The Cambrian Period of AI
lachlangray.blot.im
lachlangray.blot.im
Has anyone has tried searching for new basic operations, below the level of neural networks? We've been using these methods for years, and I doubt the first major breakthrough in ML is the most optimal method possible.
Consider the extreme case of searching over all mathematical and logical operations to see if something really novel can be discovered.
How feasible would this be?
>Tying it together, a big diversification is likely coming
I'd be interested in knowing what signs, if any, currently point to that.
What I see is a coalescing around the current paradigm. And considering that the hardware side led by NVIDIA is catering to this paradigm, I don't see us breaking out of this local optimum any time soon.
Why would we? It's proven that neural networks can learn arbitrary functions, and deep networks can apparently learn all the things we don't have functions for. Backpropagation is probably biologically implausible but it's a very good and fast method to train such networks. What could another implementation of intelligence offer that we don't already have? Backpropagation is very old and maybe it's not optimal but it s good enough that it doesnt matter
Re the question of other methods, Jeffrey Hinton proposed what I think is called a "forward-forward" alternative to back propagation somewhat recently. I don't think it's been proven to be better in any way, but it shows people are searching for new methods.
Sure it does, it's practical for pretty much any problem we want to solve, in fact i can't think of something it has not been shown to be practical about. Gradient descent is pretty straightforward and the most efficient generic method we have, even though it's not proven to be the fastest. Forward-forward is proposed as an alternative that is biologically plausible, not as an efficient alternative (it isnt)
Gradient descent is only used for NN because we initialize all the parameters to random values, so the assumption of the JtJ approximating the Hessian (required by LM) doesn't hold, since residuals are too high. I suspect there might also be something about the learning rate effectively being too high when using quadratic descent.
Even for large problems there are techniques like LBFGS which converge a lot faster than gradient descent, that don't require O(N^2) memory on the number of parameters.
> Can learn arbitrary functions says nothing about efficiency or any practical considerations.
This is absolutely true. A turing machine can perform any computation, but they are not really practical for anything outside of an academic exercise.
> A vanilla neural network can't learn a sine function or anything periodic, for inputs outside the range of the training data.
This is also true. A vanilla neural network isn't that useful outside of very basic tasks with a lot of training data. It's only when you start changing their shapes and using specialized techniques to fit the problem that they become practical. Even a simple task like classifying MNIST digits is done far more efficiently using a basic CNN than a vanilla NN.
Wondering where I saw that pattern before…
From everything we do know, what else would it be? The arrogance of consciousness is our blind-spot sometimes.
Even the words my brain just typed are combinations of words and arguments from everything you’ve ever agreed or disagreed with.
So, in my view. Traceability in AI is possible, but so should traceability in neurology be possible. It might not be as important as it seems to us. And that humility hits people in various ways.
My idea to calm everyone down about traceability is to have the AI write a log/wiki in parallel with its activities.
Putting toothpaste back in a tube has never worked, so don’t light your own hair on fire, right?
The confetti has left the cannon.
https://yoheinakajima.com/task-driven-autonomous-agent-utili...
In a way, you can also look at language models as being the new basic operation. Activations and floating point math are replaced by words and symbolic reasoning.
I would really appreciate any feedback on clarity, or anything that causes good or bad gut reactions.
Happy blogging!
> The specificity matters here because each innovation in the chain required not merely the discovery of the principle, but also the design and an economically viable use-case to all line up in order to have impact. The steam engine is an excellent example of this problem. Early tinkering with the idea of using heat to create steam to power rotary motion – the core function of a steam-engine – go all the way back to Vitruvius (c. 80 BC -15 AD) and Heron of Alexandria (c. 10-70 AD). With the benefit of hindsight we can see they were tinkering with an importance principle but the devices they actually produced – the aeolipile – had no practical use – it’s fearsomely fuel inefficient, produces little power and has to be refilled with water (that then has to be heated again from room temperature to enable operation).
