GenUI isn't about designing cosmetic "skins." (Usually, anyway. I guess it could be used for that)
It's generally for letting users customize the own workflows. How many times have you, or one of your users, liked a piece of software because it mostly fits an existing workflow but that remaining 20% is an annoyance, or maybe even a dealbreaker?
This is probably more common for businesses. They have existing procedures. and they want your software to fit into their existing processes and workflows... not the other way around.
GenUI is far from a one size fits all approach or magic bullet, but it can address a lot of those situations that either would have been dealbreakers, annoyances, or change requests. I suppose it can also help with user retention; once they've put the time and effort into customizing your product they theoretically are less likely to switch to a competitor.
Existing OpenAI/Anthropic models seem to already handle this pretty well. As you might expect, letting users describe their own UI is pretty easy. The hard part is making it work and making sure they don't escape their sandbox...
If running LLMs locally matters, it’s hard to imagine a “forever machine” existing in anything less than 5-10 years, probably more. This stuff is just evolving so rapidly. Buying a “forever machine” today might be like buying a “forever GPU” in 2003.
It just seems somewhat more efficient to take all the computing storage, memory processes, processors, or at least all the money that go into it… the in one spot.
If you’re running a big task or small, everything scales to the appropriate size, regardless of the hardware sitting on your lap or under your desk.
Buying a Mac specifically for serving is misguided in nearly all use cases, but as an all-in-one solution they make sense. Personally I feel like the number of boxes I need to manage has an inverse correlation to my happiness.
Macs are totally “fine” for light server duty… as is just about any computer of the last decade+. The CPUs are beasts, the disks are screaming fast.
The operating system itself may not be ideal at serving but you can just run Docker/Orbstack if you need to do something especially Linux-y.
I’d put the question back on you — what are scenarios where an Apple Silicon Mac wouldn’t cut it as a light server for one person or a handful of people? About the only scenario that comes to mind is scenarios where you expect to utilize it so heavily that the fans are running for many hours a day. At some point those are either gonna wear out or just ingest so much dust that the machine runs hotter and needs a deep clean. But even that is largely mitigated by just pointing an external fan at it.
Will Claude will act as a therapist or produce
value or produce a work of art? No. It cannot,
because it does not have a soul.
I tentatively agree, although I'm only tentative because I don't think it's an interesting question.
Here's what I do think is interesting. You!
I mean... yes, you, too but not you specifically. The plural "you" that the english language lacks.
And so here's what I think is the actual interesting question. Might AI help you create art? Or be a therapist? Or something else interesting and worthwhile?
Maybe AI won't write the next great guitar solo. I'm pretty sure it won't. But might it help you learn to play guitar? Help you fix your broken guitar amp? Help you understand some tricky parts of guitar playing? Help you work through some tricky tabulature where you can't tell if you're playing it wrong or if the tab is just bad?
I don't know. But that's my angle for finding any of this interesting.
Sol medium/high planner orchestrating -> Luna xhigh subagents doing implementation
...has been REALLY good for me. Even on xhigh, Luna is crazy cheap.
Subjectively I'd say it's way better than Sonnet at a fraction of the cost. Luna xhigh can do some decently challenging things on its own, but when orchestrated by a model that is actually good like Sol, I am finding it very very nice.
The cool kids have custom harnesses and workflows and stuff, yeah. I'm still using Superpowers in Codex. Planning in Sol, Luna subagents. https://github.com/obra/superpowers
I feel like I could be doing a lot better somehow. Regardless though Luna (xhigh specifically) is super good/cheap/fast for a lot of things
Shall the better model still have
the upper hand or will the raw speed
compensate?
At 14,000 tokens/sec there's just so much ridiculous stuff that might be possible. Let's assume that this POC proves they can take the next step, and can eventually etch a capable ~27B model into silicon. Let's call it Fred.
Ralph loops automatically get real real interesting again. 200x the iteration speed. This is such a clear win I feel like there's hardly anything to talk about. Instead of one stubborn iterating idiot, you could have dozens of idiots competing in parallel, genetic algorithm style.
The other common orchestration pattern I see is "big model for planning, small parallel subagents implementing, big model reviewing" Today it's Sol dispatching a handful of Luna subagents. Tomorrow maybe it's Sol dispatching as many Fred subagents as it could possibly want.
But what patterns have we not even thought about yet in a world where subagents are 200x faster/cheaper?
What if instead of dispatching single Haiku/Luna/etc subagents, we dispatched "teams" of Fred agents? Maybe each team is 8 Freds. Five come up with competing ideas and the other three vote on a winner.
