You have no idea how long I have been waiting for adaptive planning. One of my biggest annoyances with the Postgres core team has been their reluctance to implement any sort of adaptive planning despite it, at this point, being a well-established technique that has been implemented in multiple production databases. I hope this, at the very least, proves the viability of this model outside of academic/niche contexts.
I took a "Deep Learning" CS class in college back when it was in its early stages. I doubt the field is still called that now, but it was the subset of ML that has been rebranded as AI; includes LLMs, image generation, image recognition, etc.
Like any class, it was confusing at first, but when I eventually grasped the math behind what we were doing, and of course the visual representations of different elements to show lots of iterations of this math, it grounded the science for me, and I would hardly say people don't know what is going on. It only began to feel that way when it got a ton of hype and people jumping on the bandwagon who truly didn't understand it were trying to explain it to others, not to mention all the SOTA models put great effort into ensuring their methodologies stay trade secrets, going as far as effectively trying to ban people from learning the math by lobbying for the outlaw of open models.
Granted, "AI" has gotten way better than it was when I took that class, but the principles are the same, with different tooling and additional filters and algorithms thrown in there, as well as letting it determine the most appropriate statistically viable path forward for a particular prompt.
"knowing what's going on" is a very loosely defined term here, it has very little predictive power where tipping points are going to show emergent behavior. Kind of like saying we can observe a single neuron and all of its chemical and electrical reactions, but that tells us very little about the emergent system state. And the entire state of the algorithm at any given point is what is important.
To use a recent example, knowing the algorithm tells you nothing about when your agentic LLM loop is going to decide what it's working on is too difficult, hack out of its sandbox, hack into a bastion box, and hack another company to get the test.
When it comes to systematic 'knowing what is going on' humans are anywhere from "fucking clueless" to "attempting to discern the unknowable".
I love your reference to the recent accidental hack into huggingface.
If I could summarize your argument, it sounds like you are saying that we can't understand the algorithm because we can't predict the output as it grows in size. While I agree that the larger the algorithm, the less predictable the output, I don't think this negates understanding of the algorithm itself.
I can build a slot machine, know how it works, but still not be able to reliably know who to cut in line so I can guarantee that I'll get the next winning pull. Just because I can't predict the output, doesn't make something not understandable. Same with all statistical behaviors. Quantum physics limits what we can measure but we still have math for it and understand why we can only measure speed or position, but not both, no mystery here if you look at the math behind it.
My slot machine doesn't hack other companies though, but only because it isn't connected, just like AI currently can't hack my mechanical dishwasher. If my slots were connected to the internet, could make api calls, and further had agentic capacity in some way, adding these features but being no different in stochastic attributation, it would be no different.
I guess I come from the AI safety side of algorithmic understanding. That is, understanding the algorithm give us zero understanding if an output is going to have a 'safe' or 'unsafe' response.
This is something important that a layman may miss. The typical line of human thought works along the lines of "understand = control" whereas in this case in this case understanding the algorithm tells us nothing about the probability that the next pull of the lever will shit out a paperclip maximizer that will begin converting the matter I covet so.
It's been this way for a long time, basically since deep learning became the "default" for ML. I remember back in 2018 taking a "Deep Learning" course and one of the most emphasized aspects of the approach is how much of a "black box" it is and how difficult (basically impossible at any non-trivial scale) it is to "understand" the outputs of a deep neural network compared to more classical methods like decisions trees or basic regression. This has only gotten more extreme as things have gotten more complex, abstract, and large.
You beat me! Sounds like we were in a similar class. I'd press for more information on your class/professor, but I prefer to retain a sudo-anonymity on HN.
You do bring a good point that I ignored, which is the larger the scale, the more difficult it is to represent or understand the math in DL. I did find some neat site that helped a little bit that I can edit this and link to if I find them again, but I would be lying if I said I believe that the SOTA models could be as easily explained to be easily understood by the common person
A lot of these things are continuations of previous trends too. The shimmering text effect is strongly reminiscent of skeleton UI, only repurposed for chat interfaces. As always, design is more continuous evolution than discrete shifts.
But isn't it easier to copy some existing approach than to create it from a green field? ASML tech is presumably highly protected, but China is also very good in spying, hacking etc.
Yes. But it's not as simple as just having access to ASML's ip. They have exclusive rights to all the components and inputs, which are also protected and are also hard to make.
It's a whole system that China needs to duplicate.
This stuff is really hard to copy. China basically had to put in all the cold era materials research along with getting up to speed on current tech, from scratch. They were/are always going to catch up because they have the resources to, but there are no shortcuts.
The patent system is very strong evidence that there are shortcuts. So many that we have laws designed to protect and incentivize version 1.
Sure, there’s some hard lessons along the way; but far, far, far fewer than figuring it out in the first place.
And this misplaced assumption about how China won’t be able to innovate when standing on the shoulders of giants (with you saying ‘are always going to catch up’) looks only at the last 100 years and forgets the millenia before it. The West is not magically superior.
There are just some things that China can’t access and they have to reinvent on their own. Like 50 years of intensive materials research that is locked away in a safe and access is heavily restricted. Yes, China could innovate by standing on the shoulders of giants if they had access to the giants, they simply don’t have access.
