What'll be interesting is how it performs relative to Qwen3.8 27b when quantised to the same size. Which, of course, is a comparison they don't want to do.
Well, obviously. These newer American models aren’t competing with Qwen yet. They are competing for American businesses and organizations that can’t/won’t use Chinese models.
They are playing catch-up, but I’m happy to see new players to the game, and it will be interesting to watch how they develop
What you're missing about the attraction of humanoid robots is that you can, software permitting, drop them into an existing factory and entirely replace the human workforce without retooling. They might only be good for a single generation, but they're ideal for that purpose.
Slightly difficult to disentangle is that they might be using the verbalisation as a way to explore depth, rubber-duck style, in a way that they wouldn't over the board.
Efficiency and resilience are always in trade-off. Even if there was no other argument it would be worth having local inference as a guard against systemic SPOFs in the big providers.
> I disagree. They had to attend class and pass a test to get a CDL. They were trained to drive a truck, they can be trained to perform another task. Truck drivers are not lobotomized.
I don't think the difficulty is with the drivers, more with the logistics of retraining them all at once. As a cohort they came up through their class and training nicely spread out over, what, a 40 year period? Making sure there are that many training spaces for other trades over a much more condensed period is non-trivial.
> I don't think the difficulty is with the drivers, more with the logistics of retraining them all at once.
The only events that happen "all at once" are in our heads as hypotheticals.
In the real world, every change takes time to propagate.
It stands to reason, all of the "experts" of the world are going to be in less demand than they were before, because some percentage of people, will replace them with AI, for some percentage of tasks.
Stated another way, each of us will be consulting with AI more, and assigning some number of tasks to a non-expert, with the help of expertise from AI, and talking to human experts less frequently.
Interactions will decrease in frequency. In the commercial space, this means transactions will decrease in number and amount. This is "bad" for everyone, in the financial sense. Corporations lose revenues, governments lose revenues, employees get pay cuts, etc.
But downward trajectories on financial charts have been a problem for every fiat based economy we know of, long before AI came into the picture. The only thing people universally agree upon when sales, taxes, salaries, etc start going down, is that "confusion" is the first response, followed shortly by "panic", followed by "doing something to fix it" followed by "pretending it worked, even if it didn't."
Maybe AI will teach the world what to do when the charts turn down instead of up.
Infinite growth doesn't exist in a closed ecosystem, and we arent expanding into space yet.
For decades magazines and other media have promoted the idea that some day, technology will relieve us of the base burden of existence, and free us to do all those wonderful leisure activities we enjoy.
Now that AI seems to be able to deliver on the promise, the thing we seem to fear the most is the deflation of our own egos.
A conversation with an AI agent carries as little as $0.00 transactional cost, with a few cents in electricity expense, after the small fixed capital outlay for the computer.
A conversation with a consultant costs $200 an hour, and benefits taxing authorities, gas stations, phone companies, etc etc etc.
Again, it stands to reason, if we replace an expensive form of communication, with one that is almost free, that people are going to use it, and all those "expenses" they save on, were someone elses income.
The end game ... most of us will adapt, and some of us will self destruct during that process, and in 1 or 2 generations, everything will be "normal" again, and AI will be taken for granted, and we'll all be worrying about the new crisis.
Lets all remember, "the economy" as studied by economists, is the study of human behavior at its core, and its generally agreed that the human animals greatest natural defense is its intelligence and ability to adapt to its environment.
Keep calm. Carry on. This isn't a scarcity problem. It's an abundance problem.
Well yes, but you're still talking ~5 years versus 40. If the desire is for the state to step in somehow and smooth the transition (which it is, in a "UBI maybe?" discussion) then that's still providing an additional 700% training capacity. I don't think we've got a good example from history of that working in a way that lends itself to "Keep calm. Carry on." Certainly not outside military build-ups or post-war reconstruction.
"A tad messy" really undersells it. You need a solid process to manage the (sticky, carcinogenic, smelly) liquid resin both as an input and waste product, along with having to dispose of (carcinogenic, smelly, flammable) cleaning solvent.
The problem isn't so much the process of a successful print as how you deal with a failed one. At that point you've got a vat full of resin with some solid parts stuck to the bottom, so you need to filter and clean the damn thing which means taking it out of whatever containment you've got. And that step you can't use the usual solvent for because it damages the sheet. But you'll have ample time to figure all this out because you'll get loads of failed prints as you figure out how slicing and supporting for resin works: it's different to FDM, some knowledge carries over but not that much.
Focusing on the 10-30 minute curing process ignores the part prep step: SLA is much more reliant on supports than FDM in general, so there will be either rough surfaces hopefully out of sight, or (if it's not a model that orients well) scraping/sanding/filling. And you want to do support separation and get that as clean as possible before curing, while the resin is still a bit soft, but that means it's only partially cured and you need to be PPE'd up.
When I'm doing anything with my resin printer I don't even go into the room without a breathing filter, don't touch anything without gloves. It really is a very different experience to FDM.
Not the GP but the experience of being on the wrong end of a vociferous false positive is not great, and there are enough readers to make that not an unlikely outcome.
I'll probably get a kiosk panel next time. But this time, yeah, I've just not configured the network. All content comes from the console. Which is a different bundle of problems and no better in a lot of ways, but it's more of a known quantity.
Positive note from me: one thing I've had good experience with is following the path from a project that does something similar to what I want; extract a tech-neutral prose spec from the source; convert that prose spec into a formal spec; use that spec to generate valid traces, a language-specific interface layer, and conformance checks to test whether an implementation of the interface generates good traces. Once you've got that, implementing the system is a matter of colouring the interface in by numbers, which is really easy for a model.
The bonus here is that you can cut in at the prose spec stage and change it to whatever you need, as well as being able to target it at specific technical choices at the last couple of steps.
Where I've seen this work really well is with privilege escalation management, where you do actually want to be careful about the states the system can get into, and the APIs available for managing privileges in a system might be super specific to that system. But I can imagine the workflow being quite a useful one any time you've got anything state-machine flavoured.
A trace is basically a sequence of states. If you're using TLA+ or something built on it (quint in my case) then you can get it to dump out valid sequences that you can then compare to output produced by your implementation. You have to explicitly instrument where the implementation performs the updates, so it's not free and there's potential for error there, but it clearly signals "this bit of code is responsible for that bit of the model".
The point is that even if we did get to net zero today heating would continue to rise for another 15 years. That's what the lag is. Whenever we get there, add 15 years.
True, but not much of an issue I think. I'd assume we reach net zero very gradually and fifteen years before net zero will be pretty close to zero already. But oh well, first we need a clear downward trend of global emission.
reply