> we could probably agree that human intelligence is Turing-complete (with a slightly sloppy use of terms).
> So any other Turing-complete model can emulate it
You're going off the rails IMMEDIATELY in your logic.
Sure, one Turing-complete computer language can have its logic "emulated" by another, fine. But human intelligence is not a computer language -- you're mixing up the terms "Turing complete" and "Turing test".
It's like mixing up the terms "Strawberry jam" and "traffic jam" and then going on to talk about how cars taste on toast. It's nonsensical.
Game of life, PowerPoint, and a bunch of non-PL stuff are all Turing-complete. I don't mix up terms, I did use a slightly sloppy terminology but it is the correct concept - and my point is that we don't know of a computational model that can't be expressed by a Turing-machine, humans are a physical "machine", ergo we must also fall into that category.
Give my comment another read, but it was quite understandable from context. (Also, you may want to give a read to the Turing paper because being executable by a person as well was an important concept within)
Again, you're going wildly off the rails in your logic. Sure, "executable by a human" is part of the definition for Turing machines, but that's only talking about Turing-specific capabilities. If you want to argue that a Turing machine can emulate the specific definition of Turing machine capabilities that humans can perform, that's fine. But you're saying that because humans can ACT LIKE Turing machines, they must BE Turing machines, and are therefore emulatable.
This is the equivalent of saying "I have set up a complex mechanical computer powered by water that is Turing complete. Since any Turing complete system can emulate another one, it means that any other Turing complete system can also make things wet and irrigate farms.
Human intelligence is not understood. It can be made to do Turing complete things, but you can't invert that and say that because you've read the paper on Turing completeness, you now understand human intelligence.
That's not a computation, it's a side effect. It just depends on what you wire your "computer" up to. A Turing machine in itself is just a (potentially non-returning) mathematical function, but you are free to map any input/output to it.
Actually, the way LLMs are extended with tools is a pretty much the same (an LLM itself has no access to the internet, but if it returns some specific symbols, the external "glue" will do a search and then the LLM is free to use the results)
You can build this today exactly as efficiently as you can when inference is 1000x faster, because the only things you can build with this is things that absolutely don't matter. The first bored high schooler who realizes that there's an LLM between them and the database is going to WRECK you.
It’s actually extremely motivating to consider what LLM or similar AI agent could do for us if we were to free our minds from 2 decades of SaaS brainrot.
What if we ran AI locally and used it to actually do labor-intensive things with computers that make money rather than assuming everything were web-connected, paywalled, rate-limited, authenticated, tracked, and resold?
That sucks if that's your experience, but it's not the universal, or even the common, experience.
For reference, I get a sore-ish shoulder the next day, and that's it. Also for reference, when I got Actual Covid, I was knocked on my ass for almost two weeks. So for me, at least, the choice is easy.
It's my unfortunate experience, when I've had covid its a 6-12 hour affair that happens once every 12-24 months. My 3rd vaccine shot had me in bed for 3 days. Leading to continued vaccination being unsustainable. My wife has a similar experience to yours, and gets moderate to severe covid. She gets the vaccine every year to help avoid it - but still gets moderate COVID roughly once per 6 months.
It's unfortunate that the vaccine has such radically different outcomes within a single household, if it was a flu shot like experience I'd happily get it once per year.
COVID is a nasty virus. I need my brain way to much to FAFO.
COVID-19 may Enduringly Impact Cognitive Performance and Brain Haemodynamics in Undergraduate Students - ScienceDirect https://share.google/49ER4VjJUwipGotZO
Flu shot experience varies too. The last several have been very low response, but the first few were a miserable couple days and I stopped getting them because certain misery was worse than a chance of misery that I'd never know if it was flu or not, because testing was inaccessible.
Last year I skipped the flu vac (I had a zillion for tropical diseases so I though come one not another one) and lo, I got a flu about every 4 weeks, so like over 6 the whole season. I'm on a way to get it this year.
At least testably/symptomatically, I'm asthmatic as well - so it's surprising that the impact is so small. My wife gets it for 1-2 weeks whenever she comes down with it.
As a data point, my experience with the shot was a sore arm and chills for a couple days.
When I got Covid later, it was slightly worse chills for 3 days. By the 4th time I got Covid, it was just chills for a day.
If I knew that would be the experience, I'd probably have skipped it. That said, it's completely possible it was having the vaccine that made getting real Covid not so bad.
By the time it was my turn to get Covid I’d been twice vaccinated. It’s the most exhausted I can remember ever feeling. Let me tell you, the whole time I kept thinking: How much more miserable would this have been without the vaccine to blunt the impact? Felt grateful and humbled
You said: "When I got Covid later, it was slightly worse chills for 3 days. By the 4th time I got Covid, it was just chills for a day. If I knew that would be the experience, I'd probably have skipped it."
I'm saying that's not an apples to apples comparison due to the growing evidence of how much long term damage a COVID infection can cause.
Ah I see, thanks. Yep, it's definitely not apples to apples in either event. As in, not having the vaccine could have made getting it, at least the first time, way way worse to deal with.
This is incorrect for a lot of reasons, many of which have already been explored, but also:
> with every new iteration of the AI, the internal code will get better
This is a claim that requires proof; it cannot just be asserted as fact. Especially because there's a silent "appreciably" hidden in there between "get" and "better" which has been less and less apparent with each new model. In fact, it more and more looks like "Moore's law for AI" is dead or dying, and we're approaching an upper limit where we'll need to find ways to be properly productive with models only effectively as good as what we already have!
