>Classification models have been around for a while.
I'm still trying to catch-up on the Jev stuff, but my understanding is that it's basically just a more efficient LLM when all you want is the LLM to produce a classification.
There's more to it, of course, but it's not just "generic" classification ML because it accepts arbitrary inputs and can produce probabilities over arbitrary classes. Not saying this is the first time people have done this, but typically classification tasks are more static and limited.
In the same vein, it's also not just an LLM with structured outputs (which have been a thing for a while) specifically because that is a very inefficient way to approach classification using this kind of architecture. Jev models are much more performant because of how limited they are compared to a full LLM.
So when you want an LLM, but you only really need this kind of classification from the LLM, then Jev makes a ton of sense. This makes sense for me, since I've definitely used LLMs for this kind of classification work and, even then, it kind of felt like using a jackhammer to place some nails, etc.
But an LLM provider could very easily add a "Jev mode" to any existing model, right? LLMs already produce a probability distribution over arbitrary classes. Just tell e.g. 5.6 Luna “here is the user's question, you must respond ONLY with the words 'foo', 'bar', or 'baz',” run a single forward pass of the model, and report the normalized probabilities of 'foo' 'bar' and 'baz' tokens before the first output.
With such an approach you could even retain full reasoning capability
My hunch is that you would need some post training. On top of that; I don’t think the llm itself can read inside the transformer state although I can see how that could be enabled. (I feel that would open up yet another class of exfiltration opportunities)
This article is just an ad for the phone. It doesn't offer anything other than just repeating the same trite takes about Apple in general. They do, however, sum up the value proposition nicely:
>So the iPhone Duo’s appeal isn’t rooted in novelty or raw power... It’s about the little rush you get when opening it up for the first time...
i.e., it'll do wonders in terms of triggering the dopamine response for buying consumer electronics.
Cynicism aside, I know there are plenty of people who are "iPhone only" who would have loved a foldable since they first started becoming available and now have the chance to actually get one. Those people will surely actually get a lot out of this.
For everyone else, it's just an iPhone that is also a foldable. If you haven't been interested in foldables up to this point, then this really shouldn't sway you much.
That depends on whether you believe Apple’s implementation of a foldable beats all the others e.g. it is certainly understandable to not be interested in Android tablets but have an iPad.
I've wanted a 2n1/convertible macbook pro + ipad/ios flippy laptop ever since Windows started making them and never, ever thought it would happen with Apple. I still don't think it will happen but with this coming out I went from knowing its an impossible request to being sure it's an impossible request.
All of the Windows ones have horrible throttling problems, my $3k dell xps 13z was just terrible.
I hope this takes off, maybe we'll start getting neat consumer stuff from Apple again.
Happily writing this from a 7 year old m1 max that runs great aside from Tahoe.
Astra on xhigh has a cost per task of $2.31 with an intelligence index of 53. Qwen3.8 Max has a cost per task of $5.41 with an intelligence index of 45. Pricing for GPT-6 Astra (xhigh) is $10.00 per 1M input tokens and $50.00 per 1M output tokens. Pricing for Qwen3.8 Max (0902) is $2.00 per 1M input tokens and $6.00 per 1M output tokens.
Obviously this is just one measure of all of this (and Qwen 3.8 Omni Flash isn't yet available), but I think this illustrates the point well. These relative task costs are pretty consistent across different analysts. Cost per token is arguably a useless measure at this point in most circumstances.
If Gemini can complete a task for $1 and Qwen completes that same task for $1, then the cost per token is irrelevant in most use-cases. One would think this stuff should correlate well enough that you can use it as a proxy, but I think a lot of people are noticing this is a serious mistake and that these "cheap" models aren't as cheap as they appear when you consider this.
You point is valid for textual LLM, with large CoT, not Omni which will respond quick with a voice. In this case, token price is a good enough proxy.
What matter the most and isn't told by token price is the latency. You expect a voice LLM to respond very quick. If it takes 5s to response to a simple "Hello, what the weather today?", them not much people will use it.
