key elements of work are done by people "on the edge", "in the penumbra" of those alignment start-ups.

In the previous post I mentioned the key new understanding of large Transformer models as simulators. That work has been done "while at Conjecture", but is not listed as directly coming from Conjecture (one of those "prosaic alignment" start-ups). I think the key people involved are still at Conjecture, but they seem to be trying to keep some distance between Conjecture and this work. I am continuing to take notes of those materials and commit them to GitHub (see links in the comments to the previous post).

Here is another one of those stories. Grokking is a phenomenon, where small Transformers look at a part of a mathematical structure for quite a while, and then rather suddenly transition to understanding the whole of that mathematical structure including the part they never see in training. It has been discovered in 2021 and has been a subject of a number of follow-up attempts to understand it.

The recent breakthrough has been done in mid-August by Neel Nanda who left Anthropic (perhaps the most famous of the "prosaic alignment" start-ups) a few months ago. And it looks like he has more or less solved the mysteries behind this phenomenon. I am going to continue studying his writings more. The links are in the comments.

Вот, наконец, кажется возник правильный подход к пониманию природы моделей вроде GPT-3 и разнообразного волшебства, с этим связанного:

www.lesswrong.com/posts/vJFdjigzmcXMhNTsx/simulators

Он говорит, что надо перестать думать про эти модели в терминах более старых AI-систем.

openai.com/blog/our-approach-to-alignment-research/

"Our approach to aligning AGI is empirical and iterative. We are improving our AI systems’ ability to learn from human feedback and to assist humans at evaluating AI. Our goal is to build a sufficiently aligned AI system that can help us solve all other alignment problems."

github.com/salesforce/CodeGen

One can also run one of these models via HuggingFace; it is based on "A Conversational Paradigm for Program Synthesis" paper, arxiv.org/abs/2203.13474

Someone has even created a fake GitHub Copilot based on that (useful for those who prefer VSCode): github.com/moyix/fauxpilot

Anthropic AI is an organization which has been created approximately a year ago by former OpenAI people who (I believe) have been unhappy about the current direction of OpenAI.

"Anthropic is an AI safety and research company that’s working to build reliable, interpretable, and steerable AI systems. Large, general systems of today can have significant benefits, but can also be unpredictable, unreliable, and opaque: our goal is to make progress on these issues."

They have just published their first major paper directed towards better understanding of Transformers. I am going to accumulate various links in the comments.