I feel angry that it is so hard to publish machine learning contributions on small data. Most problems are not images and NLP, and most data scientist work on relatively small tabular data. And most of the _value_ is there. And we know very little about that regime.
What do you think about this argument?
GPT-3 is obviously not intelligent. It was fed by a nontrivial percentage of all text produced by humans which cannot grow by several orders of magnitude. Therefore the paradigm of feeding data into function learning is a dead end for AGI.
@ylecun
I don't think it's that simple. The most awesome property of living cognition is not the absorption of 1E15 bytes of video "data" the child sees but its filtering. How little it actually needs to "see" to build a world model. And the key is that a child is not a mere observer,
@ylecun
You can only learn causal models by embodying the AI and let it manipulate the world, running experiments. Watching videos and reading text will not get you to causal world models, whatever the architecture and the task you define.
We are hiring at Huawei's Paris AI research!
- PhD in 2020 or 2021
- permission to work in France
- strong publication record
- preferable domains: RL, bandits, physical systems, time series, AutoML, combinatorial and online optimization, transfer learning
RT please
@ylecun
@EduardoSlonski
@DrJimFan
@RichardSSutton
We also have 10s to 100s millions intero and exteroceptive neurons (estimate by GPT :) ).
Also, we don't just given a set of data and learn a function, we live, the paradigm is way more complex than anything we do in ML. The rest of the body, may be as important as the network
I think "Bayesian" is mostly a synonym of "subjective", probabilities unsupported by third-person scientific data. I think there is nothing wrong with quantifying feelings and even asking for validation, just let's be clear about what we're doing.
Two types of Bayesians
1) Beer pub Bayesian: People who want to be seen as intelligent by using words like "Bayesian", "prior", "posterior", "updating" to refer to the fact that people have opinions and sometimes change their minds.
2) Actual Bayesian who analyzes data
@davidasboth
The surprise of the last year for me that you can train neural nets on 200 points. OK, it's smooth low-dim data. But you _do_ need NNs, linear or forests do not work. And the experiments generated strange puzzles over heteroscedastic NN regression.
@tszzl
Yes. Banks seems to overlook the spiritual way though that we also have a tendency to follow facing technocratization. Culture is changing, but so do we, individuals. With food and shelter assured, more and more people will have time to explore enlightenment, which will
.
@vervaeke_john
proposes here an alternative to the Turing test wherein we let two copies of the subject to talk to each other. His scientific prediction is that within a minute the conversation, and I quote, "goes batshit crazy". I tested GPT-3.
👇
@GaryMarcus
Given the enormous, unimaginable suffering in the 20th century, should we have stopped the Enlightenment? Had Luther foreseen the death of God, would he have gone ahead with his proposals?
I've been in this ML business for about 30 years now, and still can't get around the fact that the optimal random forest overfits like crazy. This is not how the world supposed to be.
A major milestone: our
#RAMP
preprint paper is out. Don't hesitate to comment or discuss our vision on the openreview site if you are interested in
#transparency
and
#reproducibility
in scientific
#datascience
workflows.
@ylecun
True.
What you fundamentally cannot learn by mere observation is what data you should learn from. What prediction is important/useful to care about. We are trying to "teach" AI everything to compensate this lack, and quantity covers it up superficially. But this is the main
I love that ICML accepts these kinds of papers, actually it's pretty surprising, we didn't even think about aiming.
Just look at this plot and promise me never again using standard error as your confidence estimate on heavy-tail RL losses.
Andrew Patterson, Samuel Neumann,
Anybody else noticed that in reinforcement learning, performance across seeds is non-Gaussian? How do you cope with it? This means that reporting mean and even standard deviation is not that meaningful, we're averaging qualitatively different things.
I don't know about "batshit crazy", but as much as I can tell, it is the right scientific term. More precisely, it's like Sheldon trying to pick up a woman in a bar.
Let's celebrate the 30th anniversary of Rich Sutton's Dyna paper.
