In the NYT today, Cade Metz implies that I left Google so that I could criticize Google. Actually, I left so that I could talk about the dangers of AI without considering how this impacts Google. Google has acted very responsibly.
Dishonest CBC headline:
"Canada's AI pioneer Geoffrey Hinton says AI could wipe out humans. In the meantime, there's money to be made".
The second sentence was said by a journalist, not me, but you wouldn't know that.
I thought I had a very good idea about perceptual learning and accepted several invitations to give talks about it next week. But I have just discovered a fatal flaw in the idea, so I am cancelling all those talks. I apologize.
Suppose you have cancer and you have to choose between a black box AI surgeon that cannot explain how it works but has a 90% cure rate and a human surgeon with an 80% cure rate. Do you want the AI surgeon to be illegal?
Andrew Ng is claiming that the idea that AI could make us extinct is a big-tech conspiracy. A datapoint that does not fit this conspiracy theory is that I left Google so that I could speak freely about the existential threat.
Yann LeCun thinks the risk of AI taking over is miniscule. This means he puts a big weight on his own opinion and a miniscule weight on the opinions of many other equally qualified experts.
Extrapolating the spectacular performance of GPT3 into the future suggests that the answer to life, the universe and everything is just 4.398 trillion parameters.
The X factor: When I was an undergrad at Kings College Cambridge, Les Valiant who won the Turing award in 2010 lived in the adjacent room on X staircase. He just told me that Turing lived on X staircase when he was a fellow at Kings and probably wrote his 1936 paper there!
Caterpillars extract nutrients which are then converted into butterflies. People have extracted billions of nuggets of understanding and GPT-4 is humanity's butterfly.
New paper:
Companies are planning to train models with 100x more computation than today’s state of the art, within 18 months. No one knows how powerful they will be. And there’s essentially no regulation on what they’ll be able to do with these models.
@JeffDean
@ylecun
@TheOfficialACM
Thanks to my graduate students and postdocs whose work won a Turing award. Thanks to my visionary mentors Inman Harvey, David Rumelhart and Terry Sejnowski. And thanks to Jeff Dean for creating the brain team that turns basic research in neural nets into game-changing products.
Fei-Fei Li has written a book. She was the first computer vision researcher to truly understand the power of big data and her work opened the floodgates for deep learning. She delivers a clear-eyed account of the awesome potential and danger of AI.
My Coursera MOOC "Neural Networks for Machine Learning" was prepared in 2012 and is now seriously out of date so I have asked them to discontinue the course. But the lectures are still a good introduction to many of the basic ideas and are available at
I suspect that Andrew Ng and Yann LeCun have missed the main reason why the big companies want regulations. Years ago the founder of a self-driving company told me that he liked safety regulations because if you satisfied them it reduced your legal liability for accidents.
We recently developed a new, unsupervised version of capsule networks (with
@sabour_sara
,
@yeewhye
, and
@geoffreyhinton
). I hope that my new blog will make it easier to understand some ideas that led to this work. Enjoy!
Introducing SimCLR: a Simple framework for Contrastive Learning of Representations. SimCLR advances previous SOTA in self-supervised and semi-supervised learning on ImageNet by 7-10% (see next).
Joint work with
@skornblith
@mo_norouzi
@geoffreyhinton
.
@pmddomingos
and for a long time, most people thought the earth was flat. If we did make something MUCH smarter than us, what is your plan for making sure it doesn't manipulate us into giving it control?
Canadian politicians are talking about setting up a Canadian DARPA. Tying precious research dollars to military applications is not an efficient way to innovate and will lead to obscenities like the self-healing minefield. Encourage start-ups, not the military industrial complex.
Since leaving Google, I have received many requests to join the advisory boards of start-ups and, until now, I have declined them all. However, I have decided to join as an advisor for
@VayuRobotics
to work with
@nitishsr
again. (1/3)
The University of Toronto made a video for a non technical audience in which I explain how deep learning works, the enormous promise of this technology and some of the potential risks.
Shooting someone on Fifth Avenue and getting away with it seems like a minor indiscretion compared with killing thousands of New Yorkers by refusing to order the manufacture of ventilators.