He goes on to say that the very first "commercial" steam engine only happened to be commercially useful because of the particular situation in which it was invented: England had cut down most of their trees, but luckily had lots of coal. The engine wasn't quite useful enough to use to (say) pump water out of an iron mine, because it was so resource-hungry that it the cost of the fuel plus getting fuel to the engine would be too much. But it's just barely useful enough to pump water out of a coal mine, if you can provide it coal without having to transport it. And that gave it just enough of a commercial toe-hold to fund its optimization.
It sounds like "dumb AI" of the 2000's may have performed a similar function; and we are, perhaps, on the edge of the "AI revolution", similar to the Industrial Revolution, where we reach a hockey-stick proliferation.
EDIT: Fixed link
[1] https://acoup.blog/2022/08/26/collections-why-no-roman-indus...
I think GPT-4 was enough of a leap over GPT3.5 that I'm not sure we've hit that point yet but it'll be interesting if the next GPT is less of a leap.
It turns out that consumer protection around these "new technologies" is basically non-existent, allowing corporations to charge people for fantastic sounding ideas that they know for a fact are not and are not going to be practical within the lifespan of the item they just sold you.
> I think GPT-4 was enough of a leap over GPT3.5
Compared to what, though? The cost of training? The cost of running the model? The accuracy of output? The increase in real world utility?
It's the "Wild West Snake Oil" period of AI.
I'm not a lawyer, doctor, etc, so I can't speak for what the impact might be in those fields but I think at the very least we're looking at quite striking incremental improvements on existing processes.
I think the exciting stuff is just barely ahead of us. Things like better integration with home assistants so that Alex actually gets my commands correct. Things like integration with first gen robotics. Consider this prompt:
Me: Assume I have a device called a Roomba which is a robotic vacuum cleaner which is a small disk shaped object that vacuums an area and returns to a dock platform. You are an AI that runs the household and have the power to active the Roomba and send it to clean either the entire house or a specific room. I will give you a list of parsed voice commands, please take your best guess at what each might mean and if they are relevant or something you can do with the Roomba.
* "yo, why is the floor dirty? Get cleanin"
* "please turn on the vacuum"
* "Is woobie working?"
* "Omigod I just stepped on a chip, you need to clean this up now"
* "Why is it so loud in here? Who turned on the hockey puck?"
* "Yo AI did the Room thing run today? The floor still looks sticky".
ChatGPT:
"yo, why is the floor dirty? Get cleanin" - This command could be interpreted as a request to clean the floor. As an AI that runs the household, I could activate the Roomba to clean the entire house or a specific room based on the user's preference. However, the command is not specific about which action to take, so I might clarify with the user before activating Roomba.
"please turn on the vacuum" - This command is a clear request to start the Roomba's cleaning process. As an AI that runs the household, I would activate Roomba to clean the entire house or a specific room based on the user's preference.
"Is woobie working?" - This command is not relevant to Roomba, and as an AI that runs the household, I would inform the user that Roomba is not named "woobie."
"Omigod I just stepped on a chip, you need to clean this up now" - This command is a clear request for immediate cleaning in response to a mess caused by someone stepping on a chip. As an AI that runs the household, I could activate the Roomba's spot cleaning feature or manually control Roomba to clean the specific area where the chip is present.
"Why is it so loud in here? Who turned on the hockey puck?" - This command is not clear about what the speaker wants to accomplish. It mentions a loud noise and a "hockey puck," which could refer to the Roomba. However, it is unclear if the speaker wants the Roomba to stop cleaning or if they are just making a comment. As an AI that runs the household, I might ask the user for clarification before taking any action.
"Yo AI did the Room thing run today? The floor still looks sticky." - This command is a question about whether the Roomba has cleaned the room today and a comment on the current state of the floor. As an AI that runs the household, I could check the logs of the Roomba to see if it has cleaned the room today, and if not, activate the Roomba to clean the room. Alternatively, I could instruct the Roomba to spot clean the area where the floor is still sticky.