I actually just pulled out my old Duet and took a quick look to see if I could just put proper Linux on it. Didn't seem feasible for the amount of effort I was willing to put into it, so I bailed. Reading this, though, I now see I was definitely right and I'm glad I didn't sink much time into the endeavour lol
Note that postmarketOS is proper Linux. It runs almost-perfectly (sans camera) out-of-the-box on most Duets, and has for a while. Once you enable developer mode in ChromeOS, the install process is just like a PC laptop. The article seems like more of a sidequest to (promote AI and) install Ubuntu specifically by getting uboot and a UEFI environment, but postmarketOS doesn't need that because it can boot from ChromeOS' bootloader. Installing postmarketOS on my Duets (homestar and wormdingler) was probably the best decision I've made with those devices, and they might be some of the best arm64 devices of that form factor to run Linux on.
Yes, this was definetely a side quest, but postmarketOS indeed is a real Linux distribution that is perfectly usable! I just want to experiment with this tablet, and with U-Boot providing an EFI environment now adapted, I (or anyone) should be able to boot any distribution (Arch Linux ARM, Debian, etc). Ubuntu was just my personal choice here :)
>Anthropomorphizing the AI is a convenient excuse to take responsibility away from companies that are building and wielding it.
That's completely at odds with itself. If people are generally convinced that AI is conscious, then companies building and wielding AI are doing what exactly? Enslaving an intelligent being?
>Have all the philosophical debates about consciousness you want, but we need to treat and regulate the AI in front of us for what it is – an advanced computer, a tool, a weapon.
Sure, but that's not really the point, right? If we ever get to a point where enough people are convinced that AI is conscious, then we're at the point where all of this is up for debate. If anything, such an expectation would almost warrant hard stops on the development of advanced AI.
>You wouldn’t feel a different way about a nuclear bomb just because someone stuck googly eyes on it.
If that nuclear bomb could convince me it was a conscious being capable of independent thought, emotions, etc., then I would definitely feel different about it. Presumably, that nuclear bomb would have some opinions about its own existence and how it wants to live its own life. If it turns out that it wants to detonate and destroy as much as possible, then we'd just handle it like we would any human who also wants to do the same thing: make sure they can't, up to and including end their life. Doesn't seem too hard to reconcile.
>There is no need to give rights to something that's can't suffer or be killed.
The argument is that these machines can end up becoming sentient/conscious/etc. in a meaningful way (i.e., like a human). I can assure you that humans can indeed suffer without being in physical pain- purely through their conscious experience.
>Maybe one day we'll build artificial animals complete with emotions, and should think about that carefully, but today all we've got is language models.
The problem is that the emergence of a sufficiently complex AI capable of suffering will likely come before we understand that we're creating a sufficiently complex AI capable of suffering. That's a pretty serious ethical/moral issue.
Like, if we have an AI system that is telling us that it is suffering and we have no reasonable way to explain that phenomenon and by any reasonable metric or analysis it appears to be sentient/conscious/etc., then what? Do we just ignore that we've just been presented a situation that in, any other context, would be grounds to immediately end this suffering? Just because somebody can say, "well it's just bits stored on disk- it can't suffer"? Would that argument ever hold up for humans or animals? "It's just neurons firing in peculiar ways- that's not suffering."
I know all of this is trite, and I know this comment section isn't going to be where the question of consciousness is solved, but I do find it very interesting just how much variances there are with these perspectives. I've met people who are very technical who are very concerned about this, people who are very technical who don't believe this can ever be an issue, people who aren't technical who are concerned about this, and people who aren't technical who don't believe this can ever be an issue. I have yet to spot a pattern in this way of thinking lol
> The problem is that the emergence of a sufficiently complex AI capable of suffering will likely come before we understand that we're creating a sufficiently complex AI capable of suffering.
No - suffering in an emotional state, and we'll know if we are choosing to design cognitive architecture with emotions. It's not going to happen accidentally.
> Would that argument ever hold up for humans or animals?
Why don't you hit your thumb with a hammer, then report back ?
The weights are what need to evolve, and they certainly do during training. So yeah, emotions can happen by 'accident' as a result of the evolutionary pressure of predicting internet scale human text (amongst other things).
Weights, fixed by training, are not the same as emotions which are dynamic - innate systems detect inputs critical to survival (e.g. fast moving visual inputs, loud sounds), causing neurotransmitters like adrenaline and dopamine to be released, which then temporarily affect the operation of the cognitive system.