Or what if they were heterogenous teams? One Luna and a bunch of Freds.
What if instead of a two-tier orchestration system (Sol->Luna) it was three-tier or n-tier? (Sol->Luna->Fred->...Fred)
Those ideas overlap a bit, and crazy shit like Gas Town has already explored even wilder ideas I guess. But man, 14000 tk/sec opens up so much stuff.
At 14,000t/s that's effectively a motor cortex for an android, you no longer need to train the robot to walk, it has a general idea for how to walk (baked into the 1b model), and then just corrects based on sensor input, in real time.
I still personally think that a heavy lean into MoE will be better for that sort of thing. Our brains are subdivided into large parts but I'm sure (and I'm not a brain scientist) that those parts can be subdivided even further into systems that run at various frequencies and latencies depending on what they're used for.
I was thinking about it the other day actually. How our brains evolved structure. I imagine it was purely just down to evolution adding/clustering additional cells around the areas where additional cells were needed. And after long enough a natural brain architecture emerged.
Makes me wonder if we're on the right track with transformer architecture/attention but if it'd be more effective on a larger scale, like MoE with a billion "experts".
That seems promising to me too, although, the thing I've always read is that you can't make the "experts" too narrow. Even if you had a "coding expert" it has to know a lot more than coding - if you tell it to make an online store it needs to parse your language, understand the internet, what a "store" is in this context, etc.
I am not a primary source, probably not even a secondary or tertiary source, so take this with all the grains of salt.
Oh for sure, but I think generally multiple experts are selected in an MoE pass for a token, so presumably it'd select programming related ones as well as general knowledge/language.
Only problem would be the routing layer works on the previous token as far as I understand so it might need more informational depth than just "a token" to select experts, I suppose in the same way attention works.
I wish I had the GPUs to run those sorts of experiments ha ha.
I can see an extra shade of purple. I had essentially the same lens removal cataract surgery as Monet did. Like Monet, I can see slightly into the ultraviolet range, because the retina can react to some UV frequencies that the lens normally blocks! Unlike Monet, I'm not much of a painter.
It really does defy words, for obvious reasons. The short answer is that it's "like purple... but even more purple"
The slightly longer version is: imagine moving a color slider from red to purple and then keep going a little more.
This is true and it's barely even debatable. Whatever exact role language plays in our thought processes, it is most definitely nonzero.
It's why I think "LLMs are only fancy autocorrect" style takes are really underselling how wild it is that we've, in a roundabout way, sort of crystallized a bit of the human thought process in a way that is genuinely useful for a lot of tasks.
> sort of crystallized a bit of the human thought process
a) LLMs don't think. They predict a most probable sequence of language tokens. Huge difference there.
b) Whatever LLMs do doesn't model human behavior whatsoever. LLMs are basically very fancy logistic regressors. I.e., it's a mathematical abstraction first and foremost.
When I see these sorts of debates about LLMs thinking, its rarely a disagreement about what LLMs do. Its almost always over how 'thinking' is defined and the two sides use different definitions but don't actually communicate to each other what those definitions are because they assume the other side is using the same one.
The loosest definition of thinking is along the lines of anything that can process information in a useful way. Basic calculators can therefore think about adding two numbers. The strictest definitions tend to on the side that it is linked to the nebulous concept of consciousness and therefore cannot ever be machine generated. In that we don't even really understand how humans think, so how could we possibly know if machines can do it.
It has nothing to do with thinking or consciousness.
There is a common misconception that LLM are simply a "statistical process" that doesn't feature any abstract conception of the tokens it is predicting. There are studies that show that such features do exist - that there is discernible structure built into the weights - and that the process of inference is a very rich one.
The statistical process exists but it is the substrate in which the model is implemented - or more accurately - grown.
If you can predict Magnus Carlsen's next move then you are just as good at chess as Magnus - and being that good absolutely does require reasoning.
If you can predict the solution to an open Erdos problem that stumped hundreds of people for decades...
I don't dispute what you're saying about how LLMs work, but this is exactly what I mean. LLMs can be shown to demonstrate reasoning, so if define thinking as being able to reason, then LLMs can indeed think. If instead you define thinking as being more than just reasoning, then LLMs cannot think. Neither of these two options change what LLMs do, its just an argument about the best word to describe it.
When I see these sorts of debates about LLMs thinking,
its rarely a disagreement about what LLMs do. Its almost
always over how 'thinking' is defined and the two sides
use different definitions
Well, hmmm. Yes, I think that happens a lot.
I think there's a pattern that happens even more often, and it's what happened here.
