There are people interested in overcoming aphantasia (or hypophantasia, an extremely weak form of imagination).
Today I have medium-ish hypophantasia, but I remember when I was doing phantasia exercises, in particular "snapshotting" and "memory streaming", at least two times, there was a subtle shift in my perception and all of a sudden I could remember a ton of things, as if I opened a door. It would only last maybe 10-20 minutes (I would practice 30 - 60 minutes per day).
It wouldn't surprise me a bunch of those memories are in the brain but you just don't have access to them in everyday waking consciousness.
So, for me, it feels like a lot of my memories are visually indexed, and if I can't visualize then I can't remember, but once I configure my mind through meditation and these exercises, it is like I can "tune my mind" to mind's eye access (radio/TV analogy here) and with it the memories.
Then once I stopped the exercise (for the day) it would go away in around 10-20 minutes (kind of like how a muscle pump goes away rather quickly after exercising).
It looks like these models work pretty well as natural language search engines and at connecting together dots of disparate things humans haven't done.
They're finding them very effective at literature search, and at autoformalization of human-written proofs.
Pretty soon, this is going to mean the entire historical math literature will be formalized (or, in some cases, found to be in error). Consider the implications of that for training theorem provers.
I think "pretty soon" is a serious overstatement. This does not take into account the difficulty in formalizing definitions and theorem statements. This cannot be done autonomously (or, it can, but there will be serious errors) since there is no way to formalize the "text to lean" process.
What's more, there's almost surely going to turn out to be a large amount of human generated mathematics that's "basically" correct, in the sense that there exists a formal proof that morally fits the arc of the human proof, but there's informal/vague reasoning used (e.g. diagram arguments, etc) that are hard to really formalize, but an expert can use consistently without making a mistake. This will take a long time to formalize, and I expect will require a large amount of human and AI effort.
It's all up for debate, but personally I feel you're being too pessimistic there. The advances being made are faster than I had expected. The area is one where success will build upon and accelerate success, so I expect the rate of advance to increase and continue increasing.
This particular field seems ideal for AI, since verification enables identification of failure at all levels. If the definitions are wrong the theorems won't work and applications elsewhere won't work.
Hassabis put forth a nice taxonomy of innovation: interpolation, extrapolation, and paradigm shifts.
AI is currently great at interpolation, and in some fields (like biology) there seems to be low-hanging fruit for this kind of connect-the-dots exercise. A human would still be considered smart for connecting these dots IMO.
AI clearly struggles with extrapolation, at least if the new datum is fully outside the training set.
And we will have AGI (if not ASI) if/when AI systems can reliably form new paradigms. It’s a high bar.
Maybe if Terence Tao had memorized the entire Internet (and pretty much all media), then maybe he would find bits and pieces of the problem remind him of certain known solutions and be able to connect the dots himself.
But, I don't know. I tend to view these (reasoning) LLMs as alien minds and my intuition of what is perhaps happening under the hood is not good.
I just know that people have been using these LLMs as search engines (including Stephen Wolfram), browsing through what these LLMs perhaps know and have connected together.
The strength of the glasses alters the size of the image that lands on your retinas. More (-) means a smaller image thus you stop seeing movement much closer to the focal point.
Halting is sometimes preferable to thrashing around and running in circles.
I feel like if LLMs "knew" when they're out of their depth, they could be much more useful. The question is whether knowing when to stop can be meaningfully learned from examples with RL. From all we've seen the hallucination problem and this stopping problem all boil down to this problem that you could teach the model to say "I don't know" but if that's part of the training dataset it might just spit out "I don't know" to random questions, because it's a likely response in the realm of possible responses, instead of spitting out "I don't know" to not knowing.
SocratesAI is still unsolved, and LLMs are probably not the path to get knowing that you know nothing.
> if LLMs "knew" when they're out of their depth, they could be much more useful.
I used to think this, but no longer sure.
Large-scale tasks just grind to a halt with more modern LLMs because of this perception of impassable complexity.
And it's not that they need extensive planning, the LLM knows what needs to be done (it'll even tell you!), it's just more work than will fit within a "session" (arbitrary) and so it would rather refuse than get started.