@pmddomingos
is that the whole story? Isn't it also relevant that Bill Gates has given away a lot of his wealth whereas Steve Ballmer would rather own a basketball team than fix malaria.
I have one N95 mask. After using, I put it in a plastic bag, wash my hands and bake at 170F for 2 hours. My guess is that a half-baked mask is better than none. I would love to know whether baking at 170F reliably kills COVID-19 and whether it degrades the filtering by the mask.
A common argument against taking inspiration from the brain when designing neural networks is that it's like taking inspiration from feathers when designing flying machines. Drones need blades that will not damage things they hit and can be easily repaired with a quick preen.
The Google Brain team in Toronto has openings for several research scientists who have already made exceptional contributions to research on deep learning or its applications in NLP, vision, or reinforcement learning. To apply, go to
@AndrewYNg
I used to talk to Andrew a lot and it was great to catch up again and get his take on the various risks posed by recent developments in AI. We agreed on a lot of things, especially on the need for the researchers to arrive at a consensus view of the risks to inform policy makers.
There is a UK company called Profit Crunch Ltd that is fraudulent. They claim i am a director and use my reputation to reassure investors. They are crooks. I have nothing to do with them. They also forged my signature on an insurance certificate.
This is a really neat way of using neural networks to get rid of polygonal meshes. The interaction of computer graphics with neural networks is really exciting.
Finding the natural parts of an object and their intrinsic coordinate frames without supervision is a crucial step in learning to parse images into part-whole hierarchies. If we start with point clouds, we can do it!
📢📢📢 𝐒𝐮𝐧 𝐞𝐭 𝐚𝐥. "𝐂𝐚𝐧𝐨𝐧𝐢𝐜𝐚𝐥 𝐂𝐚𝐩𝐬𝐮𝐥𝐞𝐬: 𝐔𝐧𝐬𝐮𝐩𝐞𝐫𝐯𝐢𝐬𝐞𝐝 𝐂𝐚𝐩𝐬𝐮𝐥𝐞𝐬 𝐢𝐧 𝐂𝐚𝐧𝐨𝐧𝐢𝐜𝐚𝐥 𝐏𝐨𝐬𝐞"
A network design that realizes the concept of "𝐦𝐞𝐧𝐭𝐚𝐥 𝐩𝐢𝐜𝐭𝐮𝐫𝐞" for unsupervised 3D deep learning.
📢📢📢 Introducing 𝐍𝐞𝐮𝐫𝐚𝐥 𝐃𝐞𝐬𝐜𝐫𝐢𝐩𝐭𝐨𝐫 𝐅𝐢𝐞𝐥𝐝𝐬 (𝐍𝐃𝐅)
That's right, we 𝐭𝐞𝐚𝐜𝐡 𝐚 𝐫𝐨𝐛𝐨𝐭 to manipulate unseen objects, and unseen poses from 𝐣𝐮𝐬𝐭 𝟏𝟎 𝐞𝐱𝐚𝐦𝐩𝐥𝐞𝐬 🤯
Wanna know more? See this thread
One evening this summer, my partner and I saw the Loch Ness monster swimming in Lake Huron. It was at least 10 feet long and swam in an undulating fashion with big black humps that were each several feet long and went up and down as it moved across the lake.
Organizers of data science and machine learning conferences (NeurIPS, ICML, AISTATS, ICLR, UAI, ...): Allow remote paper & poster presentations at conferences - Sign the Petition! via
@Change
@ylecun
Let's open source nuclear weapons too to make them safer. The good guys (us) will always have bigger ones than the bad guys (them) so it should all be OK.
We've just released the first version of our Deep Learning Tuning Playbook! This is our attempt to distill our process for actually getting good results with deep learning. We emphasize hyperparameter tuning since it has been a large pain point.
@sonicshifts
I now predict 5 to 20 years but without much confidence. We live in very uncertain times. It's possible that I am totally wrong about digital intelligence overtaking us. Nobody really knows which is why we should worry now.
@pmddomingos
There is so much wrong with that argument that I do not know where to start. Maybe just look at how smart AI was 10 years ago and how smart it is now. Then project that 10 years into the future.
