Apple's "strategic vision" for AI is to add a computer use agent (AI assistant) to the OS to perform tasks on behalf of the user, plus contextually surface AI capabilities in many specific contexts they've got utility (copy editing, image generation, photo organization, translation, coding).
What's missing here? What else should they be doing? What are their competitors doing, in any space relevant to their markets, that's much different? None of these critiques ever seem to say.
If AI ends up being another 'normal' technology, Apple's advantages in distribution (~2B active devices, with a user base that installs updates pretty reliably), ability to give their AI tools access to your existing data and apps, and general facility with packaging tech so consumers actually understand what it's good for, put them in an extremely strong position to capture value from it.
If AI ends up being something other than a 'normal' technology, if we really are a few years from building the sand god, well, all bets are off, and it's a little silly to evaluate the strategic planning of an individual company against that backdrop.
I feel that any vision needs to have some actual execution behind it, otherwise it's just a twinkle in someone's eye. "Built for Apple Intelligence" is literally the slogan for Apple's 2024 lineup. The vision is there front and center. But everything they layed out in their vision (and explicitly advertised!) has been a joke so far, falling somewhere between half-baked, trivial, or nonexistent. They've had to pull that ad showing contextually-aware interactions because that's nowhere near ready. https://www.techradar.com/computing/artificial-intelligence/...
After a certain point, it becomes a "put up or shut up" situation for those making wild claims. That's where all the criticism is coming from, and rightfully so. Sure, set a course for the future, but until there's something real to show in the present, it's all empty hype until proven otherwise.
The material finding of this paper is that reasoning models are better than non-reasoning models at solving puzzles of intermediate complexity (where that's defined, essentially, by how many steps are required), but that performance collapses past a certain threshold. This threshold differs for different puzzle types. It occurs even if a model is explicitly supplied with an algorithm it can use to solve the puzzle, and it's not a consequence of limited context window size.
The authors speculate that this pattern is a consequence of reasoning models actually solving these puzzles by way of pattern-matching to training data, which covers some puzzles at greater depth than others.
Great. That's one possible explanation. How might you support it?
- You could systematically examine the training data, to see if less representation of a puzzle type there reliably correlates with worse LLM performance.
- You could test how successfully LLMs can play novel games that have no representation in the training data, given instructions.
- Ultimately, using mechanistic interpretability techniques, you could look at what's actually going on inside a reasoning model.
This paper, however, doesn't attempt any of these. People are getting way out ahead of the evidence in accepting its speculation as fact.
While I agree overall, LLMs are pattern matching in a complicated way.
You transform your training data in a very strange and high dimensional space. Then when you write an input, you calculate the distance between that input and the closest point in that space.
So, in some sense.. You pattern match your input with the training data. Of course, in a very non intuitive way for humans.
Now, it doesn't necessarily imply things as 'models cannot solve new problems not seen before' we don't know if our problem could get matched to something completely unrelated for us, but in that space it makes sense.
So with your experiments, if the model is able to solve a new puzzle never seen before, you'll never know why, but it doesn't imply either that the new puzzle was not matched in some sense to some previous data in the dataset.
I think Moravec's Paradox is often misapplied when considering LLMs vs. robotics. It's true that formal reasoning over unambiguous problem representations is easy and computationally cheap. Lisp machines were already doing this sort of thing in the '70s. But the kind of commonsense reasoning over ambiguous natural language that LLMs can do is not easy or computationally cheap. Many early AI researchers thought it would be — that it would just require a bit of elaboration on the formal reasoning stuff — but this was totally wrong.
So, it doesn't make sense to say that what LLMs do is Moravec-easy, and therefore can't be extrapolated to predict near-term progress on Moravec-hard problems like robotics. What LLMs do is, in fact, Moravec-hard. And we should expect that if we've got enough compute to make major progress on one Moravec-hard problem, there's a good chance we're closing in on having enough to make major progress on others.
Leaving aside the lack of consensus around whether LLMs actually succeed in commonsense reasoning, this seems a little bit like saying “Actually, the first 90% of our project took an enormous amount of time, so it must be ‘Pareto-hard’. And thus the last 10% is well within reach!” That is, that Pareto and Moravec were in fact just wrong, and thing A and thing B are equivalently hard.
Keeping the paradox would more logically bring you to the conclusion that LLMs’ massive computational needs and limited capacities imply a commensurately greater, mind-bogglingly large computational requirement for physical aptitude.