Additionally, there's a relevant adage in computer science: "Debugging is twice as hard as writing the code in the first place. Therefore, if you write the code as cleverly as possible, you are, by definition, not smart enough to debug it." If the code being written is already at the frontier capabilities of these models, how the hell are they supposed to fix the bugs that crop up, especially if we can't rely on them getting twice as smart? ("They won't write the bugs in the first place" is not a realistic answer, btw.)
Just because you're not writing code where you can see that the new models are appreciably better doesn't mean they aren't. LLM progress now isn't in making it magically appear smarter at the top end (that's in diminishing returns as you imply), but at filling in weak points in knowledge, holes in capability, improving default process, etc. That's relevant because it turns out most of the time the LLM doesn't fail at coding because it's not a general super genius, but because it just had a hole in its capabilities that caused it to be dumb in a specific scenario.
Additionally, while the intelligence floor is shooting up and the intelligence ceiling is very slowly rising, the models are also getting better at following directions, writing cleaner prose, and their context length support is increasing so they can handle larger systems. The progress is still going strong, it just isn't well represented by top line "IQ" style tests.
LLMs and humans are good at dealing with different kinds of complexity. Humans can deal with messy imperative systems more easily assuming they have some real world intuition about it, whereas LLMs handily beat most humans when working with pure functions. It just so happens that messy imperative systems are bad for a number of reasons, so the fact that LLMs are really good at accelerating functional systems gives them an advantage. Since functional systems are harder to write but easier to reason about and test, this directly addresses the issue of comprehending code.
The argument they are making is that if a bug is discovered, the agent will not debug it, instead a new test case is created, and the code is regenerated (I suppose if a quick fix isn't found). That is why they don't need debugging agent twice as capable as coding agent. I don't know if this works in practice, as in my experience, tests are intertwined with the code base.
I forget where I read it originally, but I strongly feel that AWS should offer a `us-chaos-1` region, where every 3-4 days, one or two services blow up. Host your staging stack there and you build real resiliency over time.
(The counter joke is, of course, "but that's `us-east-1` already! But I mean deliberately and frequently.)
It may be a bad metric, if you're being proscriptive, but it's a great heuristic. If I see a 500-line function where I'm not expecting one, I'm going to pay a lot more attention to it in the PR to try to figure out just why it isn't shorter.
The thing about this, though - cars have been built before. We understand what's necessary to get those 9s. I'm sure there were some new problems that had to be solved along the way, but fundamentally, "build good car" is known to be achievable, so the process of "adding 9s" there makes sense.
But this method of AI is still pretty new, and we don't know it's upper limits. It may be that there are no more 9s to add, or that any more 9s cost prohibitively more. We might be effectively stuck at 91.25626726...% forever.
Not to be a doomer, but I DO think that anyone who is significantly invested in AI really has to have a plan in case that ends up being true. We can't just keep on saying "they'll get there some day" and acting as if it's true. (I mean you can, just not without consequences.)
While you are right about the broader (and sort of ill defined) chase toward 'AGI' - another way to look at it is the self driving car - they got there eventually.And, if you work on applications using LLMs you can pretty easily see that Karpathy's sentiment is likely correct. You see it because you do it. Even simple applications are shaped like this, albeit each 9 takes less time than self driving cars for a simple app.. it still feels about right.
> another way to look at it is the self driving car - they got there eventually
Current self driving cars only work in American roads. Maybe Canada too, not sure how their roads are. Come to Europe/anywhere else and every other road would be intractable. Much tighter lanes, many turns you have a little mirror to see who's coming on the other side, single car at a time lanes that you need to "understand" who goes first, mountain roads where you sometimes need to reverse for 100m when another car is coming so it's wide enough that they can pass before you can keep going forward, etc.
Many things like this that would require another 2 or 3 "nines" as the guy put it than acceptable quality in American huge roads.
Waymo has promised to launch In London and Tokyo next year. New York, London, Tokyo probably covers the entire spectrum of difficulty for self driving cars, maybe we need to include Mumbai as the final boss but I would be happy saying self driving is solved if the above 3 cities have a working 24/7 self driving fleet
The final boss could be something like Scottland mountain roads, or some of the million beaches on a cliff in Greece where this "you have to first reverse" kinda situation happens every 30 seconds.
No no, the final boss is a dozen of them have to race around such roads, no crashes allowed, and they must all finish at exactly the same time. Blindfolded.
Give the Waymo guys some credit - San Francisco isn't the suburbs of Houston. It might not be quite the same as a 1000 year old city in Europe, but it's no snack either.
> another way to look at it is the self driving car - they got there eventually.
No they did not. Elon has been saying Tesla will get there “next year” since 2015. He is still saying that, and despite changing definitions, we still are not there.
Karpathy talked about Waymo, and he said they aren't there yet. They still have humans in the loop via telemetry and there are parts of cities they won't go to.
Karpathy is biased when it comes to self driving. Example: You can't both have humans teleoperating like he claimed, AND have cones disabling Waymos. Waymo's mistakes such as one where they were going around a parking lot honking at each other tells you they don't have humans in the loop except in the most extreme cases. He's likely correct that it's not self driving 100% of the time, but what if it's 99.999% and in the 0.001% the humans have to tell the Waymo how to get out of a very tricky situation?
i guess the comment you replied proves the actual point "we may never get there, but it will be enough for the market".
sigh, i guess it's time to laugh on that video compilation of elon saying "next week" for 10yrs straight and then cry seeing how much he made of doing that.
