ux266478
It's all tradeoffs. The same power that grants the ability to reverse transactions also grants the ability to reverse a transaction in bad faith and commit fraud. Will the system correct that? Would it surprise you to know that it doesn't have to? That same power also grants the ability to lock an entity out of any kind of monetary transaction, and things like debanking. Usable for good? Sure. I don't think it's controversial to say that it being leveraged against Hunter Biden was at all justified, however. Would you like to be on the receiving end of that? Would you like to see a world where an LLM gets to decide that kind of thing based on your presence in the surveillance panopticon?
To that sense, it's really just a disagreement on what tradeoffs are acceptable. People who have a utopian view of cryptocurrency are looking at some tradeoffs as preferable to others, it's a different failure mode. To be clear, I don't necessarily agree with it myself, but you shouldn't be so uncharitable to people to assume disagreement is always a matter of delusional ignorance.
The problem is that this sneaks in a postulate by negating "optimal", which is that underutilization is the opposite of that which is desirable, that it is implicitly undesirable. If we don't take that premise, say that it's value neutral and neither desirable nor undesirable by itself, then the argument doesn't really stand. This neutral valuation is far more reasonable to me. Anybody who works in logistics intuitively already understands this. We don't have cargo ships bow to stern across every single shipping lane, yet this is "underutilization" of those shipping lanes. Sure they're bottlenecked at the ports, but the ports can always be bigger, more numerous and possibly even work faster. The slop turns out doesn't matter. Actually it's beneficial for a few reasons, but my point here is sufficient I think.
You're right this is ultimately a power struggle at the core of it, what I think you could stand to acknowledge is that's all it is. It's just dressed up as economic rhetoric.
Not that it defeats your point, but an 8x MI355X node is $350k-400k. The only reason you're paying that much is for the VRAM, too. You could run it with much less compute than what you get in a single card.
What, you don't want to save a few bucks on a train ticket? Come on, you'll love it!
The struggle you have with pinning down moral relativism I think betrays the fact that your understanding of it is low quality. A good ear mark is, can you name a single passage from a treatise on moral relativism that you like? If you can't find a single aspect of a philosophical construction to advocate for, that means you don't actually understand it.
You also keep sliding between conflating moral relativism with moral anti-realism and even moral nihilism at one point. These are different axes.
> Realism + Absolutism
All moral facts converge for every single person. As a matter of fact, they aren't facts at all. Morals are a hinge of the Wittgenstein variety.
> Realism + Relativism
Moral facts are derived from frameworks, or contexts, which are themselves grounded facets of reality.
> Anti-Realism + Absolutism
Kantian Constructivism. There are no facts which are coherent without a framework, but moral claims are still necessarily only capable of being universal.
> Anti-Realism + Relativism
Moral reality is constructed by the framework, grounded only to the framework.
[1] - https://www.eastwestcenter.org/sites/default/files/private/i...
[2] - https://www.usitc.gov/publications/332/journals/chinese_semi...
You can tell the post-training makes all the difference though. Gemini has a different stink to it than DeepSeek.
> Some way of informing the model whether the output it is producing is good or bad
That's what post-training is for. At its most basic, you're giving it examples of inputs and outputs and then doing reinforcement learning to calibrate its adaptation to your examples. You need much less to fine tune a model than you do to pretrain the weights. You can build a really excellent LoRA for a stable diffusion model, for example, with 50 high quality images. LLMs are "a bit" more complicated and costly to fine tune, and you have to be mindful of the agentic loop, but the principle is the same. There's more to it than just LoRAs. Steering vectors, projection layers, custom encoders, etc. There's a fair amount to learn, but it sounds a lot scarier than it is.
Here's something to chew on: chain-of-thought doesn't exist until after pretraining! It's basically created by having <think>...</think> blocks directly in some example outputs, and this is fine-tuned into stability. It's literally not much more than a parlor trick and some careful calibration. A powerful parlor trick to be sure, though.
As the compute to build adaptions for multi-billion to trillion parameter models becomes more and more available (and affordable), and the artistic techniques of fine tuning and harness engineering spread throughout the public, I think we'll come to see the "one size fits all" model for the non-starter it really is. Anybody who's even toyed around with LoRAs (let alone made their own) already knows this. That's a very deep rabbit hole, and the ceiling is determined by target domain knowledge and systems engineering.
I'm bullish on LLMs as expert tools in the same grain that computers are. You have to learn all about them to use them effectively. But what really makes the difference is how well you know what you're going to be pointing them at. There is very little room for companies like OpenAI or Google to live between us and our tools.
On a broader timescale, we can take a look at the fact that goods which were designed to be affordable and mass-produced have been elevated to refined luxury goods, like the Herman-Miller Eames, as an indicator that material standards have fallen very far. Not the least of which because of an ever-dwindling labor share of capital (beginning in the 1970s), which silently impoverishes working households and degrades their comfort while being completely invisible to an uncareful eye on paper.
Much of the major boons of comfort were established prior to the transition to shareholder capitalism. Things like e-commerce, IoT, online banking, these are for the birds. Meaningful advances are things like HVAC and refrigeration, ubiquitous motorized transport, electrical lighting.