It's far from obvious that thought space is much less complex than physical space. Natural language covers emotional, psychological, social, and abstract concepts that are orthogonal to physical aptitude.
While the linguistic representation of thought space may be discrete and appear simpler (even the latter is arguable), the underlying phenomena are not.
Current LLMs are terrific in many ways but pale in comparison to great authors in capturing deep, nuanced human experience.
As a related point, for AI to truly understand humans, it will likely need to process videos, social interactions, and other forms of data beyond language alone.
I think the essence of human creativity is outside our brains - in our environments, our search spaces, our interactions. We stumble upon discoveries or patterns, we ideate and test, and most ideas fail but a few remain. And we call it creativity, but it's just environment tested ideation.
If you put an AI like AlphaZero in a Go environment it explores so much of the game space that it invents its own Go culture from scratch and beats us at our own game. Creativity is search in disguise, having good feedback is essential.
AI will become more and more grounded as it interacts with the real world, as opposed to simply modeling organic text as GPT-3. More recent models generate lots of synthetic data to simulate this process, and it helps up to a point, but we can't substitute artificial feedback for real one except in a few cases: like AlphaZero, AlphaProof, AlphaCode... in those cases we have the game winner, LEAN as inference engine, and code tests to provide reliable feedback.
If there is one concept that underlies both training and inference it is search. And it also underlies action and learning in humans. Learning is compression which is search for optimal parameters. Creativity is search too. And search is not purely mental, or strictly 1st person, it is based on search spaces and has a social side.
Good points. Came here to say pretty much the same.
Moravec's Paradox is certainly interesting and correct if you limit its scope (as you say). But it feels intuitively wrong to me to make any claims about the relative computational demands of sensi-motor control and abstract thinking before we’ve really solved either problem.
Looking e.g. at the recent progress in solving ARC-AGI my impression is that abstract thought could have incredible computational demands. IIRC they had to throw approximately $10k of compute at o3 before it reached human performance. Now compare how cognitively challenging ARC-AGI is to e.g. designing or reorganizing a Tesla gigafactory.
With that said I do agree that our culture tends to value simple office work over skillful practical work. Hopefully the progress in AI/ML will soon correct that wrong.
The US has a 'Do Not Call' registry for unsolicited phone calls, but technically doesn't need one for texts because it's illegal to send marketing texts without prior consent in the first place. Thing is, 'consent' often just means failing to notice a checkbox during a signup flow or something, so people end up getting junk anyway.
Even more annoyingly, politicians wrote in an exception for themselves. In combination with the way campaign finance works in the US, this means that if you've ever give your number to any political campaign, it will be passed around forever and you'll have multiple politicians begging you for money for months leading up to every election. Each individual campaign/organization seems to respect 'STOP,' but once your number is on an e.g. 'Has ever donated to a Democratic candidate' list, there's seemingly no way to get it off for good. Thanks, Obama. (I gave him $50 in 2008.)
Even worse if someone else signs up somehow using your contact info. I got signed up (via email thankfully) for a political party in another country and no amount of "mark as spam", unsubscribe or replying would get me off the list. Eventually I just had to create a filter that dumps those messages in the trash.
It must be something with non-U.S. English speaking countries because I get numerous semi-spam messages in email and text for services in Australia and the U.K. casinos with account numbers or PINs, two step notifications for national car registries, banking, contractors asking about work or sending invoices. Maybe it's just English speaking countries have a lot of people named "iamthepieman"
My wife had someone do sign up for a bank account with my wife's gmail address. She told the bank they got it wrong, and they went away for a bit and then they re-signed up AGAIN. So she told the bank to close the account. It didn't re-occur after that.
A number of elderly folks have had this issue as well. I'm really at a loss on how to fix it, some times there are bad actors but generally it seems folks are clueless and the signup flow doesn't adequately account for this.
I have a common-ish first initial, last name Gmail account. The number of people who think they have my address is staggering. Hundreds over the years.
In one case, the manager of a large factory was forwarding me an email with remote access credentials and VPN software every month.
Is the email in question something along the lines of firstnamelastname at gmail? I'm guessing your email address is a really common name that someone else keeps forgetting how their email actually deviates, or someone typos writing theirs.
