Shane Legg

My favorite application is Marcus Hutter's AIXI[1], a provably optimal theoretic intelligent agent for any environment, though uncomputable. And then Shane Legg's use of a Monte Carlo AIXI approximation to define a universal measure of intelligence[2]. [1]https://en.wikipedia.org/wiki/AIXI [2]http://www.vetta.org/documents/Machine_Super_Intelligence.pdf

His blog from before DeepMind was interesting then, though he seems to have stopped posting. http://www.vetta.org/

Interestingly, he once called MIRI[1] (formerly SIAI) “the best hope that we have”[2]. [1] https://intelligence.org/ [2] http://www.vetta.org/2009/08/funding-safe-agi/

> Can we equate it as the ability to conceptualize a Turing machine? This would be the same as the ability to follow arbitrarily complex instructions, given enough time and memory. I don't quite follow (who is being compared to whom?), but FWIW, humans are not always very good problem-solvers. There is a very obscure vein of research developing IQ tests which are basically just Turing machines; see http://www.vetta.org/2011/11/aiq/ - one of whose developers says: > There is an agent called Manual.py in there. It’s pretty primitive, but I did use it to see how well I could do. I seem to recall that I was around 70… but I didn’t manage enough samples to get a good statistic (I spent a few hours doing it before I started going crazy). Anyway, if you’re used to using Python I’m sure you can hack up something better…

Here are a set of links: http://www.vetta.org/recommended-reading/ I'd second the recommendation of Bishop if you can hack the math, and also Elements of Statistical Learning, though I wouldn't attempt to learn techniques from the latter so much as look at a very interesting mathematical take on them. gl

It's difficult not to get caught up in the excitement around Deep Learning. The unsupervised nature of it at its best really does seem to approach human levels of learning (if not intelligence). Also, I thought this was hilarious: In a 2011 interview that predated DeepMind, co-founder Shane Legg said he gave only a 50 percent chance that human level-machine intelligence would exist by 2028. ONLY?!? By 2028? That is a pretty radical prediction. Indeed, Legg says there is a 10% chance of human level intelligence by 2018, and 90% by 2050[1]. Legg has multiple papers published in the field of measuring machine intelligence, so I guess he has a pretty good view of the field. If that 2028 prediction is even close to correct then I think setting up an ethics board now probably is the correct thing to do. (the notion that DeepMind asked that Google create an internal ethics board as a condition of the acquisition, as reported by The Information, had some AI researchers griping) Edit: Legg's posting on his blog in 2011 deserves to be read: I’ve decided to once again leave my prediction for when human level AGI will arrive unchanged. That is, I give it a log-normal distribution with a mean of 2028 and a mode of 2025, under the assumption that nothing crazy happens like a nuclear war. I’d also like to add to this prediction that I expect to see an impressive proto-AGI within the next 8 years. By this I mean a system with basic vision, basic sound processing, basic movement…

Link without ssl (notice the http at the beginning): http://www.vetta.org/documents/Machine_Super_Intelligence.pdf

Basically Shane Legg (who is probably not the Google AI chief), chief AGI scientist at DeepMind, who said in 2011/2012 (https://www.vetta.org/2011/12/goodbye-2011-hello-2012/) that > I give it a log-normal distribution with a mean of 2028 and a mode of 2025, under the assumption that nothing crazy happens like a nuclear war. I’d also like to add to this prediction that I expect to see an impressive proto-AGI within the next 8 years. By this I mean a system with basic vision, basic sound processing, basic movement control, and basic language abilities, with all of these things being essentially learnt rather than preprogrammed. It will also be able to solve a range of simple problems, including novel ones. is saying the same thing but pushed his timeline back a few years. I assume if you ask Shane 3 years ago before GPT he would look away and murmur something like "kurtosis".

Shane Legg is a cofounder of DeepMind and an AI researcher. He was pretty casual about predicting human level AGI in 2028. https://www.vetta.org/2011/12/goodbye-2011-hello-2012/ He doesn't say so publicly any more, but I think it is due to people's negative reaction. I don't think he changed his opinion about AGI.

"SciPy – the embarrassing way to code" http://www.vetta.org/2008/05/scipy-the-embarrassing-way-to-code/

Here is an interesting definition and way to measure intelligence http://www.vetta.org/documents/AIQ.pdf an informal definition from the paper: “intelligence measures an agent’s ability to achieve goals in a wide range of environments” There is also a formal definition and a method of measuring it. EDIT: a blog with a summary of the paper and source code http://www.vetta.org/2011/11/aiq/

http://www.vetta.org/2008/05/scipy-the-embarrassing-way-to-code/ ;)

> Shane Legg is known for being a co-founder of DeepMind, in a business role (as I understand) You are quite mistaken. He leads the applied AI team there, and has significant history in research. http://www.vetta.org/publications/

Shane Legg isn't just a business guy - his career has been pretty much focused on AI research since uni - http://www.vetta.org/about-me/

There's an algorithm developed by Marcus Hutter called AIXI, which makes provably optimal decisions. Unfortunately(?) it's also uncomputable, but computable approximations exist including a Monte Carlo variant: http://www.vetta.org/2009/09/monte-carlo-aixi/. As the paper notes it scales extremely well; to get better results you just throw more computing power at it.