Four pages, zero experiments, but one full page on limitations, 600 citations, and the founding of a sub-branch in reinforcement learning on which I am betting for AGI.
Forget ChatGPT, here is what's coming.
As all devices will be equipped with self-optimizing AIs, they will naturally develop their personalities (since their experiences will differ). This means that engineering will go from designing "machines" to "shepherding" little AIs.
I think "Bayesian" is mostly a synonym of "subjective", probabilities unsupported by third-person scientific data. I think there is nothing wrong with quantifying feelings and even asking for validation, just let's be clear about what we're doing.
@QiaochuYuan
I like him, like his approach, and happy that he has success. He inspired me to start my podcast. You don't need to listen to him. You can also do what he does, follow your own instincts and manifest the way you think conversations should be done. It's enlightening.
The difference between planning in
#reinforcementlearning
and in
#LLM
is that in RL we check where we really are after each action, and replan. This means that the agent needs to be connected to the world. This rhymes with what
@ylecun
has been arguing.
Catchy title (GPT-3 is not a god to be believed in), but the article is hyper-precise. I think GPT-3 is a nice tool but missing grounding. Also nice to see how my vision articulates with
@ylecun
's.
IA : Pourquoi Huawei ne croit pas à GPT-3 via
@zdnetfr
As we become less tribal, a major source of conflict is perceptual diversity. That we cannot imagine that others perceive the world differently from us.
@anilkseth
Well, I think that's exactly the point of
@GaryMarcus
. Images that fool us need to be carefully engineered, and play no role in our survival (on the contrary, our "illusions" are probably necessary for survival), whereas deep nets have an arbitrary relevance landscape.
I agree with
@drmichaellevin
that we have been producing high-level agential intelligences aka kids, without asking for guarantees about their alignment.
That raises the following poll.
Retweet if you're interested in the answer.
Do you have
@BeingMIAkashs
@ylecun
Model-predictive control on a learned model _is_ reinforcement learning. It is called model-based reinforcement learning. This is like saying: "Abandon Machine Learning, in favor of optimization".
@OkbaLeftHanded
That's true but in my experience e.g. small data reinforcement learning has a huge market, trillions per year. It's not the IT sector, more heavy (control) engineering and transportation so it's not the usual high-liquidity companies that dominate ML research today. But it's big.
@moreisdifferent
@tdietterich
Yes, we have the same experience in model-based reinforcement learning. I hate it because it's dumbly inefficient, but retrain from scratch on the full data each time a new data arrives is unbeatable for now.
@casebash
@drmichaellevin
Best to listen to
@vervaeke_john
.
We don't align kids by giving them and enforcing rules (purely), but by embodied stories and rituals, and by bringing them in contact both with others and with reality, so they bind to both.
The
@SaclayCDS
Solar wind
#RAMP
event of tomorrow is open. The goal is to detect variable-length events in multiple aligned time series. Don't hesitate to sign up if you would like to participate tomorrow or in the next two weeks.
@CRSTNNSL
@DavidDeutschOxf
I think it may be purity vs mixing. There was less population density, languages developed "on their own" for 50K years. Then with the Indo-European explosion, there was much more mixing, and mixing meant to get rid of fragile complexities. You can see it in modenr-day English,
On the way to AGI, I think pure model-free RL is a dead end. Yes it is difficult for an agent to figure out what slice of the world matters for reaching its goals but this is not an excuse for not learning a model for that slice. Every living thing builds such a model. Pls refute
My conversation with
@AnnaCiaunica
about embodied cognition, feeling like a zombie, depersonalization, meditation, movement therapy, 1st person science, and much more.
Pls give it thumbs up, share, comment, and sign up!
@TonyZador
Hm, you put the genome in a test tube, it won't hunt. You put the colt in water, it won't stand. The point is that there are a lot of environmental and developmental affordances and constraints that produce the behavior, and those are also information.
I am often surprised by polling but this one scares the shit out of me.
1 out of 5 people 18-29 believe the holocaust is a myth. How did it get to be this way?