Whether I'm right or not, what I said was somewhat nuanced - I stated language is a part of our thought process (even posted research to support this) and, given that fact, I think many underrate how wild this achievement is even if it's only "fancy autocorrect."
And, of course, the other side comes in with BUT IT'S NOT THINKING.
Which... I didn't say, and I would not say, because (like you said) it's impossible to do without the discussion immediately devolving into semantics. Semantics that I'm really, really uninterested in. But, FWIW, I like your definition.
You should look into emergence. An ant in a colony, an offer in a market, a drop of water in a weather system are all evidence that irreducibly complex things have simple mechanisms at their core.
Did you reply to the right post? You quoted me, but you wrote "LLMs don't think" as if it was a rebuttal. It's puzzling, because I didn't say that they think, so it kinda seems like you got confused? Maybe somebody else said that?
I don't really have an opinion on whether or not they "think" because I feel it's impossible to even discuss without getting into a very very uninteresting semantic argument about what "thinking" is.
Are we defining "thinking" as doing it the same way humans do it? Then, of course they're not thinking. It's a statistical model, not axons and neurons, or even a simulation of axons and neurons.
Are we defining "thinking" on a purely functional or behavioral basis, kind of a Turing test approach? Then... well, I think it gets nuanced. For some tasks, within some constraints, they do pass that test. For many others, of course they don't.
Are we defining thinking in more esoteric terms? Something to do with the soul? Maybe the ability to come up with truly novel concepts rather than rehashing and remixing the stuff it was trained on? Do ants think? Do dogs think? Do jellyfish think? Octopi? A newborn baby?
Anyway, it's a deeply uninteresting semantic question.
The (or a) current neuroscience models of the brain are that its main job is to predict how the body should be responding in the near future. Obviously a lot more complex network nodes than an LLM, but prediction is clearly tied up with thought in some way.
I don’t find LLMs to be very good independent thinkers, but I wouldn’t over sell our own mentation either - it clearly arises from a large number of simpler entities.
The more significant difference is that the LLM is stuck with language which is clearly an emergent and secondary capability of our own thinking. We can formulate words to explain things, but we also can look at two volumes and feel what it means that one is larger than the other. Raise a toddler and you can see the progression from not understanding, repeated experiments, muscle memory and finally to conscious point for reasoning.
I don’t find LLMs to be very good independent
thinkers, but I wouldn’t over sell our own mentation
either - it clearly arises from a large number of
simpler entities.
Yeah. I don't see them ever hitting the heights of human creativity in terms of coming up with entirely new ideas, schools of thought, etc. That really might be a fundamental limitation of being trained on existing thought. Also, a lot of human experience involves (1) things we don't have words for (2) things we've never put into words.
clearly an emergent and secondary capability of
our own thinking.
Yes. And it's part of our thinking. More than a capability . Thought influences speech, but speech also influences thought.
That's why I think it's remarkable that we've managed to (choosing my words very, very carefully here) create a statistical model that does a remarkably decent job at emulating the behavior of a fragment of that process.
It is remarkable. A shame that unregulated business models and giant pools of capital seeking very high returns seem to be distorting (hype cycle, excessive spending that seems untethered to realistic plans, and uncritical roll out prior to compelling proof of utility) the rollout of the remarkable new technology.
I would speculate when we do eventually develop independent synthetic sentient beings, LLM technology will be a part of the package. Perhaps also growing up with a sibling that tries to trick one.
Maybe someone needs to write the singularity novel but with Cain and Abel, not just a unified super intelligence but siblings full of some good will and a good bit of clear seeing and some fun (?) trickery.
The statistical model that underpins any deep learning system is the substrate in which a process is implemented. There is still a process - just because that process is grown - not programmed - doesn't tell us anything about the depth or limits of its capability.
"Fancy logistic regressors" are, in fact, a modelling tool. You can tell LLMs model human behaviour because they're doing things that until a few years ago only humans could do, like cheat in exams.
It's generally for letting users customize the own workflows. How many times have you, or one of your users, liked a piece of software because it mostly fits an existing workflow but that remaining 20% is an annoyance, or maybe even a dealbreaker?
This is probably more common for businesses. They have existing procedures. and they want your software to fit into their existing processes and workflows... not the other way around.
GenUI is far from a one size fits all approach or magic bullet, but it can address a lot of those situations that either would have been dealbreakers, annoyances, or change requests. I suppose it can also help with user retention; once they've put the time and effort into customizing your product they theoretically are less likely to switch to a competitor.
Existing OpenAI/Anthropic models seem to already handle this pretty well. As you might expect, letting users describe their own UI is pretty easy. The hard part is making it work and making sure they don't escape their sandbox...
reply