Another possible scenario is that Gmail is getting wires crossed. I have had the account firstname.lastname@gmail.com for 20 years now. About 5 years ago, some dude in Australia (who coincidentally has the same rare last name as me) started using firstnamelastname@gmail.com. Based on the emails I've seen I believe that Gmail let him do this for a while, but eventually started delivering his emails into my inbox. I don't know if there was a technical change in Gmail for how they handled these addresses or what, but it's very odd.
firstname.lastname@gmail.com and firstnamelastname@gmail.com are the same address, according to gmail documentation. If this is what is actually happening (and there isn't a subtle typo, etc.), then something is more wrong than "wires crossed" & you should report it as a security vulnerability.
I still get 5-10 texts a day from trumpy candidates because someone used my number like 5 years ago when they were spamming signups for trump rallies so the rally would be empty
>this means that if you've ever give your number to any political campaign, it will be passed around forever and you'll have multiple politicians begging you for money for months leading up to every election
They really should learn to not do that, my carrier routes most of those to spam already and the few that it doesn't, I mark as spam, so presumably they'll start getting routed to spam for other people with the same carrier.
What's worse is if someone accidentally uses your phone number when they sign up for something, then you're on the list and never able to get off of it.
If only we had the mobile numbers of numerous politicians. We could make a small donation to their opposing party and add a phone number from that last.
In his recent "Intelligence Age" post, Altman says superintelligence may be only a few thousand days out. This might, of course, be wrong, but skyrocketing demand for chips is a straightforward consequence of taking it seriously.
This is actually quite clever phrasing. "A few thousand days" is about ten years, assuming normal usage of 'few' (ie usually a number between 3 and 6 inclusive).
Now, if you, as a tech company, say "X is ten years away", anyone who has been around for a while will entirely disregard your claim, because forward-looking statements in that range by tech companies are _always_ wrong; it's pretty much a cliche. But phrasing as a few thousand days may get past some peoples' defences.
The mistake isn't thinking 'scaling is the solution to AGI'.
And the mistake isn't thinking more generally about 'the solution to AGI'.
The mistake is thinking about 'AGI'.
There will never be an artificial general intelligence. There will never artificial intelligence, full stop.
It's a fun concept in science fiction (and earlier parallel concepts in fantasy literature and folk tales). It's not and will never be reality. If you think it can be then either you are suffering from 'science fiction brain' or you are a fraud (Sam Altman) or you are both (possibly Sam Altman again).
Demand for compute will skyrocket given AGI even if AGI turns out to be relatively compute-efficient. The ability to translate compute directly into humanlike intelligence simply makes compute much more valuable.
Since AGI isn't here yet, the eventual implementation that breaks through might be based on different technology; for example, if it turns out to need quantum computing, investing lots of money to build out current fabs might turn out useless.
Input and output, given that they must connect with the physical world, seems to me to be the likely limiting resource, unless you think isolated virtual worlds will have value on to themselves
An AGI can presumably control a robot at least as well as a human operator can. The hardware side of robotics is already good enough that we could leverage this to rapidly increase industrial output. Including, of course, producing more AGI-controlled robots. So it may well be the case that robot production, rather than chip production, becomes the bottleneck on output growth, but such growth will still be extremely fast and will still drive demand for far more computing capacity than we're producing today.
And I suppose you are assuming that the robots will mine and refine the metal ore themselves, and then also dig the foundations for the factories that house their manufacturing?
Non-general AI won't cause mass unemployment, for the same reason previous productivity-enhancing tech hasn't. So long as humans can create valuable output machines can't, the new, higher-output economy will figure out how to employ them. Some won't even have to switch jobs, because demand for what they provide will be higher as AI tools bring down production costs. This is plausible for SWEs. Other people will end up in jobs that come into existence as a result of new tech, or that presently seem too silly to pay many people for — this, too, is consistent with historical precedent. It can result in temporary dislocation if the transition is fast enough, but things sort themselves out.
It's really only AGI, by eclipsing human capabilities across all useful work, that breaks this dynamic and creates the prospect of permanent structural unemployment.
We do have emplyoment problems arguably caused by tech, currently the bar of minimum viable productivity is higher than before in a lot of countries. In western welfare states there aren't jobs anymore for people who were doing groundskeeper ish things 50 years ago (apart from public sector subsidized employment programs).
We need to come up with ways of providing meaningful roles for the large percentage of people whose peg shape doesn't fit the median job hole.
The irregularities of many real-world problems will keep even humans of low intelligence employable in non-AGI scenarios. Consider that even if you build a robot to perform 99% of the job of, say, a janitor, there's still that last 1%. The robot is going to encounter things that it can't figure out, but any human with an IQ north of 70 can.
