Yes, this is the crunchy bit and we're already well underway on that curve. I haven't touched a frontier model in months because I can do with smaller ones without wondering/worrying when I'll get banned next because what I'm doing is dual use. Qwen was the the first time I thought oh, ok, so this is 90% of what I need. Then DS4 made that 95% and with GLM 5.3 it's somewhere between 99% and 100%. Now, I'm not trying to solve any Nobel prize winning stuff here but if that's the kind of customer that will remain I don't think the frontier model providers are going to like the next six months or so.
That's the crux of the issue. You think that GLM 5.3 is all you need.
However, the expectations for AI are not to make your software engineering work 100% faster.
It's, for example, to build software more or less end-to-end with minimal input and feedback from you.
Then next, robotics and boosting manufacturing. Not by 10% but let's say 200%.
Such are the ambitions. Can GLM 5.3 do it?
The crucial question is whether eg. Fable 6 can, and with what lead over the competition.
Open-weight models still have not beat February's Mythos.
I am actually doing robotics stuff and yes, GLM 5.3 can do it.
As for the manual feedback: I want that manual feedback, but if I let it rip it will do just fine all by itself, but then I have no clue what is going on under the hood.
At some point you have to let go of trying for full understanding and control, delegate to the AI, and unlock the real productivity boost.
I have no clue about the state of robotics.
For software, I know that Opus 5.5 can handle more than GPT 5.6 Sol. I am willing to pay for the difference rather than use the cheaper DeepSeek Flash.
> At some point you have to let go of trying for full understanding and control, delegate to the AI, and unlock the real productivity boost.
The prerequisite here is that you can trust it not to do something that gives criminals access to your bank account or allows users to consume an unlimited amount of services it costs you money to provide for free or bankrupts you through legal liability etc.
It's basically this:
"A computer can never be held accountable, therefore a computer must never make a management decision."
Notice what this is actually saying. It's not saying that the computer can't generate a given output. It's saying that if the computer makes a mistake and you don't bother to check, you're the one who has to eat it.
> "A computer can never be held accountable, therefore a computer must never make a management decision."
That's a great way of putting it. So far I always used 'you can outsource work but you can't outsource responsibility'.
> "A computer can never be held accountable, therefore a computer must never make a management decision."
Legendary quote from an internal 1979 IBM training document [1]
https://simonwillison.net/2025/Feb/3/a-computer-can-never-be...
I'm getting plenty of mileage without that and anything with a motor in it requires that you do your utmost best to ensure it works within the design envelope.
We aren't going to get useful humanoid robots while understanding the implementation fully.
If you can't leverage AI to build a safe and useful robot without fully understanding the implementation, the model is not good enough.
There are so many assumptions in these two lines that I don't know where to begin. I wish you best of luck.
The humanoid robot was an example.
My argument is that if safety and liability make the use of AI without fully understanding the implementation impossible, the model is arguably not good enough.
A long time ago we decided as a society that in order for an engineer to sign off on physical designs, they have to take legal liability for them.
No such process exists for a model to take a design seriously that can kill people.
For serious engineering work not only physical designs, software also.
When we manually tested nuclear power plant controlling software, each test case was printed out on paper and tested and then signed by the tester. A lot of paper was produced and archived for each software release. (Archived for as long as the software was running in any nuclear power plant).
Everyone understood, that in a case of significant accident at the nuclear power plant, accident analysis would be done and if the root cause was an software error, this could have possible legal consequences for the developers and testers. (Of course nuclear power plants have many independent and redundant hardware controlling mechanisms, but incorrectly working controlling software could damage the power plant in a very costly way).
With a larger project, I do not think there is a requirement that the engineer signing off understands every implementation detail.
I imagine that the requirement is that after testing, you are confident that the product is safe.
Respectfully: you are out of your depth. Maybe pick another subject?
It was you who shifted the conversation away from software to make an argument that "anything with a motor" requires full understanding of the software implementation, and that for this reason GLM 5.3 is good enough.
I think you're entirely wrong, and we can leave it at that, seeing that you're clearly unwilling to engage constructively.
The practice of deliberately engineering software without full understanding of the implementation, especially when it comes to robotics, should be illegal. Technology serves society alone and it should progress at a pace that is sustainable for adaption. Your excitement for a “singularity” does not outweigh someone else’s health and safety.
Nah man, no you don’t. There are very few absolutes in the world and that certainly isn’t one of them.