“So I asked Ilya Sutskever, OpenAI’s chief scientist, for a reading list. He gave me a list of like 40 research papers and said, ‘If you really learn all of these, you’ll know 90% of what matters today.’ And I did. I plowed through all those things and it all started sorting out in my head.”
What papers do you think were on this list?
[1] https://dallasinnovates.com/exclusive-qa-john-carmacks-different-path-to-artificial-general-intelligence/
John Carmack’s ‘Different Path’ to Artificial General Intelligence - https://news.ycombinator.com/item?id=34637650 - Feb 2023 (402 comments)
RIP. If it's any consolation, it sounds like the list is at least three years old by now. Which is a long time considering that 2016 is generally regarded as the date of the deep learning revolution.
In my experience when it comes to learning technical subjects from a position of relative total ignorance, it's the older resources that are the easiest to bootstrap knowledge from. Then you basically work your way forward through the newer texts, like an accelerated replay of a domain's progress.
I think it's kind of obvious that this would be the case when you think about it. Just like how history textbooks can't keep growing in size to give all past events an equal treatment, nor can technical references as a domain matures.
You're forced to toss out stuff deemed least relevant to today, and in technical domains that's often stuff you've just started assuming as understood by the reader... where early editions of a new space would have prioritized getting the reader up to speed in something totally novel to the world.
I thought it was 2012, when AlexNet took the imagenet crown?
(Gosh, it's been less than a decade. Time sometimes doesn't fly, considering how much it's changed the world since then...)
So, slavery?
"Lena" is a bit of different case because it's not AGI. Probably ripe for the "forced prison labor" suggested by your sibling as the moral cop-out. Imagine being sentenced to being a cloud VM image!
Even breathing and drinking water are sourced by megacorps-in-charge
The context uses human-like/human-level a lot, but I agree what level and type of intelligence commands human respect is tricky business.
> forced prison labor
Would be interesting if we found ways to convince ourselves the AIs had it coming.
Generally speaking, slavery has been morally acceptable and popular before, and I will also not be surprised if we return to those ways.
The people given a second chance at life will be the ones who are quickest at identifying traffic signals or fire hydrants from a line up of images.
I would've hoped he'd be exploring weirder alternatives off the beaten path. I mean, neural networks might not even be necessary for AGI, but no one at OpenAI is going to tell Carmack that.
Otherwise you may end up walking the ditch beside the beaten path. It is slow and difficult, but it won't get you anywhere new.
For example, you may try an approach that doesn't look like deep learning, but after a lot of work, realize that you actually reinvented deep learning, poorly. We call these things neurons, transformers, backpropagation, etc... but in the end, it is just maths. If you end up finding that your "alternative" ends up being very well suited to linear algebra and gradient descent, once you have found the right formulas, you may realize that they are equivalent to the ones used in traditional "deep learning" algorithms. It help to recognize this early and take advantage of all the work done before you.
I mean, any idiot can go off-trail and start blundering around in the weeds, and ultimately wind up tripping, falling, hitting their head on a rock, and drowning to death in a ditch. But actually finding a new, better, more efficient path probably involves at least some understanding of the status quo.
Oh man, you had me going with such a vivid metaphor. I was really hoping for a payoff in the end, but you abandoned it. The easy close would be "probably involves at least some understanding of the existing terrain" but I was optimistic for something less prosaic.
It would have been interesting seeing someone like Carmack going in this direction, but from the little details he gave he seems less interested in cells and Kjeldahl flasks and more of the same type-a-type-a on the ol' QWERTY.
† 'simply' might involve multiple decades of research and Buffett knows how many billions
[1] Human neurons implanted in mice influence behavior, https://www.nature.com/articles/s41586-022-05277-w
> Their survival here is incredible enough; but even more fantastic, to me, is the fact that they have gone unnoticed during this century, until now. Lately there have been men capable of appreciating their potential value– and not only myself. What Thon Kaschler might have done with them while he was alive!– even seventy years ago."
