On Sunday, 14 September 2025 10:13:30 Central European Summer Time Marc wrote: > Why don't you chat a bit with AI. My impression is that AI is good at > teaching you what you want to know. Quite often it messes up, but for > broader knowledge acquirement it should do fine. > https://copilot.microsoft.com/
I think this is a good idea, but only for the broad strokes. There is a recent conversation with Copilot that comes to mind. I had a practical question to Copilot when I was introduced to this newfangled GPT-5 thing in MS Edge's update changelog. So I opened the chat, and thank heavens that I don't need to be signed into a Microsoft account to use it anymore. And I saw these three modes, Fast (2-3s), Think Deeper (~10s) and Smart (GPT-5). And so I asked what determines the complexity of the model for that Smart mode. Is it like my closed beta into GPT-3, where it's just a model change (Ada, Babbage, Curie, Davinci), where Davinci was most advanced but also slowest and most token-intensive? So the answer it gave was that it all uses the same model, that GPT-5 is not a collection of models like the GPT-3 series were (though I don't think many people are aware of that for GPT-3 either). That it has internal weights that determine the complexity, and that it can be set for each conversational "turn". What it also did was immediately bring up `reasoning_effort`, an API parameter. So I am tempted to think of "Smart" as just "shoot my turn into the API with lowest complexity, and ask which value to set reasoning_effort to, then make the front-end send the prompt again with that level of reasoning_effort". Logical, simple, arguably not worthy of much hype but marketable, and reasonable to consider during a recession with so many engineers fired. But the model would not budge. The release is more advanced than that, and for sure it's just the both of them -- internal and external weights. I said thank you and moved on. But what if I didn't know where to probe further, what to believe and what to discredit as e.g. hype, what if I couldn't make comparative questions that demand the model to make a decision on its own conflicting "beliefs"? What if I was still completely in the dark, and had to take the generated responses as gospel? AI is helping me become more productive, and to learn about new subjects faster. Another example may be how I learned Arduino using it. But that took existing knowledge on Bash and Python, as well as understanding variable types and a bit of memory management (no heap but memory is counted in kB at best). And to program ATtiny85 chips, it made me waste so much time on burning their bootloader. You don't need to use avrdude for it, the IDE just has it and frequency presets. 2MHz which I made the AI settle on during a travel abroad, turned out to be invalid (only 1, 8, or 16MHz are valid, and 20MHz with external crystal). But I had no way to confirm that, so I spent a considerable amount of time on figuring out.. a dead end. So if you want to use AI to help you be more productive, do make sure to cross-verify your findings, and to hold the AI accountable for it. If necessary, go back to an earlier "turn" that undoes all of the references to the incorrect statements. It won't listen to you disagree with it, if it's just seen that incorrect statement 5 times before in its backlog. -- [Met vriendelijke groet] [Best regards] [Michael De Roover] --- --- --- --- [Mail] [*@nixmagic.com] [michael@a...@de.roover.eu.org] [Web] [https://michael.de.roover.eu.org] [Forge] [https://git.nixmagic.com] [Weather] [Antwerpen] [23:00] [13.9°C] --- --- --- --- [0] [2025-09-15 23:56 CEST] [~] [v...@workstation.vm.ideapad.internal] [$] [/usr/bin/sign-mail] [>_] --- --- --- --- -- Visit https://lists.isc.org/mailman/listinfo/bind-users to unsubscribe from this list.