------
If you can't see how having an AI that can perform reasoning like demonstrated above is extremely useful then, I dunno, just wait and see what comes.
The sky's really the limit if you consider what can be done with language and text and how much of our world revolves around text.
Edit: formatting
Is there any objectivity or measurements in your analysis?
> at the very least we're looking at quite striking incremental improvements
I'm not even sure what this should mean. This sounds like something I would find in a marketing pamphlet.
> "yo, why is the floor dirty? Get cleanin"
From.. possible striking improvements to the legal profession.. to.. a conversational mode embedded in my Roomba. You've actually perfectly captured the gulf between what this technology is marketed to be and what it is barely capable of doing.
"You are an auditor/lawyer looking for suspicious conversations or patterns of deceit by employees of company Z. I will feed you documents and you will give your analysis of each document or collections of documents. Pay special attention to X Y Z. Consider these sorts of actions as deceptive (a,b,c,d). If you find anything of note, create a summary of the findings in the form ['wording indicates deceptive intent', document 37, page 93].".
You'd then feed it 20 pages at a time and for less than $100 you've scanned a mountain of legal documents in a way in which you'd normally require a team of legal assistants and 100s of hours of work.
Again, I'm not a lawyer but I can conjure this very simple example that I could implement in a month if I had the time. At this point I think you're just squeezing me for ideas.
Perhaps you should explain what the technology is marketed as that you see as lying or misrepresenting. That would make this more productive.
I think that if we consider the accuracy rate as measured in various ways being roughly that of a human, then you're trading human mistakes for AI mistakes in exchange for dramatically lower costs and a dramatically higher speed of processing. You might even say a higher level of reasoning. In my own interactions it's been fantastic at reasoning clearly and quickly outside of complex trick questions. Most scenarios in life aren't generally full of trick questions.
To some up my experience with basically all these products...YMMV.
No, I simply disagree with your assessment of the technology and you are unwilling to admit you could be wrong. It's a very rude way to start a conversation with a stranger.
> Each time something new is discovered, the goal post has moved.
This isn't anything new, it's merely a large scale version of technology that has existed for more than 30 years powered by training methods too questionable to be seriously utilized previously. People misunderstanding any of this as "invention" is a large source of the problem, I think.
> In previous century billions were spent on trying to automate transation to other natural languages, without good results.
Why were those billions spent? What entities spent that money? Was it consumer driven?
> Today an LLM can do this mere as part of a side effect
Not particularly well, and the difference in cost between making GPT3 and GPT4 compared to the difference in their accuracy paints a bleak future for continued scaling of this methodology. It is, quite literally, all downhill from here.
> Non-natural languages as a bonus.
If you're willing to blindly trust that output. Which is why I refer to this as "snake oil." It really does appear to be the same phenomenon. People are all excited about what the label says it can do, they forgot to measure if it actually does that at a level that justifies it's cost.
GPT-4 is already practical (I use it every day and it speeds up my workflow drastically in some cases), and most areas it'll assist in aren't regulated in any way. I don't see it making much of a dent in medical or law in the near term for similar reasons as self driving, but the rest ought to go way faster. It's not vaporware, shit works.
Then the cracks began to show...
The difference is that GPT has had some successful demos that look pretty amazing. Time will tell if we have merely invented a more clever Clever Hans.
People swooning over expected change is usually a pretty good signal reality will fall short.
This is a bit different.
But no one was there in the Cambrian period. We have no idea what the circumstances really were. I'm not dismissing historical sources. Just saying the timescale and comparative lack of sources equate to a lot of unreliability.
The real question is how much AI is a step up in tooling. The fact that humans develop tools using their intelligence is well documented. And doesn't require a detour into evolutionary theory.
The society with the best tools for the problems it faces has a competitive advantage.