The clearest description of what CS researchers talk about when they talk about "intelligence" (see http://www.vetta.org/definitions-of-intelligence/) is probably a production function from economics.

Shane Legg and Marcus Hutter of IDSIA did some recent work on a metric for machine intelligence. ( http://www.vetta.org/documents/ui_benelearn.pdf ) Marcus Hutter's PhD thesis produced a theory of "Universal Artificial Intelligence" called "AIXI" based on Solomonoff induction. Sadly, AIXI is incomputable, and the time and space bounded version "AIXItl" is still impractical. http://www.hutter1.net/ai/aixigentle.htm But you're right: we don't know much about intelligence nor how it should be embodied on a computer. This is why we should encourage some really smart people to focus on the problem! I sincerely think this is an area that would yield to research. Might take a long time and a lot of brainpower though... Lenat's Eurisko is pretty cool, and I don't know much about it, but it doesn't sound like Eurisko is as general as is needed.

The call for AI safety has existed since before we broke through the Turing test with LLMs. And I personally wouldn’t call things like code generation or content-generated learning experiences for advanced topics “basic”. Not to mention where we’re headed with multimodal integration. Many have argued for safety for decades. They’ve predicted and built the AI trajectory, they’ve been right, and we should listen. > If one accepts that the impact of truly intelligent machines is likely to be profound, and that there is at least a small probability of this happening in the foreseeable future, it is only prudent to try to prepare for this in advance. If we wait until it seems very likely that intelligent machines will soon appear, it will be too late to thoroughly discuss and contemplate the issues involved. ~ Co-Founder of Deepmind, 2008 https://www.vetta.org/documents/Machine_Super_Intelligence.pdf

Researchers aren't a monolith. Just because Gary Marcus (a complete fraud by the way, look up his "XProp" magitech he sold to Uber) pooh-poohed connectionist AI until ChatGPT came out doesn't mean nobody predicted gains from scaling; certainly many people deep in the know expected things to click. Certainly they expected that the basic trick of associative learning will be cracked. E.g. here's Shane Legg, December 2009 [1]: > Conclusion: computer power is unlikely to be the issue anymore in terms of AGI being possible. The main question is whether we can find the right algorithms. > One of the big things influencing me this year has been learning about how much we understand about how the brain works, in particular, how much we know that should be of interest to AGI designers. I won’t get into it all here, but suffice to say that just a brief outline of all this information would be a 20 page journal paper (there is currently a suggestion that I write such a paper next year with some Gatsby Unit neuroscientists, but for the time being I’ve got too many other things to attend to). At a high level what we are seeing in the brain is a fairly sensible looking AGI design. You’ve got hierarchical temporal abstraction formed for perception and action combined with more precise timing motor control, with an underlying system for reinforcement learning. The reinforcement learning system is essentially a type of temporal difference learning though unfortunately at the moment there…

> If easy instances are enough, what's the relevance of the provably ideal uncomputable problem, besides rhetoric? It shows that there's a path to AI that's gated on solving formal math problems instead of imitating wishy-washy emotional humans. This point may be obvious to you, but apparently it wasn't obvious to Stross. > If humans used easy instances, and they're the only example we know of that achieved I, how is that not weak evidence that AI through easy instances will pass through human-like I? Humans are also the only animals that can play chess, but chessplaying software didn't pass through a human-like stage where they got nervous, forgetful, etc. My intuition says math can't be that radically different from chess (both involve search on trees with a complicated evaluation function), and we're only a handful of new insights from having computers decisively beat humans at math. And at that moment, by implication from my previous point, they might just also beat us at everything else :-) Remember that, on an absolute scale, humans are about as bad at math as you can be and still build a civilization (otherwise civilization would've happened at a lower level of development) and also that we probably suck at math because math is a conscious task (see Moravec's paradox). Related link that should make you scared if you buy my reasoning: http://www.vetta.org/2009/09/monte-carlo-aixi/ .