Something as simple as output length is a hard linear floor for productivity, even putting aside the obvious context problems that you're intuiting, and it's far from being the biggest cost that arises from steering skill. Learning to make a smaller, faster model do the same work with less tokens is a technical domain that a lot of people don't seem capable of learning. I'm not just talking about "context engineering", but learning how to fine tune, post-train, create better harnesses, design inference setups, etc. If we're both using AI, but I'm beating you to market every single time and with a better product, what is your AI usage actually buying you? Yes, competency and skill is this meaningful right now, and it's highly technical. Not the least of which because you know how to describe the problem in way that gives it a smaller solution and requires less iteration.
Most of the labor who understand the technology enough to do those things lives at the companies selling you these services, but you can absolutely learn to do these things yourself right now. It's actually really fun! A hell of a lot more fun than fucking prompting that's for sure.
Where we're at, I would equate it to the early mainframe era where the programmers came with the computer. I'm placing calls that we follow a similar track and the two will end up decoupling, that "model engineers" are going to move in-house. OpenAI will have a ring to it like IBM does today.
I'm thinking more the Noam Chomsky and John Searle variety. For the stochastic parrots people, which I assume you to mean the no-skin-in-the-game bloggers, I'm not concerned about them. I find a lot of the rhetoric around LLMs to be eye-roll worthy, most people slinging it often lack a coherent theory of semantics to begin with, let alone an understanding of logical induction/statistics. Take the definitional entailment that these models are ampliative. This small fact undermines quite a lot of the naive mental model people have of what on transformers even are. No intentional theory, no position driving the argument, the opposition isn't substantial. The virtue signal is valid, but from strangers is uninteresting.
As you mention the humanities is very diverse, linguistics is no exception. It's not that nobody was on the corner of distributional semantics, but it's been a long road and for much of its life results were routinely dismissed in the mainstream out of dogmatism. Noam Chomsky is a good pull because I believe he's the most prominent example of this chauvinism. Vague hand-waives about explanatory power, etc.
> I'm thinking of post-structuralists especially.
I understand what you mean. It's probably not coincidence that Chomsky is a vocal critic of the school. I think many different post-modernist camps even beyond post-structuralism actually are amenable to the implications of distributional semantics, and are probably the better equipped for it. I think there are nuanced problems unifying the two, but it's a work day so I won't elaborate.
> It would depend a lot on the specific school and instructor but in general I really don't think you would. I've taken classes like this and people were far more open minded and critical than you might assume.
I was just being cute with that, really. Although in this case, I think critique is itself the problematic lever, acceptance is more an exercise of apologetics, which is the weak point of post-structuralism in many ways.
> By "associative logics" I hand-wavingly mean all manner of categorical thinking, "amygdala" thinking, mapping, putting things into buckets, hedging.
I assumed so, but it's more or less a long phrase to repeat the same concept. Relations (IE logic), mappings, morphisms, associations, patterns, etc. Same-side, same-coin. Informal, formal, take a position of drawing no line here and the shadows scatter. The extraneous qualifiers aren't what confused me though, it's that gerund "seeking". It sticks out enough to imply additional structure, but one that isn't contextually indicated. Taken in a conservative form, I considered pattern-seeking to be interchangeable with pattern-constructing, which then returns it to just being an extraneous qualifier (hence the "bit incoherent", it's a mild interpretive non-confidence).
There's also not very much I would say is technically impressive about it at all, and definitely not anything people normally reach for. The bandwidth of the asset streaming they accomplished on the PS4, with such a small team that still had to build the rest of the game, is the only thing that makes me raise an eyebrow. In fact, that may as well have been the real gain of crunching down how planets are described rather than the huge number of planets themselves. Even that's not particularly groundbreaking if you take away the context of the small team. And at that point we have to ask ourselves if it's groundbreaking, or just very intelligent and clever pragmatism. Impressive in a labor lens, not a technical one.
If you walked into a film theory class and posited that you could derive every single encoded interpretation of a film by memorizing the positioning of the actors and objects frame-by-frame, plus the audio track in another language that you do not speak, for every single piece of video ever made, you'd probably be asked to leave. Not that I'm advocating for the position of critique here, I don't think the anti-distributional semantics crowd is ever going to recover from their humiliation that's been accelerating over the last 8 years. It's just that from the position of critique it requires a coherent narrative that human brains are capable of ingesting (IE not maximal information overload).
If you really want to go the lower level route, I think Francois Laruelle's non-philosophie touches on what you might be thinking of in a much more robust way, shining a light on the unexamined consequences of decision and dialectics of-themselves. If you can stomach the writing of continentals, that is.
> yet nothing prevents you or your agent from changing invariants to fit the algorithm and making it incorrect.
If the thing we're trying to prove is ungrounded, it's trivially true that any verification method falls apart. You're still encoding priors in your testing methodology. No matter how robust you think your testing suite is, I can still trick it if given free reign over the codebase.
Watching this next generation closely, because there are just a shitload of red flags starting to crop up that don't seem to be going away.