So you're now looking at TODOs, and hierarchical plans, and all this unnecessary pre-work even when the task scales horizontally very well (if it just jumped into it).
I would consider that detecting your own limits when trying to solve a problem is preferable to having the illusion of thinking that your solution is working and correct.
Ah yes, the function that halts if the input problem would take too long to halt.
But yes, I assume you mean they abort their loop after a while, which they do.
This whole idea of a "reasoning benchmark" doesn't sit well with me. It seems still not well-defined to me.
Maybe it's just bias I have or my own lack of intelligence, but it seems to me that using language models for "reasoning" is still more or less a gimmick and convenience feature (to automate re-prompts, clarifications etc, as far as possible).
But reading this pop-sci article from summer 2022 seems like this definition problem hasn't changed very much since then.
Although it's about AI progress before ChatGPT and it doesn't even mention the GPT base models. Sure, some of the tasks mentioned in the article seem dated today.
But IMO, there is still no AI model that can be trusted to, for example, accurately summarize a Wikipedia article.
Not all humans can do that either, sure. But humans are better at knowing what they don't know, and deciding what other humans can be trusted. And of course, none of this is an arithmetic or calculation task.
> Yes machine learning vision systems hallucinate, but so do humans.
When was the last time you had full attention on the road and a reflection of light made you super confused and suddenly drive crazy? When was the last time you experienced objects behaving erratically around you, jumping in and out of place, and perhaps morphing?
Well there is strong anecdotal evidence of exactly this happening.
We were somewhere around Barstow on the edge of the desert when the drugs began to take hold. I remember saying something like, “I feel a bit lightheaded; maybe you should drive . . .”And suddenly there was a terrible roar all around us and the sky was full of what looked like huge bats, all swooping and screeching and diving around the car, which was going about 100 miles an hour with the top down to Las Vegas. And a voice was screaming: “Holy Jesus! What are these goddamn animals?” [0]
[0] Thompson, Hunter S., „Fear and Loathing in Las Vegas“
One hopes so. Many of the comments assume an ideal human driver, whereas real human drivers are frequently tired, distracted, intoxicated, or just crazy.
Many accidents are caused by low-angle light dazzle. It's part if why high beams aren't meant to be used off a dual carriageway.
When was the last time you saw a paper bag blown across the street and mistook it for a cat or a fox? (Did you even notice your mistake, or do you still think it was an animal?)
Do you naturally drive faster on wide streets, slower on narrow streets, because the distance to the side of the road changes your subconcious feeling of how fast you're going? Do you even know, or are you limited to your memories rather than a dashcam whose footage can be reviewed later?
etc.
Now don't get me wrong, AI today is, I think, worse than humans at safe driving; but I'm not sure how much of that is that AI is more hallucinate-y than us vs. how much of it is that human vision system failures are a thing we compensate for (or even actively make use of) in the design of our roads, and the AI just makes different mistakes.
If the internal representation of Tesla Autopilot is similar to what the UI displays, i.e. the location of the w.r.t. to everything else, and we had a human whose internal representation is similar, everything jumping around in consciousness, we’d be insane to allow him to drive.
Self-driving is probably “AI-hard” as you’d need extensive “world knowledge” and be able to reason about your environment and tolerate faulty sensors (the human eyes are super crappy with all kinds of things that obscure it, such as veins and floaters).
Also, if the Waymo UI accurately represents what it thinks is going on “out there” it is surprisingly crappy. If your conscious experience was like that when you were driving you’d think you had been drugged.
I agree that if Tesla's representation of what their system is seeing is accurate, it's a bad system.
The human brain's vision system makes pretty much the exact opposite mistake, which is a fun trick that is often exploited by stage magicians: https://www.youtube.com/watch?v=v3iPrBrGSJM&pp
I wonder what we'd seem like to each other, if we could look at each other's perception as directly as we can look at an AI's perception?
Most of us don't realise how much we mispercieve because it doesn't feel different in the moment to percieve incorrectly; it can't feel different in the moment, because if it did, we'd notice we were mispercieving.
Honestly, on an iPhone, I wish I could completely mute/disable the Phone app, so when it calls it doesn't overtake whatever you are doing, or just delete it.
reply