Congratulations to my and
@AKAZAKCI
's PhD student
@mehdidc
defending his thesis on
#DeepLearning
and novelty generation. If you want to understand why we count letters generated by a model trained on digits when we don't even want to generate letters, ask him a copy of his thesis
DALLE is definitely out of sync :)
"Realistic photo of a young tourist-looking guy smuggling weights of a large language model in his backpack at US airport customs in Chicago."
@bernhardsson
I have an upcoming preprint (brutally downscored by NeurIPS :) ) that shows that some of the Ray Tune engines beat random search by a factor of 2x to 3x. Best ones seem AX, BlendSearch, Hebo.
OK, I've seen stuff. But 6000 French CEOs celebrating my 17-year-old son with a standing ovation still feels surreal. What a day! I'm immensely proud, bravo Martin! Thank you
@tweets_Apm
!
@DrJimFan
@tegmark
No debate, I agree :). What I would add is that "world" is a relative concept. It's what we sense. So LLM has implicit information about our world model. A WiFi antenna's world model will be quite different. In that sense LLMs will be helpful for an agent whose world model
The four reviews we got on our NeurIPS paper on model-based RL are excellent. The paper will be rejected but it's rather because of systemic issues and the lack of a clear scientific paradigm in ML than any misreading or not reading of the paper by the reviewers.
@vgr
I built a radiator in my bedroom out of the cooling grid of an old fridge, a plastic pipe, and some corks during the 97 ice storm in Montreal. Electricity was out for a week but gas water heater worked in the bathroom.
@Aella_Girl
I'm with David Deida on this. That desire is like the air I'm breathing, life energy that I can turn into creation. No shame to have it and no suppression. But I don't act on it because I'm monogamous and what we're building with my partner could not exist if we were poly.
@sentientist
Think about them this way: you are their frontal lobes. They are born immature. So their behavior is a mirror of your toddler self. My worst rage is coming when they do something I was not allowed to. My inner toddler shrieks, IT'S NOT FAIR!!! Golden material to bring into
Entropy is ~ the log number of microscopic states that produces the same macroscopic state, right? So can we say that without postulating _levels_ (at least a micro and macroscopic level of analysis), entropy loses its grounding?
I'm super excited to announce my podcast: I, scientist, where we explore AI, body, and soul.
My first guest is
@_GPaolo
whom we discuss agency in AI, reinforcement learning, and much more. Pls give it thumbs up, subscribe, comment, and share.
Our next single-day
@SaclayCDS
#RAMP
hackathon will be on October 10 at the INRIA Turing building on detecting Solar storms from spacecraft measurements (event detection in time series).
If you would like to participate, please sign up before October 3:
@hardmaru
@slashML
It's painstakingly slow because there are so many moving parts and they all depend on each other. Also takes a couple of years to build a good experimental stack. Third: benchmarks are all over the place, experimental settings are not fixed.
Are we rushing ahead into the unknown, scaring people, while also producing fake shelters for them to retreat into artificial small worlds? Are we losing our grip on the world or realizing our humanness in the mirror of AI? Will AI take our agency away? Are we willingly giving it
Our
#ICLR2021
paper on model-based RL with
@albertcthomas
and Gabriel Hurtado
Blog:
We compare generative models on Acrobot. Both multimodality and heteroscedasticity are crucial, even though the environment is noiseless.
1/n
We've just submitted a position paper on embodied AI. One of my realizations during writing is that social media AI is more embodied than LLMs, making them more natural to interact with, which may explain why there is less buzz about their alignment, even though their impact on
@fchollet
You underestimate the reliability of the source when we judge information. Our whole brain is about judging epistemic authority. I can't check the veracity of every piece of information, it defeats the purpose of distributed intelligence.
@Philip_Goff
Can you elaborate how theists deny suffering? My understanding would be more like they make suffering meaningful, thus bearable. Avoiding suffering is not at the top of the value hierarchy, but suffering exists.