They could, but why? US and China has poured Trillion of USD into training and any semblance of getting those money back seems like some far fetched dream. Currently there is no realistic path to profitability with these models.
What I see as usable product in the future are smaller specialized models which are able to run and be trained with fraction of resources what goes to current frontier models.
I see it exactly the same, self hosted LLM will be the future. They may not be SOTA, but it is better to have a Trabant in your garage than being denied use of shared Ferrari, because somebody outside of your control had hissy fit.
I like a good car analogy, but it doesn't hold up in this case. What am I going to do with a Ferrari that's durable and makes me money, the same way that Mythos/Fable can output me better code than Opus 4.x? I mean I guess I could take a picture of me in a Ferrari, and then when I show off that picture but can't produce the Ferrari, I look dumb, but Mythos generated code/artifacts are still downloadable and runnable, if you got far enough before the cut off. Judging from the quality/availability of the few days we had it, it's a shared resource anyway. The usage limits were too low to give everything to Fable to do.
I am completely calm regarding AI and development.
First nobody sane want to give their domain IP to OpenAI/Anthropic. That's why local AI will eventually prevail and flourish because people who actually have some IP will have no problem to buy 10k+ EUR machine to run some pretty good models on it. However if your main job is just doing CRUD stuff, then you are screwed.
Secondly hallucination is really Achilles heel of every LLM. Sure you can recreate an application which exists in thousand of variations on the internet, but the moment you will try to go more into domain knowledge you will start struggling more and more.
Try to make CAN driver for ESP32, easy it is probably going to work. Try to make CAN driver for STM32F7xx now the AI will start having a problem but probably will be able to produce something what is working after a lot of debugging. Now let's make CAN driver for MPC5555. AI will start writing fairy tales about registers which do not exist. All of processor above have reference manuals and sometimes example git repositories available on open internet.
> First nobody sane want to give their domain IP to OpenAI/Anthropic. That's why local AI will eventually prevail and flourish because people who actually have some IP will have no problem to buy 10k+ EUR machine to run some pretty good models on it. However if your main job is just doing CRUD stuff, then you are screwed
Replace OpenAI/Anthropic with AWS and this is not too dissimilar to the arguments in 2009 about cloud providers.
It’s not that there's nobody for whom this is true, it’s just that there’s enough of everyone else to build an empire with.
> It’s not that there's nobody for whom this is true, it’s just that there’s enough of everyone else to build an empire with.
But those everyone else are racing to the bottom because all their ideas are being soaked up by AI and then being given to their competitors on a silver platter as AI output.
These companies were/are blatantly breaking copyright, going as far as downloading data from illegal torrents. Only differentiator between these AI companies is who has access to more data and create better models from them. Please don't be naive that some disclosure policy or retention agreement would stop them from getting more data and thus getting edge above competition.
Did you try this by giving it access to the materials? Human programmers also don't memorize all this stuff. If this is the reason for your calmness it's quite shortsighted.
There are problems when you rely too much on AI generated code, but these shallow dismissals are quite annoying.
1. There can be massive differences between chips which sounds plausibly same and thanks to the way how LLM is working, models are mangling these variations together
2. Registers are often named in very way similar across different manufacturers so models are making up registers in MPC5555 which are coincidentally registers in Renesas processors doing same thing.
3. There are no standard in reference manuals, sometimes there are literally missing chunks of knowledge thanks to translation to English or there are pieces which you can only get from Application Notes which has code as a screenshot.
And then you will find out that all those descriptions are wrong and through trial and error you will get it working in 2 weeks time.
Bonus point: Random people having public Git repositories for obscure processors, but with bad or completely non working implementation of drivers for them. However LLM will just output variation of this garbage on you, because there are 3 public repositories on the whole internet. Sometimes I have a feeling that this must be on purpose to poison the well.
Ok but it takes weeks of trial and error for expert humans. If you allow the llm to run the code and see the results and do trial and error, it will also likely figure it out, no?
That's not how it works in embedded world. If you set your registers wrong, you can have i.e. double the clock, then your quanta on CAN is not correct (they are half) and you have double the baudrate or set point off. The thing is that LLM does not have implementation of target processor, so it can't run the code nor it has implementation of peripherals so it can't know if CAN driver is implemented correctly or incorrectly.
Why can't it do that? How do you do it then? An LLM-based agent can execute any commands in a terminal, including running whatever commands that will send commands to an external device through whatever wires you have set up. Is the claim that you need to have hands-on access to fiddle with the embedded device? That there is no way to just send wire signals out from the workstation towards the device that could load the program, see the results etc?
Just Gabbard trying to push forward Russian propaganda. We have Ukrainian super soldiers being bred in US biolabs, or we can have Ukraine using black magic to fight invasion.
The biolab conspiracy is so old, that nobody even cares.
I am not reading either for pleasure. I am reading so much during my daily life (Documentation, coding, manuals, logs) that reading for pleasure sounds like a bad joke.
I can empathize with the commenter you're replying to. I still read for pleasure and have to read constantly for work of course. On some days, I feel like I can't. It is like exercising the same muscle every day which will lead to injury in a short amount of time. Switching to audiobooks or just taking a break does resolve that.
From what I am seeing around me though, reading for pleasure or even to gain knowledge has decreased a lot. I noticed it in adults first and this is being reflected in their children.
Except you don't get injured. You know what happens when I read a book at the end of the day after working all day? I fall asleep about 20 pages in, and sleep like a rock. You know what happens when I read HN at the end of the day after working all day? I'm up till 3am and getting 5 hours of rest. Same for any screen time really. People think screentime is some big rest but it isn't really.
Just rip the bandaid off. Get over the inertia of doing the task. You won't injure yourself reading a damn book.
This is why I prefer audibooks piped through my whole house sonos. It allows me to consume long form while walking around the house completing various projects. Reading isn't a chore, but the thing I engage in while doing chores.
It is exactly same like when OEM will make you sign agreement that you won't try to reverse engineer the car, but if you will flip it without the restriction, then all is clear.
What I see as usable product in the future are smaller specialized models which are able to run and be trained with fraction of resources what goes to current frontier models.
reply