No need stupid moving texts. CS231n: Convolutional Neural Networks for Visual Recognition - https://cs231n.github.io/ The Unreasonable Effectiveness of Recurrent Neural Networks - https://karpathy.github.io/2015/05/21/rnn-effectiveness/ Understanding LSTM Networks - https://colah.github.io/posts/2015-08-Understanding-LSTMs/ ImageNet Classification with Deep Convolutional Neural Networks - https://papers.nips.cc/paper/2012/hash/c399862d3b9d6b76c8436e924a68c45… Deep Residual Learning for Image Recognition - https://arxiv.org/abs/1512.03385 Multi-Scale Context Aggregation by Dilated Convolutions - https://arxiv.org/abs/1511.07122 Identity Mappings in Deep Residual Networks - https://arxiv.org/abs/1603.05027 Recurrent Neural Network Regularization - https://arxiv.org/abs/1409.2329 Deep Speech 2: End-to-End Speech Recognition in English and Mandarin - https://arxiv.org/abs/1512.02595 Order Matters: Sequence to Sequence for Sets - https://arxiv.org/abs/1511.06391 Neural Machine Translation by Jointly Learning to Align and Translate - https://arxiv.org/abs/1409.0473 Pointer Networks - https://arxiv.org/abs/1506.03134 Attention Is All You Need - https://arxiv.org/abs/1706.03762 The Annotated Transformer - https://nlp.seas.harvard.edu/annotated-transformer/ Neural Turing Machines - https://arxiv.org/abs/1410.5401 A Simple Neural Network Module for Relational Reasoning - https://arxiv.org/abs/1706.01427 Relational Recurrent Neural Networks -…

https://colah.github.io/posts/2015-08-Understanding-LSTMs/ https://papers.nips.cc/paper/2012/hash/c399862d3b9d6b76c8436e924a68c45… https://papers.nips.cc/paper_files/paper/2012/file/c399862d3b9d6b76c84… https://arxiv.org/pdf/1512.03385 https://arxiv.org/pdf/1511.07122 https://arxiv.org/pdf/1603.05027 https://arxiv.org/pdf/1409.2329 https://arxiv.org/pdf/1512.02595 https://arxiv.org/pdf/1409.0473 https://arxiv.org/pdf/1506.03134 https://arxiv.org/pdf/1706.03762 https://nlp.seas.harvard.edu/annotated-transformer/ https://arxiv.org/pdf/1410.5401 https://arxiv.org/pdf/1706.01427 https://arxiv.org/pdf/1806.01822 https://arxiv.org/pdf/1704.01212 https://arxiv.org/pdf/2001.08361 https://arxiv.org/pdf/1811.06965 https://www.cs.toronto.edu/~hinton/absps/colt93.pdf https://arxiv.org/pdf/math/0406077 https://scottaaronson.blog/?p=762 https://arxiv.org/pdf/1405.6903 https://onlinelibrary.wiley.com/doi/10.1002/047174882X.ch14 https://github.com/Bladefidz/information-theory/blob/master/books/Elem… https://arxiv.org/pdf/1611.02731 https://www.vetta.org/documents/Machine_Super_Intelligence.pdf https://karpathy.github.io/2015/05/21/rnn-effectiveness/ https://cs231n.github.io/

a backup just in case https://nlp.seas.harvard.edu/annotated-transformer/ https://scottaaronson.blog/?p=762 https://karpathy.github.io/2015/05/21/rnn-effectiveness/ https://colah.github.io/posts/2015-08-Understanding-LSTMs/ https://arxiv.org/pdf/1409.2329.pdf https://www.cs.toronto.edu/~hinton/absps/colt93.pdf https://arxiv.org/pdf/1506.03134.pdf https://proceedings.neurips.cc/paper_files/paper/2012/file/c399862d3b9… https://arxiv.org/pdf/1511.06391.pdf https://arxiv.org/pdf/1811.06965.pdf https://arxiv.org/pdf/1512.03385.pdf https://arxiv.org/pdf/1511.07122.pdf https://arxiv.org/pdf/1704.01212.pdf https://arxiv.org/pdf/1706.03762.pdf https://arxiv.org/pdf/1409.0473.pdf https://arxiv.org/pdf/1603.05027.pdf https://arxiv.org/pdf/1706.01427.pdf https://arxiv.org/pdf/1611.02731.pdf https://arxiv.org/pdf/1806.01822.pdf https://arxiv.org/pdf/1405.6903.pdf https://arxiv.org/pdf/1410.5401.pdf https://arxiv.org/pdf/1512.02595.pdf https://arxiv.org/pdf/2001.08361.pdf https://arxiv.org/pdf/math/0406077.pdf https://www.vetta.org/documents/Machine_Super_Intelligence.pdf https://www.lirmm.fr/~ashen/kolmbook-eng-scan.pdf https://cs231n.github.io/