> The sea of monks' faces was alight with smiles upon hearing so favorable a reaction to the Memorabilia from one so gifted as the thon. Paulo wondered why they failed to sense the faint undercurrent of resentment– or was it suspicion?– in the speaker's tone. "Had I known of these sources ten years ago," he was saying, "much of my work in optics would have been unnecessary." Ahha! thought the abbot, so that's it. Or at least part of it. He's finding out that some of his discoveries are only rediscoveries, and it leaves a bitter taste. But surely he must know that never during his lifetime can he be more than a recoverer of lost works; however brilliant, he can only do what others before him had done. And so it would be, inevitably, until the world became as highly developed as it had been before the Flame Deluge.
-- A Canticle for Leibowitz
Neuralink is the only place where this pattern seemed to break a bit but then seems like Elon came into his own path with trying to push for faster results and breaking basic ethics.
The denial of obviously fertile paradigm feels like such a useless self-defeating loss to indulge in an intellectual status game.
We could be all better off right now if connectionists were given DOE-grade supercomputers in the 90s, and were supplied with custom TPUs later in the 00s as their ideas were proven generally correct via rigorous experimentation on said DOE supercomputers. This didn't happen due to what amounts to academic bullying culture: https://en.wikipedia.org/wiki/Perceptrons_(book)
The sheer scale of cumulative losses we suffered (at least in part) due to this denial of the connectionism as a generally useful foundational field will be estimated somewhere in the astronomical powers of ten in the future, where the fruits of this technology will provide radically better lives for us and our descendants.
I see you have a knee-jerk reaction to hype and industry, and we are all fearing replacement unless its a stock market doing the work for us ... but why do you feel the need to punch down at this prosaic field "about nonlinear optimization"? The networks in question just want to learn, and to help us, if we train them to this end - and we make any and all excuses to avoid receiving this help, as our civilization quietly drowns in its own incompetency...
Of course, there are a group of people defending the symbolic computation, e.g. see Gary Marcus, and always pushing back on connectionism (neural networks).
But this is somewhat a spectrum, or also rather sloppy terminology. Once you go away from symbolic computation, many things can be interpret as neural network. And there is also all the computational neuroscience, which also work with some variants of neural networks.
And there is the human brain, which demonstrates, that a neural network is capable of doing AGI. So why would you not want a neural network? But that does not say that you can do many things very different from mainstream.
AlexNet (2012) VGGNet (2014) ResNet (2015) GoogleNet (2015) Transformer (2017) Reinforcement Learning:
Q-Learning (Watkins & Dayan, 1992) SARSA (R. S. Sutton & Barto, 1998) DQN (Mnih et al., 2013) A3C (Mnih et al., 2016) PPO (Schulman et al., 2017) Natural Language Processing:
Word2Vec (Mikolov et al., 2013) GLUE (Wang et al., 2018) ELMo (Peters et al., 2018) GPT (Radford et al., 2018) BERT (Devlin et al., 2019)
I'm very confident that this is pretty much what any researcher, including Ilya, would recommend. It really isn't hard to find those resources, they are simply the most cited papers. Of course you can go deeper into any of the subfields if you desire.
But AGI is one of those very ambiguous terms. For many people it's either an exact digital replica of human behavior that is alive, or something like a God. I think it should also apply to general purpose AI that can do most human tasks in a strictly guided way, although not have other characteristics of humans or animals. For that I think it can be built on advanced multimodal transformer-based architectures.
For the other stuff, it's worth giving a passing glance to the fairly extensive amount of research that has been labeled AGI over the last decade or so. It's not really mainstream except maybe the last couple of years because really forward looking people tend to be marginalized including in academia.
Looking forward, my expectation is that things like memristors or other compute-in-memory will become very popular within say 2-5 years (obviously total speculation since there are no products yet that I know of) and they will be vastly more efficient and powerful especially for AI. And there will be algorithms for general purpose AI possibly inspired by transformers or AGI research but tailored to the new particular compute-in-memory systems.
When an artificial tool that is really a point solution "tricks" us into thinking it has replicated a task that requires complex multi-component functioning within our brain, we assume the tool is acting like our brain is acting.
The joke of course being that if you maliciously edited GPT's index for translating vectors to words, it would produce gibberish and we wouldn't care (despite being the exact same core model).
We are only impressed by the complex sequence to sequence strings it makes because the tokens happen to be words (arguable the most important things in our lives).
EDIT: a great historic metaphor for this is how we thought about 'computer vision' and CNN's. They do great at identifying things in images, but notice that we still use image-based captcha's (Even on OpenAI sites no less!)?