Now, initially this still looks like it's going to reduce demand for janitors by 99%. So it's still going to cause mass unemployment, right? Except, it's going to substantially reduce the cost of janitorial services, so more will be purchased. Not just janitorial services, of course. We'll deploy such robots to do many things at higher intensity than we do today, and as well as many things that we don't do at all right now because they're not cost effective. So in equilibrium (again, the transition may be messy), with 99% automation we end up with an economy 100x the size, and about the same number of humans employed.
I know this sounds crazy, but it's the historical norm. Today's industrialized economies already have hundreds of times the output of pre-industrial economies, and yet humans mostly remain employed. At no point did we find that we didn't want any more stuff, actually, and decide to start cashing out productivity increases as lower employment rather than more output.
We're quickly approaching how smart the average human can get, that's the problem and what sets this apparant from the historical norm.
This worked before because commonly people couldn't even read or do basic math. We figured that out and MUCH more and now everyday people are taught higher think for many years. People, today, are extremely smart as compared to all of human history.
But IMO we've kind of reached a ceiling. We can't push people further than we already have. In the last two decades this became very evident. Now almost everyone goes to college, but not all of them make it through.
The low-end has been steadily rising, that now for 20 bucks an hour you need a degree. That's with our technology NOW. We're already seeing the harmful effects of this as average or below-average people struggle to make even low incomes.
It's true that humans will always find new stuff to do. The issue is as time goes on this new stuff goes higher and higher. We can only push humans, as a whole, so far.
Self-driving vehicles can be symbiotic with mass transit. Autonomous taxis can make rail work better at low to moderate densities by shuttling people to stations. They can also ease the path to higher density development in car-centric suburbs by dropping people off at their destinations and then going to serve other riders or park out of the way, eliminating the need for density-killing parking lots immediately adjacent to businesses. As density rises in these areas, building mass transit will become more viable.
Nor does autonomy have to be limited to car-sized vehicles. It could be used for larger vehicles, perhaps intermediate between car and bus size — once you don't have to pay for a driver, operating a larger number of smaller vehicles is much more viable. These could be dynamically dispatched and routed, eliminating many drawbacks of existing bus service.
This is likely achievable in theory now, at least between major airports with ILS-equipped runways, but fully automated systems couldn't handle present traffic coordination procedures. You'd need a series of new standards to replace human-oriented air traffic control with a scheme in which ground computers could directly interface with aircraft flight management systems. Developing something like this and rolling it out on a global basis, in such a safety-critical application, would likely take two or three decades. Not clear it's really worth the trouble, since you'd want backup pilots for unexpected situations anyway.
What would probably make more sense is to just add a single-button auto-land feature, that sends an emergency destress call and configures and invokes existing automatic navigation, approach, and landing features to find the nearest appropriate airport and land. Given how rarely this would be used, there wouldn't be a need for the system to navigate complex traffic patterns, as ATC could just clear other aircraft out of the way. Something like this has recently become available for general aviation aircraft, but I haven't heard about anyone working on it for airliners yet.
"Code commands" is very plausibly an autocomplete flub of what was supposed to be "code commits." When I type "code comm" my iPhone offers up "commands" as the completion.
I've seen a lot of mockery of this request, but I suspect people aren't considering the wide variance in employee quality that can exist within a mismanaged organization. What Musk was asking for here wouldn't be a good way to evaluate skilled, conscientious developers, but it would be a pretty effective way to rapidly identify people who are basically incompetent or just aren't really doing anything.
> What Musk was asking for here wouldn't be a good way to evaluate skilled, conscientious developers, but it would be a pretty effective way to rapidly identify people who are basically incompetent or just aren't really doing anything.
So it's basically a FizzBuzz test, but for existing employees?
What's missing here? What else should they be doing? What are their competitors doing, in any space relevant to their markets, that's much different? None of these critiques ever seem to say.
If AI ends up being another 'normal' technology, Apple's advantages in distribution (~2B active devices, with a user base that installs updates pretty reliably), ability to give their AI tools access to your existing data and apps, and general facility with packaging tech so consumers actually understand what it's good for, put them in an extremely strong position to capture value from it.
If AI ends up being something other than a 'normal' technology, if we really are a few years from building the sand god, well, all bets are off, and it's a little silly to evaluate the strategic planning of an individual company against that backdrop.