Thanks. I think you are a correct exception to what I said. I should have known that using words like "nobody" would not go over well on HN (but tedious to type "a very large percentage"), despite that statement being a verbatim quote from one of the world's leading ML engineers and, to me, not controversial.
I do consider the cloud both widely available and near infinite in resource adding capability.
If it is really not economically feasible to add resources, then the performance gains were not as promising as thought (whether cloud or on-site).
I would not retrain such a model on all data, just do online updates. Also I still think for that use case training times and latency are negligible (nobody cares or nobody notices any difference between training a BoW and bi-LSTM.)
If you are deploying on resource-constrainted devices (IE: low-end PC's without GPU), it is not unusual to take a lot of time training a model on a very powerful computer (which nobody cares about), then distilling or transfering the result for test time.
Nobody in business cares if you are doing proper AI or dumb curve fitting. What matters is the complexity (engineering debt) and performance (accuracy, robustness).
Online learning, sample -
and energy efficiency are unrelated to training times. Like said: nobody cares if you ran Vowpal Wabbit for 1 hour or 100 hours, as long as you are not constantly babysitting it and calling that paid work (or have the unusual requirement of daily retraining while using an online model).
> simple dot product is always going to be faster than many matrix multiplies
If you care about this (because it is profitable), you rewrite in lower-level language or predict with cloud GPU (which will be at least comparable to simple dot product, while adding performance)
If you are building large-scale systems that take weeks or months to train, you are at a point where you shouldn't care about this. Throw more compute at the problem, it will pay for itself.
If we are talking days or hours: start parameter search on Friday and return best parameters on Monday.
Do research and iteration on heavily subsampled datasets.
If you are building models for yourself, or for Kaggle, you may care in as much as your laptop gets uncomfortably hot.
Nobody cares how long it takes to train a model. What matters is prediction speeds, which are comparable (and NLP less likely to require high frequency, where a few more milliseconds matters).
Besides that, the accuracy gains are not marginal anymore (BoW can't compete like it used to, especially with pre-trained models).
Something can be possible, while still technically not feasible.
I agree our knowledge currently is lacking, but see no reasons why this will never catch up.
There are fundamental limits on cognition. For one our universe is limited by the amount of computing energy available. Plenty of problems can be fully solved, to where it does not matter if you are increasingly more intelligent (beyond a certain point, two AGI's will always draw at chess). Another limit is practical: the AGI needs to communicate with humans (if we manage to keep control of it), so it may need to dumb down so we can understand it.
Even an AGI as smart as the smartest human will greatly outrun us: it can duplicate and focus on many things in parallel. Then the improved bandwith between AGI's will do the rest (humans are stuck with letters and formulas and coffee breaks).
Manually deployed atom bombs and malware can already wreck us. No difference with autonomous (cyber)weapons.
You are correct, in that AI experts may not be the best predictors for AGI. For one, they spend their lives working towards the goal of AGI, so it would require a huge amount of cognitive dissonance for them to say that AGI is impossible or very very far on the horizon.
Philosophers and futurists are better suited to hypothesize an AGI timeline.
But you take it too far by saying it is anyone's game.
Game theory, security, and economic competition makes it impossible to globally ban AI. The incentives to automate the economy (compare AI revolution with industrial revolution) and to weaponize AI (Manhattan Project for intelligence) are just too big. We are already seeing that the US focus on fair and ethical AI puts them at a disadvantage against China and Russia. AGI must require pervasive surveillance of the populace, but the Luddites are holding this back.
I suggest you learn to stop worrying about the bomb, and start planning for its arrival.
Clearly, AGI-level intelligence is possible, because human brains exist.
So unless you pose that a function has to rely on its materialization (there is something untouchably magic about biological neural networks, and intelligence is not multiple realizable), it should be possible to functionally model intelligence. Nature shows the way.
AGI will likely obsolete humanity. Either depricate it, or consume it (make us part of the Borg collective). Heck, even a relatively dumb autonomous atom bomb or computer virus may be enough to wipe humanity from the face of the earth.
Shane Legg (co-founder DeepMind) and Marcus Hutter (Schmidhuber pedigree) defined machine intelligence in this canonical paper from 2007: https://arxiv.org/abs/0712.3329
> Universal Intelligence: A Definition of Machine Intelligence
> A fundamental problem in artificial intelligence is that nobody really knows what intelligence is. The problem is especially acute when we need to consider artificial systems which are significantly different to humans. In this paper we approach this problem in the following way: We take a number of well known informal definitions of human intelligence that have been given by experts, and extract their essential features. These are then mathematically formalised to produce a general measure of intelligence for arbitrary machines. We believe that this equation formally captures the concept of machine intelligence in the broadest reasonable sense. We then show how this formal definition is related to the theory of universal optimal learning agents. Finally, we survey the many other tests and definitions of intelligence that have been proposed for machines.
General intelligence usually meant in relation to humans, but you are correct in noting that it is a spectrum, not a binary.
Your reasoning does not follow. To see why, take something humans already clearly created: Flight. Kerosene-type jet fuel propulsion has not been observed in nature. It is flight nonetheless.
Human flight is not as agile or energy-effective as a dragonfly, but it is faster and stronger. Just like artificial learning may not be as sample-effecient as the human brain. It is a learning intelligence nonetheless and we are already working with the core mechanisms of reasoning and deduction.
If AGI is possible, it already happened. If even AI experts put it a 100-1000 years out, where some human monkeys banging on digital typewriters could eventually create it, then, in the vastness of space, time, military contracts, alien intelligences, and random Boltzmann brains, it must have been reality multiple times already.
If AGI is impossible, it will never happen. We already know that perfectly intelligent AGI's are not physically possible: Per DeepMind's foundational theoretical framework, optimal compression is non-computable, and besides that, it is not possible for an inference machine to know all of its universe (unless it is bigger than the universe by at least 1 bit, AKA it is the universe).
Remains being more intelligent than all of humanity. To accomplish that, by Shannon's own estimates, there is currently not enough information available in datasets and the internet. Chinese efforts to artificially increase the intelligence of babies is still in its infancy too (the substrate of AGI is irrelevant for computationalism, unless it absolutely needs to run on the IBM 5100).
So until that time travels, we will have to make due with being smarter than/indistinguishable from a human on all economic tasks. We're already there for some subset of humanity, you may even be a part of that subset, if you believed this post was written by a human.