I completely agree, but I also think that an industry disrupting architecture tweak akin to the “Attention is all you need” is VERY possible to emerge at any moment.
It feels like the cognitive gaps on current LLMs are indeed structural, but also that if we solve that structural issue with a new or extended transformer type of architecture, we’ll be looking at a whole new ballgame.
I mean, basically we’re just looking at needing some type of new post training learning architecture. It’s very clear that extending context windows isn’t that. What’s needed is an honest to god, continuous learning and modification process.
Because the space of people interested in such things is relatively small and so a single article has knock on effects where a reader of the article or a blogger sees it and starts exploring the space and posts more about it, increasing the exposure some more.
I would think you could, in principle, deduce the bad key from a recorded segment of the jamming signal, and only revoke when identified.
I would have to look at the math more specifically, but I would think it would be impossible to both disguise the compromised key and take advantage of the key to interfere with the signal in the way its designed to prevent. And even if you did, I’m fairly sure the jamming would only affect users whose decryption runs through your branch, and in your geographic proximity so I don’t think it would have an advantage of countering the below the noise floor decryption designed to mitigate against jamming.
Isn’t this the exact problem tree-based broadcast encryption schemes were designed to solve? You could surgically revoke the keys of a bad actor, and I’m not exactly sure, but I think the scope of their ability to affect the jamming resistance of other users is necessarily limited by the tree.
When riding a motorcycle, you’ll encounter people that don’t see you almost every trip. The same is not true in a car.
Riding a bike is just a 100% engagement thing with higher risks and lower margins for error, for all kinds of reasons. And it’s not just traffic, minor pavement imperfections become relevant, the necessary skill floor is also higher. It just demands more attention, straight up.
In a car, you shouldn’t, and it’s not without risk, but you CAN occasionally get away with minor distractions: adjusting the radio, seat, etc. That just doesn’t work on a bike as well. I’m failing to properly articulate the why, but it really is fundamentally different in some ways. I’ve spent many years doing both, and the bike just demands more of your attention resources, independent of your vulnerability in the event of an accident.
The cynic in me thinks its an extension of the NSA having long ago switched from being defensively helpful to US companies, to deliberately introducing backdoors and issues that they can exploit.
This matches my experience. LLMs are great at speeding up the easy but tedious bits, but they can’t even come close to independently identifying and solving the issues that come up in the development process which has always been like 90% of the work.
I definitely can do more with them, but the speed up isn’t anywhere close to an OOM. I would say they let me drift into scope I probably didn’t understand when I started, and that can be dangerous waters, but I absolutely use them as a tool to speed up my understanding of that new territory. And I wouldn’t say that pushing the boundaries of my understanding on ambitious projects is in anyway unusual for me. It’s just that I can push a little further a little faster.
Or perhaps you meant Q clearance nuke stuff? That would be QUITE a bit harder to find and illegal to share. But it’s lack of availability is hardly a counterpoint to the comment you were replying to.
Admittedly my knowledge was based on work I did in academia, and I now work in transportation, so I suppose it’s possible I’m in error, but I’d be surprised if much survives the RO stage, and isn’t eaten up by the oxidation stage. My understanding was that the water needs to actually get remineralized to protect the distribution system. And that it’s very devoid of pharmaceutical contaminants by that point. I was unaware of this being an issue in real world potable reuse systems. Though, I suppose different jurisdictions may have different standards. My state was pretty strict.
That’s been a solved problem, engineering-wise, for a while.
The advanced treatment stages take care of it. Between UV, ozone, and nanofiltration, etc. we can remove the pharmaceuticals.
Actually the problem is the water comes out too pure out of a well designed water reuse system, to the point where the mineral content can be too low and you need to add some back in.
It’s pretty common in the security world to have a red team and a blue team. There is overlap in the skillset for both, but there are good reasons to have separate people develop each team, and we wouldn’t expect people to have a talent for both.
Ideally, we like it if the red team can suggest solutions, but that’s not always their job or expertise and I’ve rarely if ever heard someone express the sentiment you are within that context by suggesting a really good red team person isn’t useful if they can’t fix the holes they find.
This quote resonated with me: “ Tom Hudson told me, “If you’re going to use an LLM to write me an email, I’d much rather you just send me the prompt; at least then I’d have an idea of what you actually meant to say.”
In my personal life I use AI a lot for information discovery and high level discussion of the problem space. I use it occasionally to write some prototype code to get started on something. It makes a great debugging and problem solving tool, though I typically find that I need to have an idea of what the problem is to steer it in the right direction. It makes a poor intuition generator, but a great intuition checker and can run with an idea for much faster iteration. I use it essentially zero in my day job as an civil engineer though.
I would essentially NEVER use it to write an email. By the time I’ve specified what it is I’m trying to say, I’ve basically said it. Wordsmithing beyond that usually has almost zero value. Same frankly with writing engineering reports. By the time I’ve told it what it needs to say, I’ve basically written that section. In general, I feel like LLMs are just bad writing tools… In writing I typically find that if I can farm it out to have an LLM write something, then it frankly probably just didn’t need to be said.
Actually the fact that we don’t have immortal mice is what makes me feel like this current optimism with respect to life extension is pretty unrealistic.
You’re not wrong, and this is speculation, but I suspect you’re just missing the subtext I added in my edit: that some people are burnt out on the evangelism of agentic workflows in the same way they were about blockchain or whatnot.
It feels like the cognitive gaps on current LLMs are indeed structural, but also that if we solve that structural issue with a new or extended transformer type of architecture, we’ll be looking at a whole new ballgame.
I mean, basically we’re just looking at needing some type of new post training learning architecture. It’s very clear that extending context windows isn’t that. What’s needed is an honest to god, continuous learning and modification process.