This is why we need some kind of professional accountability for software engineers. This behavior is willful malpractice, and it only flies because they know they'll never face consequences when it goes wrong. Let's change that.
I was with you right up until the final paragraph, but this made me do a double take:
> OpenAI is too important to trust sama with.
...wat? They made a chat bot. How can that possibly be so existentially important? The concept of "importance" (and its cousin "danger") has no place in the realistic assessment of what OpenAI has accomplished. They haven't built anything dangerous, there is no "AI safety" problem, and nothing they've done so far is truly "important". They have built a chat bot which can do some neat tricks. Remains to be seen whether they'll improve it enough to stay solvent.
The whole "super serious what-ifs" game is just marketing.
Yeah the whole fearmongering is clearly just marketing at this point. Your LLM isn't going to suddenly gain sentience and destroy humanity if it has 10x more parameters or trains on 10x more reddit threads.
I'm not even sure we're any closer to AGI than we were before LLMs. It's getting more funding and research, but none of the research seems very innovative. And now it's probably much more difficult to get funding for anything that's not a transformer model.
> I'm not even sure we're any closer to AGI than we were before LLMs.
I mean this is very obviously untrue. It'd be like saying we aren't any closer to space flight after watching a demonstration of the Wright Flyer. Before 2022-2023 AI could barely write coherent paragraphs; now it can one-shot an entire letter or program or blog post (even if it's full of LLM tropes).
Just because something is overhyped doesn't mean you have to be dismissive of it.
In hindsight there's an obvious evolutionary pathway from the Wright Flyer to Gemeni/Apollo/Soyuz.. but at the time in 1903 there absolutely was not, and anyone telling you so would be a crank of the highest degree. So it may turn out that LLMs have some place on the evolutionary path to AGI, or it could turn out they're a dead end like Cayley's ornithopters. Show me AGI first, then we can discuss whether LLMs had something to do with it.
In order to get to space, you must first be capable of flight through the atmosphere. That should be apparent to anyone even then because the atmosphere is in between space and the ground.
Regardless of whether spaceflight is still 1000 or 100 or 50 years away, you are still closer than you were before you demonstrated the ability to fly.
Or we could be stuck here for decades pending a breakthrough nobody alive today can even conceive of, or we could be compute limited by a half dozen orders of magnitude. Or it could happen next week. That's the nature of breakthroughs--you just can't have any idea when or how (or if) they'll happen.
It seems to me the answer is more "stay away from Bobs". If you don't get too tangled up with those guys when the shit hits the fan you won't get much on you.
I guess a counterpoint might be Apple's "strategy". Scare quotes because I truly don't know if it was deliberate or just a happy accident. But somehow they've managed to not get so intensively exposed to the downside risk--if the wild claims about AI don't pan out they're not going to lose very much compared with the other megacorps.
Apples plan has been pretty obvious. They invested in small locally running features that provide small utility rather than massive hosted models that cost a fortune and aren’t profitable.
There also doesn’t seem to be much risk in falling behind. If you wait longer you can skip buying the now obsolete GPUs and training the now obsolete models.
I wonder what was the imagined use case? TBH I was seriously thinking about buying a framework desktop but the NPU put me off.. I don't get why I should have to pay money for a bunch of silicon that doesn't do anything. And now that there's some software support... it still doesn't do anything? Why does it even exist at all then?
At least part of it is probably Microsoft's 40 TOPS NPU requirement for their Copilot+ badge. Intel also have NPUs in their modern CPUs. Phones CPU manufacturers have been doing it even longer, though Google calls theirs TPU.
I use an older Google Coral TPU running in my home lab being used by Frigate NVR for object detection for security cameras. It's more efficient, but less flexible than running it on the GPU.
Don't know if I need an NPU for my daily driver computer, but I would want one for my next home server.
Small models aren't entirely useless, and the NPU can run LLMs up to around 8B parameters from what I've seen. So one way they could be useful: Qwen3 text to speech models are all under 2B parameters, and Open AI's whisper-small speech to text model is under 1B parameters, so you could have an AI agent that you could talk to and could talk back, where, in theory, you could offload all audio-text and text-audio processing to the low power NPU and leave the GPU to do all of the LLM processing.
That seems like a really niche use case, and probably not worth the surface area? The power savings would have to be truly astonishing to justify it, given what a small fraction of compute time your average device spends processing voice input. I'd wager the 90th percentile siri/ok google/whatever user issues less than 10 voice queries per day. How much power can they use running on normal hardware and how much could it possibly matter?
It's just an example where it fits perfectly, and it's exactly what something like Alexa or Google home needs for low power machine learning, eg. when sitting idle it needs to consume as little power as possible while waiting for a trigger word.
Any context that needs some limited intelligence while consuming little power would benefit from this.
You could always offload some layers to the NPU for lower power use and leave the rest to the GPU. If the latter is power throttled (common for prefill, not for decode) that will be a performance improvement.
You want routing to be as quick as possible, because there are dependent loads of expert MoE weights (at least from CPU in most setups, potentially from storage) downstream of it. So that ultimately depends on what the bottleneck on that part of the model is: compute, memory throughput or both? If it's throughput, the NPU might be a bad fit.
The NPU is entirely useless for the Framework Desktop, and really all Strix Halo devices. Where it could be useful is cell phones with the examples mentioned by @naasking (audio-text and text-audio processing), and maybe IoT.
You can't actually push back as an IC. Tech companies aren't structured that way. There's no employment protection of any kind, at least in the US. So the most you can do is protest and resign, or protest and be fired. Either way, it'll cost you your job. I've paid that price and it's steep. There's no viable "grassroots" solution to the problem, it needs to come from regulation. Managers need to serve time in prison, and companies need to be served meaningfully damaging fines. That's the only way anything will get done.
I'm hoping the Ladybird project's new Web browser (alpha release expected in August) will solve some issues resulting from big tech controlling most browers.
> There's no viable "grassroots" solution to the problem, it needs to come from regulation. Managers need to serve time in prison,
No, yes
Yes, giving these people short (or long, mēh) prison sentences is the only thing that will stop this.
No, the obvious grassroots response is to not use LinkedIn or Chrome. (You mean developers not consumers, I think. The developers in the trenches should obey if they need their jobs, they are not to blame. It is the evil swine getting the big money and writing the big cheque's...)
Yes, what I meant was there's no way ICs will change any of this. Using this or that extension, or choosing not to use some service won't really change anything either. The popular appetite just isn't there. Personally I use a variety of adblockers and haven't had a linkedin or anything for many years, but I fully accept that's an extremist position and most consumers will not behave that way. The only way these companies' behavior will improve is when they are meaningfully, painfully punished for it. There's very little we as consumers or ICs can do until then. Unless of course their risk management fails and they alienate a sufficiently large number of users that it becomes "uncool" to use the product. But all we need to do is look to twitter to see just how bad it'll get before then...
> soon you'll need to pay every website 5.99 a month
No, I won't. I'll just stop using them. So will almost everyone. I don't think there's a single ad-supported product that would survive by converting to a paid subscription, because they're all so profoundly unnecessary.
reply