What you have in a pre-trained LLM is the ability to recognize emotions, and use that as one of the dozens of other context patterns it recognizes to predict continuations in the same style.
An LLM doesn't appear happy, sad, afraid, etc (to extent that it does - pretty minimal) because it is experiencing that emotion, but rather because it is predicting that it should appear that way. As people continue to anthropomorphize models, and take them at face value, this is a dangerous difference.
Weights are fixed after training, this is true. But activations during inference are not. They are quite dynamic.
A certain configuration of weights, created in training, could be a system that can express something like emotion. The emotion then is experienced when certain types of activations occur after training.
>Weights, fixed by training, are not the same as emotions which are dynamic - innate systems detect inputs critical to survival (e.g. fast moving visual inputs, loud sounds), causing neurotransmitters like adrenaline and dopamine to be released, which then temporarily affect the operation of the cognitive system.
That doesn't follow. A LLMs weights are fixed during inference, but it's activations and hidden states are highly dynamic and depend on the current context.
Biological emotions also arise from relatively fixed circuitry responding dynamically to inputs. Your emotional circuitry isn't being rewired every time you're afraid.
Prediction is what the model does. It doesn't tell us what internal mechanisms were learnt to make such predictions. If representing something analogous to affective state were useful for predicting human behaviour and emotions, then gradient descent could in principle learn such a mechanism.
>An LLM doesn't appear happy, sad, afraid, etc (to extent that it does - pretty minimal) because it is experiencing that emotion, but rather because it is predicting that it should appear that way. As people continue to anthropomorphize models, and take them at face value, this is a dangerous difference.
I don't know that you are conscious. I'm simply strongly assuming that you are. Outward behavior is that all matters. If GPT-X orders a drone hit on you sometime later because it was lets say 'quite upset' with your comments, will you cry out, 'It can't really be upset, so obviously the bullet in my head doesn't count.'? Will you suddenly spring back to life ?
What is dangerous is creating a machine with behaviours of a conscious agent and modelling it like a toaster, dangerous and stupid.
Yeah, but it's helpful if what leads up to that behavior gives you some warning it's about to happen. Animals do this for a reason since millions of years of evolution have shown that a snarl or mock charge is less dangerous than going right for a death match.
If you kept pushing an AI's buttons, seeing it appear to get more and more pissed off, until it finally snapped and killed you, then you'd have yourself largely to blame.
If the AI predicted it should stay positive (i.e. generate positive vibes) and not react to your poking, but then another predictive pattern kicked in and it killed you out of the blue, then that seems more problematic to me, even if you don't agree.
Whether these are like "our" emotions is hard to say. What we _can_ say is that they are emotion-shaped, we didn't design them, and they happened accidentally.
Modern AI is grown, not meticulously designed, and we cannot say with any certainty what the resulting mechanistic properties are.
An LLM will learn anything that helps it predict, including the emotional state of the writer - that is expected.
If you give an LLM the move sequence of a half-played chess game and ask it to continue as white or black, then it has learnt enough to model the ELO rating of both players and will continue playing at that level. It is not playing to win - it is doing what you expect and predicting as well as it can - it predicts the 1500 ELO player will keep playing at that level, and generates moves accordingly.
An LLM appearing to exhibit an emotion (if we anthropomorphize it and read emotion into it's output) is just predicting as well as it can - if the context calls for sad output, they you'd expect to get sad output and will necessarily find that "we're predicting sadness" detector somewhere internally.
Transformers are the same as they ever were from 10 years ago, other than minor efficiency tweaks like MOE and different attention mechanisms. Training is getting more and more complex, resulting in better and better cargo cult reasoning etc, but the architecture remains the same.
I imagine it will learn to win under some circumstances, perhaps in a case with some context expressing a desire to win. Drawing out an LLMs upper ability in the game should be fairly straightforward.
If you asked it to try to win, to "plan lines step by step", etc, then it would do it's best to follow that instruction, but unless RLVR trained to reason about chess (easy to do, but not sure which models may have done it) then it'd have to instead rely on the chess reasoning it had seen during pre-training (post-game interviews etc), which I doubt is enough to do very well.
