> The evidence shows that there is no methodological moat for LLMS.
Does it? Then how come Meta hasn't been able to release a SOTA model? It's not for a lack of trying. Or compute. And it's not like DeepSeek had access to vastly more compute than other Chinese AI companies. Alibaba and Baidu have been working on AI for a long time and have way more money and compute, but they haven't been able to do what DeepSeek did.
They may not have been leading (as in, releasing a SOTA model), but they definitely can match others - easily, as shown by llama 3/4, which proves the point - there is no moat. With enough money and resources, you can match others. Whether without SOTA models you can make a business out of it is a different question.
> but they definitely can match others - easily, as shown by llama 3/4
Are we living in the same universe? LLAMA is universally recognized as one of the worst and least successful model releases. I am almost certain you haven't ever tried a LLAMA chat, because, by the beard of Thor, it's the worst experience anyone could ever had, with any LLM.
LLAMA 4 (behemoth, whatever, whatever) is an absolute steaming pile of trash, not even close to ChatGPT 4o/4/5/, Gemini(any) and even not even close to cheaper ones like DeepSeek. And to think Meta pirated torrents to train it...
What a bunch of criminal losers and what a bunch of waste of money, time and compute. Oh, at least the Metaverse is a success...
It's strange how people seem to think that Good Old Fashioned Journalism was objective and reliable. Bias is certainly not a modern invention. This bizarre nostalgia for an era that never actually existed seems to be pretty common, similar to the classic complaint that kids today are lazy and immoral.
Anton Chekhov was an activist journalist 160 years ago. This is not at all a novel phenomenon. And news editors who sought to defend the status quo would, of course, be just as biased as activists—they would amplify or suppress narratives according to their ideological allegiances.
The news industry has never been objective and pure. It has always been subjective and biased.
If a phenomenon is old it does not mean it's good.
If journalism isn't completely objective (or even if it can't be 100% objective) doesn't mean journalists should not be trying to be objective.
Code always had bugs - should we give up fixing them?
I am all for journalist activism if it's clearly marked as such. Disguising activism as news is the biggest sin that the modern media have committed and are committing.
The clear explanation is that neither Google nor Meta have had "ChatGPT" moments—everyone and their grandmothers have tried OpenAIs LLM so it's hardly surprising that people are excited for the follow-up.
That's the (hard) problem of qualia. You can't say for sure that ChatGPT doesn't experience anything. And I can't say for sure that you do, either. It's trivial to dismiss it on the basis of "well obviously it's like that"—far more difficult is to say something about why it's not like that.
I think consciousness has to be the product of an algorithmic process. I think this process can be summed up as "prediction". I don't think there's some secret quantum/soul/magic aspect to it. Which is why I think it's plausible that transformer models may capture some aspect of what we refer to as consciousness, though I don't quite buy it at a gut level.
You say that "it's not even comparable to the experience of a simple invertebrate," but that's just you writing down what your gut is telling you. What is the experience of a simple invertebrate like, exactly?
>You can't say for sure that ChatGPT doesn't experience anything. And I can't say for sure that you do, either.
The question rattling around in my brain tonight is - if some, many, people don't develop a theory of mind as young children, how many of the people around me are like this? Is it just the philosophers?
It seems of much more moment than the essence of ChatGPT.
Given how many people in this thread don’t seem to grasp the idea of qualia, and think the hard problem of consciousness to be easy, I’m not entirely unconvinced I’m surrounded by a bunch of imperfect p-zombies.
Why would there be a falsifiable hypothesis in it? Do you think that's a criterion for something being a scientific paper or something? If it ain't Popper, it ain't proper?
LLMs dramatically lower the bar for generating semi-plausible bullshit and it's highly likely that this will cause problems in the not-so-distant future. This is already happening. Ask any teacher anywhere. Students are cheating like crazy, letting chatGPT write their essays and answer their assignments without actually engaging with the material they're supposed to grok. News sites are pumping out LLM-generated articles and the ease of doing so means they have an edge over those who demand scrutiny and expertise in their reporting—it's not unlikely that we're going to be drowning in this type of content.
LLMs aren't perfect. RLHF is far from perfect. Language models will keep making subtle and not-so-subtle mistakes and dealing with this aspect of them is going to be a real challenge.
Personally, I think everyone should learn how to use this new technology. Adapting to it is the only thing that makes sense. The paper in question raised valid concerns about the nature of (current) LLMs and I see no reason why it should age poorly.
It's fine to think of entropy as messiness; that's the Boltzmann picture of statistical mechanics. The mistake is thinking that lowering entropy, or getting rid of the mess, is a satisfactory strategy.
Think of it as a negative feedback system, like a thermostat. Keeping entropy low means continually correcting errors. This is a successful strategy only if the world always stays the same, but it notoriously does not. Some degree of messiness is needed to remain flexible, strange as it may sound. There must be room to make the good kind of mistakes and happy little accidents (as Bob Ross would put it).
Because the author chose an analogy rooted in statistical mechanics, here's another: supercooled water. Take a bottle of purified water and put it in the cooler. It gets chilled below freezing temperature without freezing. If you give it a shake, it instantly freezes. The analogy may sound a bit vapid, but noise is the crucial ingredient for the system to "find" its lowest-energy state. The system crystallizes from some nucleation site.
It's the same with evolution. Mutations are a must. Keeping our genetic entropy low isn't a viable option, because that means we'll get stuck and die out. There must be opportunity for randomness, chance, chaos, noise; all that jazz.