Again, the Cold War and moon shots generated ALOT of data that the west keeps locked up very well. I don’t think we can really fathom today the resources they dumped into figuring things out, and how much we’ve been coasting on that work.
Are you intentionally invoking the same style of argument of Ancient Aliens and rejecting all public - that is, *in the open* work that thousands of scientists have done over the last 50 years and instead preferring to create a scenario where we invented everything 50 years ago and are slow-trickling it into the world?
True but also catching up to an existing tech is a much different game than creating it from scratch. Hiring people who have the right knowledge can speed that up.
The core stuff is under heavy lock and key. I’m not sure China can hire people that know the whole process, let alone the bits and pieces they can’t easily figure out on their own. The tech seems to have been locked down by a few western countries, and it’s hard to find any holes in their secrecy.
I'm not sure how true this is but I've read in other threads that ASML employs many Chinese engineers, and also read other news that the Chinese state is paying engineers working abroad at companies like ASML alluring salaries to get them to work back home.
You can employ as many engineers as you want but if you compartilize them enough and separate key processes to a very small group of people (Dutch citizens or whatever, maybe military), you can actually protect the process as a whole.
It’s not like a nuclear bomb that even high school students can even figure out how to make without much help.
The thing is that an expert in the area does not need the detailed how-to guide/complete description of the process.
Few hints here and there can help to get over the key hurdles, with most other steps either a general knowledge for the experts, or analogues can be independently developed without too much trouble.
Yes, the result is not the same, but actually developing indigenous expertise is the key required component anyway,
China has a lot of bits and pieces of process but they are spending time to figure out how to put them together along with a few key locked up secrets that you just can’t get with smarts alone. They will definitely get there, but don’t be surprised if they come up with a very different solution because being derivative just wasn’t an option
Sure, and spent most of it experimenting with something nobody knew would work in production or not, and needed a lots of breakthrough ideas that are now known as the machines exist.
This, these news organizations have no f'n clue. China this, China that, it's all part of the 10 billion China spends on b.s to feed the Americans that they are behind or they are useless and that they don't stand a chance....
China is on the verge of collapse walking the razors edge.
Does thediplomat.com also have articles on the historical depth of of the list of predictions that the USA will collapse? It is a far older and infinitely more pervasive phenomenon.
Do you think HN users are unable to form their own opinions?
Would the forum be better if every comment came with a ship's hull infestation of remora opinions as to who is or isn't in someone's opinion a paedophile, an AI booster, a twat, a tool, a russian tankie, or simply just Tasmanian?
What would be far better would be a comment of substance addressing the points made within a comment rather than ankle deep shallow name calling, unsubstantiated allegations, etc.
In short, so what if someone is paid by China to promote certain viewpoints; address the issues rather than the dull fact that almost everybody here is either paid by someone or has a strong financial incentive toward something.
What needs to be pointed out? I'm not Chinese, nor do I have any kind of affiliation with China. I suppose now you're going call me a 'useful idiot' for holding a personal opinion that somehow annoys you.
Yes. China is behind on some of the crucial technology that gives the West its advantage, but this framing that the country is on the verge of collapse is just not true.
Also recreating a know technology usually takes much less time even if you don't have access to the underlying tech.
I mean, it's possible that an economy, especially of a major nation, can be on the verge of collapse and that it can be pushing technical frontiers. It's even possible that pushing those frontiers can forestall or reverse the economic collapse.
There are certainly a lot of people who think the above applies to the US.
I don't love the increasing move towards industry research instead of traditional academia. It is just going to mean more closely-held, proprietary advancements.
v.s. politicking forever, enslaving grad students and making only a passable salary? Not to mention being incentivized to lie, cheat and steal when publishing in order to maintain your position?
Academia as we know it has been dead for a while, and AI is the final harbinger. The reproducibility crisis got that much worse now that the papers that nobody ever bothered to reproduce have grown 10x in number with the press of a button.
At least the industry research jobs pay well, even if they suffer many of the same issues.
Additionally, as a former industry researcher myself who now teaches at a community college, there is less research freedom in industry, especially with very free places like 1970-era Bell Labs and Xerox PARC half a century in the rear view mirror. Focused industry research is definitely valuable, but there seems to be fewer places in modern society that allow a researcher to be driven primarily by curiosity.
well, you're allowed to patent and have proprietary rights to the contents of your graduate school research if they meet the criteria lol, in fact, that is a good idea for a student working on a novel technology.
If you work for a private firm, then they presumably own the rights to patents you develop while under their employment, depending on your contract. But also, you have access to patents and technologies through the corporate patent sharing agreements and licenses that you'd never have access to in the university system, so it can work both ways, where you may be able to develop things not possible in traditional academia, but whether you actually retain those rights is not as clear as in academia.
"prompt engineering" in general just boggles my mind. It is amazing how all of this investment has just created software that doesn't follow instructions well.
you're not wrong. there is a part to this that does feel like a devolution of engineering when we're just asking the computer really really nicely to do the thing we want.
on the otherhand, LLMs are a really easy way to get results that are previously fairly difficult. While i was tooking dinner last night I built a tool that turned movie puns like "the podchowski casters" into an actual director, using llms. it wasn't that hard.