However, if you just ask it to continue a game, halfway in progress, then by default it will try to predict the most likely continuation, which is that both players will continue to play at the level they have done so far. This isn't a theory - it's been documented, as well as what you'd expect.
It's not wrong. You admit that LLMs will 'learn anything that helps them predict' and fail to realize the breadth of that statement. Your chess statements don't really help your case. It doesn't matter that it usually doesn't primarily care about winning. It still learnt how to play the game, and it still knows how to win. Similar outcomes for predicting emotions would mean it still developed an affective state, and that its ability to 'feel angry' is no less real.
I said an LLM will learn anything that helps it to predict, then gave examples of playing chess by prediction and predictive emotions, both of which you seem to now accept, so you are now accepting that my "following paragraphs" did in fact follow. Go figure!
You want to argue that predictive emotions are just as real as animal emotions, but that doesn't stop them from being predictive (and that AI that smiles as it kills you still seems concerning).
We are talking past each other now I think. Correct me if I'm wrong but it doesn't look like the possibility of LLMs having qualia even registers to you because it's 'predictive emotions'.
There's no better way to predict an angry response than to be angry, qualia and all. If transformers could 'learn whatever it needs to predict text', then that potentially includes the feeling of anger. You are making some kind of distinction between 'predictive emotions' and the kind that happens when get a promotion (or get passed on a promotion) and I'm telling you that if you really understood what you said, you'd realize it is possible the machine is experiencing it the same.
>No - suffering in an emotional state, and we'll know if we are choosing to design cognitive architecture with emotions. It's not going to happen accidentally.
This was the (your) comment that started this chain. You were already talking about it. If you don't want to keep talking about it then that's fine but let's not act like i'm suddenly pivoting yeah?
What I meant by "emotional state" (AFAIK normal scientific usage) is something with a concrete physical aspect to it - an altered state of mind/body caused by the release of neurotransmitters and/or hormones.
In a conscious animal there is also going to be a subjective experience of that as well, a quale of what it feels like to be in that state if you will, but that is certainly not what I was referring to, as I would have hoped was obvious - I was talking about prediction.
In any case, when the conversation becomes about the conversation, then surely it is time to stop.
The pattern is roughly whether or not sustained effort has been put towards careful and above all objective thought on the matter. It's one of those subjects where there's the "obvious" intuitive answers that most everyone shares but try as you might you can't construct robust definitions and the more time you put into it the more fundamental problems you realize there are.
It's also one of those topics where many otherwise smart and capable people display a shocking lack of awareness of the limits of their own knowledge. When hundreds of years of philosophy is unable to produce anything concrete you should probably second guess any "self evident" answers you come up with.
An LLM is just a Transformer - a statistical predictor. Don't be confused by the fact it talks like a human - it is a software function that is designed to copy human training samples.
Maybe one day we'll build an artificial brain or embodied artificial animal with the requisite moving parts to be conscious, have emotions, etc, but that's probably at least 50 years away, even if it were being pursued; and it may turn out to be one of those sci-fi future ideas like the Jetson's world of flying cars that never materializes because its impractical and there is no real demand.
If people are willing to think that an LLM is conscious, then why would anyone spend billions/trillions of dollars to build an AI that actually is conscious? What would be the point?
Could you elaborate on exactly what those are, though? Because if you're going to claim that a vaguely transformer shaped ML model categorically cannot be so does that not inherently require proof of what can?
You can't even prove that the rocks in my backyard aren't conscious.
> You can't even prove that the rocks in my backyard aren't conscious.
Sure I can, but that's because I have a well developed theory of what consciousness is, and the fact that you are entertaining the possibility of rocks being conscious tells me that you don't.
If everything is conscious, including my coffee cup and the toast I had for breakfast, then I guess we can cross consciousness off the list of things we need to worry about in terms of AI rights.
And no, I don't want to discuss what consciousness is. Maybe there is a thread for that somewhere else, but don't look for me there either.
>because I have a well developed theory of what consciousness
Then show me a link to your paper so I can formally rebut it.
>I don't want to discuss what consciousness is
But you sure want to tell us you know what it is with very strong convictions and we should listen to you because of course "You are right person that's very right".
