it’s ok they’ll just need to take even more of society’s money, energy, and water to make a second LLM that can quality check the first one, that will be its dedicated purpose, there’s no alternative

September 21, 2025 - 16:45 UTC

3
0
121
What I think could work is that their LLM understands the question and then connects to a math API to deliver the answer. Then again, people could just skip the LLM and save money and energy.
6
0
28
(this is,in fact, what it does)
0
0
1
This just changes the LLM's output from an on-screen answer to input for an API. If the LLM output is wrong, the API input will be wrong and you can still get a wrong answer (even if it's mathematically correct).
1
0
12
I know people are pushing back on you but you're right and this is essentially what Google Assistant, Alexa, Siri traditionally do. Interpret your intent and hand it off to something deterministic. It can work very well as human-computer connective tissue.
1
0
9
I suspect the issue there is that that just shifts the problem area to it 'understanding' that this is a math question, and if so, how to construct it correctly.
1
0
3
That would work if LLM's understood literally anything.

Just like how flying cars are totally viable if we make a tire than can get enough traction on the air.
0
0
4
What if we kept trying to use the calculator that doesn't work, but this time we use two of them?
0
0
67
I believe that that is what they refer to as a "reasoning model"
0
0
0
And that it cannot be fixed.
0
0
3