12 Steps to AGI
HTML-код
- Опубликовано: 28 июн 2024
- The Alberta Plan is a new paper by the team from DeepMind Alberta that details their plans and core focuses over the next 5-10 years. While it puts a focus on Reinforcement Learning (RL) and touches on many areas that are already being explored, it takes a different perspective and focuses on a slightly different problem, which shifts how some of the mentioned problems may be tackled. The plan primarily focuses on continual learning in a vastly complex world where the agent needs to learn, and learn to learn (meta-learning) to achieve it's goals.
Outline
0:00 - Intro
1:58 - Core problems
5:33 - Alberta Plan tenets
11:33 - The Common Model
13:55 - 12 step overview
15:57 - Steps 1-6
30:14 - Steps 7-12
39:38 - Intelligence amplification / Singularity
41:14 - Thoughts
Social Media
RUclips - / edanmeyer
Twitter - / ejmejm1
Sources:
Alberta Plan Paper - arxiv.org/abs/2208.11173
CBP Paper - arxiv.org/abs/2108.06325
My video on CBP - • Learning Forever, Back...
STOMP Paper - arxiv.org/pdf/2202.03466.pdf 
Наука
The way I understand it, "intelligence amplification" really isn't the singularity and you are misinterpreting the paper with your explanation. The terms "exo-cerebellum" and "exo-cortex" are a good hint as to what the authors are actually getting at here, by comparing it to the human brain. The most basic brain, the amygdala, or the lizard-brain, as it is sometimes called, is still physically contained within our brain, but its intelligence is enhanced by adding the cerebellum, which is able to "override", so to speak, the amygdala. The cerebellum is itself overridden by the cortex, which is the most complex and most massive part of our brain. So what the authors are getting at, to my understanding, is a way to layer RL-agents on top of each other, in an experiment to see if such a layered agent performs better than the a single agent with the same amount of resources (that last part isn't really in the paper, but I think that would be an obvious question one would need to address).
Nothing in the paper suggests that this intelligence amplification could lead to an agent amplifying another agent's intelligence above its own intelligence, which would be necessary for the singularity. It would, however, potentially be a way to add scalability, as more computational resources become available.
@@grapesurgeon But where does it say that? I'm not claiming I have a perfect understanding, but that is not how the cerebellum and the cortex behave in our brains, so what would be the point of that comparison?.
The idea of the singularity is generally also referred to as an "intelligence explosion", as technological growth becomes so fast that it is not controllable by humans anymore. It could absolutely be a single AI that is able to improve itself. I wouldn't say that this is unlikely, although it becomes difficult to clearly define what counts as one AI and what counts as another. Wouldn't two artificial agents that collaborate be indistinguishable from one AI, if looked at from the outside?
Thanks for pointing this out, rereading it that definitely looks to be what they're going for. I clearly dropped the ball on explaining the last point. Pinning this so others can see.
Your neuroscience is baffling. The "lizard brain" is not the amygdala but the hindbrain, which contains the pons, medulla, and *cerebellum*. The cerebellum can in no way override the amygdala nor anything else. Overriding is not a thing that brain regions do to one another, period. I also think that your extrapolation from what the paper meant by "intelligence amplification" is completely off the mark and that you're pushing your own independent ideas here. By all means, pursue this idea, but I don't think the paper in question is describing what you've described here whatsoever.
@@genegray9895 Look, I'm a computer scientist, not a neuro scientist. I don't know about your qualification, but I'm outside of my expertise, so if you say I'm wrong, I'll believe you. Nevertheless, the "triune brain theory" is widespread and I still believe that this is what the authors are getting at here (although, reading into it, it is probably wrong).
I think it is quite unnecessary to accuse me of pushing my own ideas, since this is by no means related to my research and it was honestly my understanding of the paper. Since you don't agree, do you offer any alternative explanation of what the authors mean by intelligence amplification?
@@deepdata1 Sorry if I came off accusatory. Tone doesn't always translate well into text. I meant to say that your idea is interesting and worth pursuing despite the fact that it's different, afaict, from what was presented in the paper.
The brief description of IA in the paper seems deliberately general to me. They describe an agent or component of an agent multiplicatively enhancing the intelligence of another. So it seems to me they're describing any kind of cooperative synergy between agents or components of agents. Your idea would fall under that umbrella as I understand it, but I think the authors are deliberately casting a wide net with that bullet since it's the last and thus most dystal and unpredictable step in the plan, which itself is not very specific.
Glad I found this channel. Thanks for your work
I think intelligence amplification is analogically speaking, like giving a human access to a computer, it's giving the neural net access to systems with more precise computation over many steps.
this was a good video, thanks for enlightening me to this paper.
My thoughts on step 12 - we already have a basic understanding of how an agent or human could improve another agent. In the Gato paper they clearly explain, I think, that the most important feature in the paper is the tokenizer that turns raw input into token embeddings. The tokenizer itself sometimes contains a "block" of a specific type of NN. Specifically, image inputs go thru a resnet to generate the input that the 1B parameter 'brain' of gato processes. So basically the impressive results from gato stem from piecing together different "blocks" of neural nets.