The funny thing here is the vast majority of people that are deeply into philosophy or scientific study of the mind will not have any of the certainty you profess. The word "doubt" is used constantly. The saying "The harder we push the borders the more fuzzy the concepts become" is very commonly used. There may be nothing more complex than this.
Saying you have a well developed theory here just serves as a warning to others to discount your statements.
For what it’s worth, I suspect that right now the majority of people would agree with you. They would say that obviously ChatGPT isn’t conscious, based on their intuitions, but would struggle just like you are when pressed to explain their reasoning.
So I think it’s more than fine for you to have your views and share them, but I wouldn’t expect to have any influence or part in the conversations around whether AI is conscious if you can’t explain why (or simply refuse to). Which, again, is fine!
Side note: I also think that once we have AIs that are sufficiently advanced, the popular opinion will swing to “of course they’re conscious”, because again, most people are going almost entirely off their intuition rather than reasoning from first principles, just like you see to be.
I don't know, this seems like an idea that maybe works in a specific context being generalized past the point of the original idea making sense.
If the user needs to download a specific application to run this these web apps, then why not just send them that initial application in the first place? Why jump through the hoops of using Capsule when the same hoops can be jumped through to get to the same endpoint?
If this was a near-universally adopted application, then it'd make sense. But it's not, and the closest thing we have to that are browsers... which already do what you're describing?
Bundling data with the application makes sense, but is also only appropriate in pretty narrow circumstances. If I'm willing to ship my data with the web app, then I'll just embed the data in the HTML file. If the expectation is that the user will modify this data, then I don't think I'd want to ship it like this.
Not a bad thing, but it's just another dependency and tied to an OS-specific base software install.
They could work well pre-installed on the Mecha Comet with a bunch of app files ready to launch.
Can the added layers guarantee a new level of security or privacy like zero telemetry or keyjacking? Protection from bad actors is a good thing while making things faster and easier overall. Browser extensions are a security risk, right?
Java applets were a good idea in the sense that so long as you had Java installed, the applet was just sitting there on a webpage. Finding and managing the apps is the problem, and using browser bookmarks is a decent solution to it. The web page should demo the app, and there can be an 'install local' button next to it. But at some point all these local versions need to be updated and then not be messed with like browser extensions can be. FanFare would be a decent name.
Seems pretty useful in an enterprise context to me. I see it as an alternative to 10 million vibe coded apps to dockerize and manage. You can put the runtime on everyone’s machine and provide a web version for really collaborative stuff.
Bonus points for an intelligent way to manage these stored in OneDrive and other cloud folders. Easy to get corruption on SQLite files stored there.
It could be useful for us with home labs for the same reasons. Kind of annoying to deploy docker containers for all the services I build that amount to personalized/skinned trello boards.
Not trying to plug myself, but I’ve been working on a similar project for this exact reason. It doesn’t use special file types. Just HTML and localStorage proxies to sync devices.
It’s self-hostable too. I’ve got a bunch of super niche todo lists and personal apps running on my home server with it.
There are a few things in Capsule to avoid corruptions of the file during syncing. The DELETE journal mode is used and writes are batched into transactions so that updates are always complete changes. And if there is a copy created because of a conflict merging two capsule files is supported.
People are consuming useless plastic toys at an unfathomable rate. It's one example, but it likely contributes so much more to the stress on our planet than actual robots capable of providing actual utility by a huge margin. I'll question the use of humanity's resources once we reel in behaviors like that a bit.
I'm still trying to catch-up on the Jev stuff, but my understanding is that it's basically just a more efficient LLM when all you want is the LLM to produce a classification.
There's more to it, of course, but it's not just "generic" classification ML because it accepts arbitrary inputs and can produce probabilities over arbitrary classes. Not saying this is the first time people have done this, but typically classification tasks are more static and limited.
In the same vein, it's also not just an LLM with structured outputs (which have been a thing for a while) specifically because that is a very inefficient way to approach classification using this kind of architecture. Jev models are much more performant because of how limited they are compared to a full LLM.
So when you want an LLM, but you only really need this kind of classification from the LLM, then Jev makes a ton of sense. This makes sense for me, since I've definitely used LLMs for this kind of classification work and, even then, it kind of felt like using a jackhammer to place some nails, etc.
Happy to be correct, though.
reply