That's because it turns out optical illusions and context-heavy images are things that CNN's really struggle at (since the problem space is bigger than 'how are these pixels arranged')
1) As I said, many people have different ideas of what we are talking about. I assume that for you general purpose AI has more capabilities, such as the ability to quickly learn tasks to a high level on the fly. For me, it still qualifies as general purpose if it can do most tasks but relies on a lot of pre-training and let's say knowledgebase look up.
2) It seems obvious to me that ChatGPT proves a general purpose utility for these types of LLMs, and it is easy to speculate that something similar but with visual input/output also will be even more general. And so we are just looking at a matter of degree by that definition.
Strikes me as the kind of thing where that last 10% will need 400 papers
-- just about everything I've shipped :-)
On models: Obviously, almost everything is Transformer nowadays (Attention is all you need paper). However, I think to get into the field, to get a good overview, you should also look a bit beyond the Transformer. E.g. RNNs/LSTMs are still a must learn, even though Transformers might be better in many tasks. And then all those memory-augmented models, e.g. Neural Turing Machine and follow-ups, are important too.
It also helps to know different architectures, such as just language models (GPT), attention-based encoder-decoder (e.g. original Transformer), but then also CTC, hybrid HMM-NN, transducers (RNN-T).
Some self-promotion: I think my Phd thesis does a good job on giving an overview on this: https://www-i6.informatik.rwth-aachen.de/publications/downlo...
Diffusion models is also another recent different kind of model.
Then, a separate topic is the training aspect. Most papers do supervised training, using cross entropy loss to the ground-truth target. However, there are many others:
There is CLIP to combine text and image modalities.
There is the whole field on unsupervised or self-supervised training methods. Language model training (next label prediction) is one example, but there are others.
And then there is the big field on reinforcement learning, which is probably also quite relevant for AGI.
Will receive Turing Award
It is being cited often
This is the weird thing - hopefully not! Hopefully there's even better NN models coming out every 5-10 years and we look back on transformers as 'just a phase' sort of like how we look back at RNN's (which were no less of an amazing achievement - look at the proliferation of voice assistants), as potentially obsolete technology today.
Fore example, attention is great and does a really good job of simulating context in language, but what if we come up with a clever way to simulate symbology? Then we actually are back on the path to AGI and transformers will look like child's play.
Aachen has quite a few companies with activity on NLP or speech recognition, mostly due to my professor Hermann Ney. E.g. there is Apple, Amazon, Nuance, eBay. And lesser-known AppTek. And in Cologne, you have DeepL. In all those companies, you find many people from our group. And then, at the RWTH Aachen University, you have our NLP/speech group, and also the computer vision group.
https://twitter.com/id_aa_carmack/status/1241219019681792010
Unlocking the Secrets of AI: A Journey through the Foundational Papers by @vrungta (2023)
1. "Attention is All You Need" (2017) - https://arxiv.org/abs/1706.03762 (Google Brain) 2. "Generative Adversarial Networks" (2014) - https://arxiv.org/abs/1406.2661 (University of Montreal) 3. "Dynamic Routing Between Capsules" (2017) - https://arxiv.org/abs/1710.09829 (Google Brain) 4. "Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks" (2016) - https://arxiv.org/abs/1511.06434 (University of Montreal) 5. "ImageNet Classification with Deep Convolutional Neural Networks" (2012) - https://papers.nips.cc/paper/4824-imagenet-classification-wi... (University of Toronto) 6. "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding" (2018) - https://arxiv.org/abs/1810.04805 (Google) 7. "RoBERTa: A Robustly Optimized BERT Pretraining Approach" (2019) - https://arxiv.org/abs/1907.11692 (Facebook AI) 8. "ELMo: Deep contextualized word representations" (2018) - https://arxiv.org/abs/1802.05365 (Allen Institute for Artificial Intelligence) 9. "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" (2019) - https://arxiv.org/abs/1901.02860 (Google AI Language) 10. "XLNet: Generalized Autoregressive Pretraining for Language Understanding" (2019) - https://arxiv.org/abs/1906.08237 (Google AI Language) 11. T5: Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer" (2020) - https://arxiv.org/abs/1910.10683 (Google Research) 12. "Language Models are Few-Shot Learners" (2021) - https://arxiv.org/abs/2005.14165 (OpenAI)