@AndrewYNg
So what is your best estimate of the probability that if AI is not strongly regulated it will lead to human extinction in the next 30 years? If you are a true Bayesian you should be able to give a number. My current estimate is 0.1. I suspect Yann's is <0.01
@ylecun
The central issue on which we disagree is whether LLMs actually understand what they are saying. You think they definitely don't and I think they probably do. Do you agree that this is the core of our disagreement?
We show that Gaussian RBMs can generate good images just like other generative models, despite the single-layer architecture. Key innovations:
1) Gibbs-Langevin sampling;
2) modified Contrastive Divergence.
Paper:
Code:
1/2
Does HSBC UK have any ML people? HSBC will not comply with my written instructions to transfer money within the UK. Fraud detection says it must be authorized by high value transfers. High value transfers say they cannot authorize it. 7 hours on the phone so far. Help!
@MasoudMaani
IMO the public perception that Google is not making much progress in AI is wrong. They invented transformers and diffusion models. The perception comes from the fact that they have been very cautious in what they release.
I am excited to support
@raquelutrasun
as she begins her next chapter with
@waabi_ai
. She has been an AI pioneer for the last 20 years and will bring new thinking to an incredibly important space.
@SpirosMargaris
Maybe I over-reacted. When I read it I thought it could easily be interpreted as implying that I left so that I could criticize Google and that is certainly not the case.
My friend
@Jordanjacobs10
of
@RadicalVCFund
asked what I thought of the
@covariantAI
team and technology. I told him I had made a rather small investment (I don't want to reinforce reinforcement learning) but now wished I had invested 100X more. So Radical did (and then some).
Applying reinforcement learning to environments with sparse and underspecified rewards is an ongoing challenge, requiring generalization from limited feedback. See how we address this with a novel method that provides more refined feedback to the agent.
@jeremyphoward
Risk management of things that are totally novel and for which you have zero experience and little understanding is very different from normal risk management. In the absence of any experience, understanding is all you have. I'd go with whoever has the best understanding.
@ylecun
@vkhosla
That is a very ambitious claim. My guess is that you base it on your idea that the errors of an autoregressive model must increase as the output gets longer. This is nonsense as soon as you allow it to add corrections to what it just said.
Students at U. Toronto have a website
for people in the Greater Toronto Area to report symptoms of COVID-19. The reported numbers of likely positive cases and vulnerable individuals are displayed for each region defined by the first half of the postcode.
@pmddomingos
Yes. It would be crazy not to consider this possibility.
Have you seriously considered the possibility it might wipe us all out and if so what probability do you put on that happening in the next 30 years. If you are really a Bayesian you should be able to give me a probability.
@ilyasut
Doing something of your own free will means the decision to do it was based on your own goals and desires. Free will is perfectly compatible with determinism.
@jeffclune
There is so much possible benefit that I think we should continue to develop it but also put comparable resources into making sure its safe.
@andywalters
Yes, I was wrong about the time scale, but AI is already comparable with radiologists for several types of images. It seems likely that in a few years it will routinely be giving second opinions and in 10 more years the second opinions will be better than the human opinions.
@AndrewYNg
There is a big difference between my best bet (which is that we are a passing phase of evolution) and a considered estimate of the probability of extinction used for making policy. For the considered estimate you need to entertain the possibility that your own best bet is wrong.
Interpreting end-to-end trainable autonomy models is crucial for fostering trust and ensuring safety. The use of transformers in Vayu's autonomy model allows us to interpret the model's thought process through attention maps.
#AI
#NVIDIARobotics
I was not far off!
Data from
@StanfordEng
Professors
@yicuistanford
& Steven Chu show
#N95masks
can be decontaminated without decreasing filtration efficiency using 70C heat for 30 min. Alcohol & bleach should not be used.
The sequence modeling group at the Toronto lab of Google Research has some really impressive new work on generating the words in a sequence in parallel. Imputers rock!
If you're interested in parallel generation of output sequences in machine translation and speech recognition, check out our new work on "Imputer", achieving 28 BELU on WMT'16 En>De just in 4 generation steps.
translation:
speech:
Someone with access to two boats should be able to recreate this and make a viral video of it. The water needs to be rough enough to disguise the wakes except when they positively interfere.