This is how China became an economic powerhouse under Deng Xioping, for instance. They experimented with various policies and if something accidentally turned out to work great, it became something of a "nucleation site". The policy that worked in, say, Shaowu, would spread all across China. But it would never have worked if they stuck with a rigid, Confucian strategy of keeping "entropy" low at all times.
Entropy isn't necessarily fatal. Entropy can be used as a strategy for adaptation.
This is why I feel wary whenever I hear the phrase 'best practices'. Although they're generally promoted with good intentions, they're often accompanied by a utilitarian certitude that rapidly hardens into dogmatic inflexibility, followed by defensiveness or outright dishonesty in response to unforeseen consequences.
Most 'best practices' are good defaults, but the superlative rhetoric comes with the unstated assumption that any deviation is necessarily inferior, and that the best cannot be iterated upon. This drains agency from actors within a system, selecting for predictable mediocrity rather than risky innovation.
> a utilitarian certitude that rapidly hardens into dogmatic inflexibility…
Fascinating perspective to me, and probably the right one. The term “best practices“ gives me a warm feeling because I view it as a good starting point, one that will save me some time because others have figured out what not to do. It never occurred to me that best practices would be a static thing.
I feel the same way in the sense of not wanting to reinvent the wheel, but I often reflect on how implementation of 'best practice', 'zero tolerance' or other aspirational standards often falls upon administrators who might not have (or want) expertise in the practice area.
Thank you very much for this very precious feedback! I think it way better explain what I wanted to say in "3rd Poison: Momentum".
But, there is an important distinction to make: The China you are describing seems to have way more "available energy" (people and resources) than needed and no really specific goals, while in my experience it's rarely the case for smaller organizations, teams and individuals.
So, how to innovate and avoid starvation while operating on limited resources? This certainly deserves its own post :)
Thanks for this comment. I work as a senior leader and I've been trying hard to help my teams understand that we should do small mutations and successful mutations are things we should scale.
A lot of people struggle with this disconnect of "the mess" of "variety" but I think your comment helped me to maybe get some more analogy that could help folks get unlocked to the thinking.
Philosophically, many problems in our society might theoretically be attributed for optimizing for local maxima or other short term goals. Incentives and goals aren't aligned, and rules are far too rigid in favor in too few of the people. Democratic policies as in benefiting the people and democratic as in we elected this policy. Innovation and mutation are the spice of life.
I've observed this personally! After finding a solution to a problem and repeating it numerous times, I'll often randomly change one parameter of the solution (I'm talking about things like opening jars, not building complex systems, but it could apply there as well) to see if it works better. This often happens randomly because I ask my self "what if I did this?" as I'm performing the action.
The result is that almost invariably, I found a new way of doing something that's better than before. It often takes multiple tries, but it's something that takes little energy because it can be done throughout the day and the stakes are small.
Applied to a larger scale, random adjustments to larger systems can be exactly what's needed.
Before I was using my teeth, then one day I randomly tried using my hands. It surprisingly worked much better and I'm shocked more people don't do it this way.
Kidding =]. I actually don't have a better way of opening jars, it was more an example to show the triviality of the type of problems this kind of trial and error can work for. Maybe a better example is figuring out a repeatable way to get a stubborn door to shut. Things like that.
EDIT: Actually, you probably already know this, but if you turn a jar upside down and hit it on a counter top, it can knock the lid loose enough to twist off. I didn't mention it because I didn't figure that one out on my own, but if you're looking for jar-opening tips, that's a good one.
Pretty much the only one, near as anyone can tell, though an easy way to encourage or help someone or something get stuck in a local optimum is also a stable habitat/environment, as it avoids weeding out problematic noise from helpful noise until it is too late for all but the best luck to save it.
Another way of wording this / looking at it is in the trade-off between "adaptation" and "adaptability" (cf https://people.clas.ufl.edu/ulan/files/Conrad.pdf ). Adaptability requires additional energy to maintain that -- in a steady state -- is wasteful. Being highly adapted to a particular niche/circumstance is fragile to change. There is an intrinsic trade-off.
> It's fine to think of entropy as messiness; that's the Boltzmann picture of statistical mechanics.
But it should be added: If one leaves the field of kinetic gas theory and focuses on scenarios where cohesion becomes relevant, such as in crystallisation, an increase in entropy means an increase in order.
A much more obvious explanation is that it involved nothing more than stream-of-consciousness rambling. You take a random sentence and you convince someone it's deep and profound. They work hard trying to understand it, and they end up having a sudden moment of insight. The sentence itself did nothing. It was the work in trying to understand it that did the trick.
Reading tea leaves? You're just looking at something random. Scrying? It's the same. Nothing about the Oracle of Delphi strikes me as odd enough to demand a deeper explanation than that. It was just a person making nonsensical statements.
This reminds me of a short story by Ken Liu (https://www.uncannymagazine.com/article/50-things-every-ai-w...). He trained an AI on his own corpus of stories and made it spit out a list, then he wrote a story around it. Favorite item on the list: "I never expected to sell my rational numbers."
They used GPT-3 and DALL·E 2. For the latter, there's a waitlist on OpenAIs website. For GPT-3 they have an API available to developers and there are various apps which use it like ShortlyAI.
Does it? Then how come Meta hasn't been able to release a SOTA model? It's not for a lack of trying. Or compute. And it's not like DeepSeek had access to vastly more compute than other Chinese AI companies. Alibaba and Baidu have been working on AI for a long time and have way more money and compute, but they haven't been able to do what DeepSeek did.