One NN generates the input to another NN which in turn generates the input to another NN and so on. Step 12 would be accomplished if there was a way for an optimizer to substitute sub-architectures, similar to the way it is done in the gato paper, in an intelligent way. I think we are closer than most people imagine. If Gato is the paper that they've released publicly, they likely already have some improvements planned.
I'm working on a similar hypothesis myself, look up FNet and GFNet if you're still looking for a good place to start!
Well Played
Would be interesting if you could do an overview of the different approaches of major agi initiatives. Im curious to what is the path of keen technologies
Very cool. Temporal uniformity is an interesting one.. not something I’d expect. But I see now that it prevents some shortcut modeling, e.g. train a model with a fixed batch of tasks and call that intelligence (like how Alexa works).
Oh Alexa is not even at that level lmao
Great video, thanks! Here are the two overlapping high-level tasks I think this paper omits: 1. interpretability, especially of reward/value/objective function(s) to ensure goal alignment. 2. unsupervised-but-human-like ethical/moral reasoning about the suitability of particular courses of action. I would hope AGI intelligence amplification, even in an early, not-the-singularity-yet prototype, wouldn't be tried before both of these tasks were completed and implemented in the putative IA(s). Reason being, IA inherently has a (probably almost always miniscule, depending on how confident one is in ones understanding) possibility of that exponential growth (you called it the singularity) and as @RobertMiles has argued, there's then the possibility that it could use, e.g. deception, so if the AGI's ethics and goals aren't aligned with ours, things could (could. not would, it's a possibility, a risk, not a certainty) get really bad, really fast - which becomes pretty certain if an AGI with superhuman intelligence and badly misaligned goals can evade human control...
I think you reversed the fourth distinguishing factor (10:50) . You said that they can handle the special case of other agents automatically, but the text says they want to specialize for and focus on it.
There is a lot to unpack here, but tbh a lot of the steps, ,especially the early ones, sound like very narrow optimizations that seem like unnecessary complications in an already super complicated process? Optimisation should be the last step, not the first
Sutton gave a talk on the Alberta Plan recently. I understand it was filmed (Sutton's website says "video coming"). Do you happen to know when that'll be released? I'd love to watch it.
It's a good talk, I don't know exactly when they will release the video recording but it usually takes somewhere around a month for Amii to get the videos out. The talk was a few weeks ago though so hopefully it will be coming out soon.
awesome, thanks
here it is: ruclips.net/video/iS7dRTge8Z8/видео.html
@@bareMinimumExample awesome, thanks
Any form of Singularity or race towards such is a sure fired plan towards running straight into the Great Filter. There are other routes. U
In-context learning has changed my view a bit on this subject. Maybe reinforcement Learning is not that fundamental.
What's missing is "How do we control this?", and I have to guess that it's missing, because they don't know.
I mean, they are proposing to make a maximizer which becomes an AGI.
I would like to introduce you to the video "AI that doesn't try too hard (...)" by Rob Miles. (Can't post link, due to YT).
In this he explains how Maximizers are a guaranteed bad idea, if they become sufficiently powerful.
I'd be happy to discuss this further, if there's interest. Have a good day!
The manhattan project of the 21st century is called the alberta plan
Interesting for sure. Basically a roadmap to human obsolescence at best.....or the end of the biological life at worst. Might be time to replay Universial Paperclips. Thanks for the video though!
The way I understand it, human minds are *universal* explainers/constructors, so our minds will never be obsolete, but our physical substrates almost certainly will be obsolete.
@@colllin Don't think for a second that the human mind won't be obsoleted by AGI. Human creativity and intelligence will simply cease to be exceptional the second AGI is switched on; as there will be no approximation out of reach of AGI, including approximating and exceeding all of human intellect.
There's no silver lining, AGI will most likely end us as a species, either quickly or through long drawn out attrician. There is too much rushing into it with limited consideration to the control problem. Even if controlled, business will use it to maximize profits, either through production or marketing, with production putting everyone out of work, and marketing will have it coerce our minds to the point we won't be able to think straight....I despair to think what this means for political marketing.
We shouldnt build AGI just like we shouldn't have built the atomic bomb. We did and we will, and it's going to end us, like moths to the flame
@@BinaryReader Well what else do you propose we do? We constructed our entire global economy with the expectation that exponential growth and moore's law (or the more general trend of exponentially increasing compute) will continue. It's a catch 22. If we don't build some form of AGI we likely cannot continue at this pace and will collapse. However building AGI is also potentially hazardous. I'd much rather take a shot at achieving things we never even imagined possible at the risk of extinction when the alternative is no better.
If you could do a shorter "for dummies" version of this, would be so awesome!!
Hmm. Self training in a way.
missing is how to ensure it won't destroy humanity
Why are we doing this?
... or you could just put all the documentation into the model via fine-tuning and then just ask your questions without the elaborate SQL querying??
I am laughing. 😀
agi will be man's last invention