@ylecun
People confabulate all the time (especially when talking about LLMs). Recent work on skill-mix by al. et. Sanjeev Arora is a pretty convincing demonstration that GPT4 has a combinatorial ability to create novel text that was not in its training data.
@VayuRobotics
is designing light, low-speed robots for local deliveries. These robots have about 1% of the kinetic energy of a 50 mph car and a much shorter stopping distance, which makes it far easier to make them safe. (2/3)
It seems to me that
@VayuRobotics
is addressing a great market niche and their approach has fewer ethical problems than many other AI applications. I look forward to working with
@nitishsr
and his team. (3/3)
So excited to share our work on RL for Chip Placement, which enables chip optimization in < 6 hrs, whereas baselines require manual efforts and can take weeks. Joint work with my co-TL
@annadgoldie
,
@JeffDean
(cont.)
Paper:
Blog:
We've trained an unsupervised language model that can generate coherent paragraphs and perform rudimentary reading comprehension, machine translation, question answering, and summarization — all without task-specific training:
Canada prides itself on being a more caring society than the US and is publicly apologetic about its past abuses. So why is it ignoring current torture in its prisons?
If two boats on a similar course went by some time ago, their wakes would be small and almost parallel. Where the wakes intersect you would get an interference pattern of bigger waves alternating with smaller waves and this pattern would move sideways in an undulating fashion.
I tried HSBC chat. Here's what they said:
I’ll transfer you to an agent now. It could take up to 6 hours to get connected because our agents are helping other customers like you. Please feel free to log out of Online Banking, and check back later for our response. Thank you.
@pmddomingos
Of course it was the most successful funder. It had by far the most money to spend. The question is would that taxpayer money have been better spent if it had not been required that every project have a military application. Are self-healing minefields a good use of money?
@OriolVinyalsML
I believe that not referencing our paper was not plagiarism because I wrote the abstract. What had not occurred to me at the time was that the cumbersome ensemble of 100 trillion parameter models was us.
When the sun is low and you are looking towards the sun across the water, the nearside of a steep wave looks black. So how could we get a pattern of waves that alternate between being steep and shallow and move sideways to the line of sight across the lake?
@roydanroy
Good point. My tinyLM from 1986 was intended to model how people understand words. Neural net models, despite their many inadequacies are still the best models we have. Understanding consists of creating features and their interactions that explain the data.
@menomnon
Not exactly. Google has had comparable big transformer models for several years, but chatGPT was the moment the general public began to understand the phase transition we are witnessing. It's a bit like Yann's convnets and the Alexnet moment in 2012.
@boazhsan
Yes, the analogy is not perfect. Also the same billion people can have their knowledge turned into many different butterflies. But the main point is that we are just the larval form of intelligence.
@rishabh16_
@hardmaru
Laura Culp, Sara Sabour and I are testing the idea that the embedding vector for a part can communicate a multimodal distribution over poses and identities to the level above and the level above can disambiguate this information by combining with information at nearby locations.
@inemshan
That's only my view about backpropagation when trying to understand the brain and even then I'm not sure. It would be silly to throw it away when doing engineering, because it works really well.
@frans__arthur
@jeremyphoward
As Raza Habib said, there is plenty of literature on this. My main reason is simply that I know of no cases where a much more intelligent agent is controlled by a much less intelligent one.
@sterlingcrispin
I don't think you understand how probabilities work. If I put equal weight on my 0.5 and Yann's miniscule, I get 0.25. I take Yann's opinion very seriously which is why my best estimate is less than the 0.5 that I would be inclined to say based on my own reasoning.
@cdavidnaylor
I see that Oscar is fashion-conscious. AI (by which I mean artificial neural networks) will do many wonderful things in healthcare, but I agree that understanding the dynamics of this epidemic is a job for good old-fashioned statistical models.
@AriDavidPaul
I am not suggesting that. I am appealing to the idea that different people have very different opinions and that for setting policy (or winning competitions) its a good idea to average opinions. Its called ensembling.