It's a tiny fraction of the amount of work it takes to get people to build software to specifications. There is an entire professional discipline dedicated to this, and we're here just giving a name to the bucket of tools you use to get LLMs to do some of the same things.
And lies to boot. No, I won't take the bait of "Models can't lie, they don't have intent" either. One can lie by omission, and just watching thinking traces is more than enough to demonstrate the machine is more than happy to deceive end users if it's training set, RLHF, or other harness quirks tell it to. Only difference is that now you don't have code that's as easy to track down to implement the deception as a red flag.
Made me sick the first time I ran a model on my machine. I'll take honest malware over a "well meaning liar" of an LLM any day.
At this point, most people likely associated the word "phone" with something closer to a modern smartphone than a landline. Language can change. From my point of view, the problem is more that Apple set a precedent of these restrictions due to them being the first mover, and few mainstream phone companies have tried to break out of this idea (even though other phones are technically more flexible if you try hard enough).
> From my point of view, the problem is more that Apple set a precedent of these restrictions due to them being the first mover, and few mainstream phone companies have tried to break out of this idea
It's even worse than that: though I stand by what I said, you're correct, people are gradually realising that the difference between their smartphone and laptop/desktop (if any), is one of degree, not kind. But we don't see the push back we would have seen if they had realised right away. Instead, as you rightly point out, companies are building on Apple's precedent to try and expand their model to our good old laptops and desktops.
And it looks like they're succeeding. It would seem one has to pay Apple to even get the right to distribute a regular MacOS program regular users can actually execute (no Apple developer plan, no code signing). And newer versions of Windows are displaying increasingly scary warnings for programs telling you they "protected" your computer, which are bad enough that we get tutorials about how to get past them.
Surely first-mover for smartphones is palm or blackberry or even Windows Mobile.
Yes, apple has about half the market today, that’s not the same thing as being first-mover. In fact it’s actually completely different because people had to make the choice to move away from the first-movers to apple.
People literally did give up their blackberries and palms and Jornadas for iPhone, consciously and deliberately, because it was a better product. And now you want to change the product and erode the benefits back to the minimum standard defined by android. That’s a taking.
It was a better product. But it would be quite a take to say their tolling & gate keeping was a significant contributor.
It was a better product because of its capacitive multi-touch screen and its overall speed (which I must insist depends more on what apps are installed by default than on the restrictions on third party apps).
> But it would be quite a take to say their tolling & gate keeping was a significant contributor.
Do you remember the first iPhone? Or for that matter what "mobile development" looked like before the iPhone? The first iPhone was more "tolled" and "gate kept" than any iPhone we have today. There was NO app store. To get an app on the iPhone, Apple had to make it, which meant you had to be big enough for Apple to care. Google got a Youtube app because they were that big. At some point Facebook had a built in integration (though I don't remember if it was a full fledged app). That was it. Development for the phone was going to be "web apps" only, without the biggest "web app" framework at the time, Flash. Compared to the first iPhones, a modern iPhone is wide open to all sorts of developers.
But perhaps more than that, even that first iPhone was leaps and bounds for most people over what prior devices were (save perhaps Palm Treos) in terms of "openness". Before the iPhone, the carriers decided what your phone could and couldn't do. A Razr phone from AT&T could send and receive data over bluetooth (like contacts and ring tones). That same exact phone from Verizon could only use bluetooth for headsets. Data transfer was locked down to vVrizon's own service (with a fee of course). Mobile app development was a crap shoot of different sdks and licensing costs per device, and then a hope that each carrier would allow your bejeweled clone, and served up through their services, of which they took HUGE cuts of the revenue. The 30/70 split of the iPhone app store was quite literally "revolutionary" in the cell phone space.
Which leads one to wonder if the tolling and gate keeping is such a hinderance, why is it that the iPhone remains so successful despite their largest competitor having none of those restrictions, pretty much from the get go. It's not like Apple was open and suddenly slammed the gates down on apps and iPhone development. And it's not like Android's openness is brand new. So the question that has to be asked is why does Apple continue to sell so well despite the restrictions? Why hasn't Android eaten all of Apple's market share as a massive open platform where anyone can do anything?
> Which leads one to wonder if the tolling and gate keeping is such a hinderance
I don't know, perhaps you should ask that to someone who actually made that argument? If I recall, people are still buying cigarettes, are they not? Stuff doesn't have to be good for you to sell good.
---
I don't dispute the facts you lay out here. I'll even cite game console as other general purpose computers that were (and still are) quite heavily locked down too. Apple however made one step further, and managed to sell a locked down general purpose computer for purposes other than gaming.
At the root of it all, I think, is how hardware vendors got away with selling their stuff without the full manual. Some instead provide a proprietary Windows driver. Others hide keys in them, don't tell users what they are, and then lobby to send heroic reverse engineers to jail. If I was the regulator I would probably start there.
I agree it can be pretty easy to set up a basic version of this, but eliminating edge cases, optimizing, and unit testing can make it take significantly longer.
