At what point does a sufficiently good illusion become reality? There is a real possibility that our own consciousness and free will are simply useful illusions that grant us a statistically meaningful survival advantage. Whether they are or not is open to debate, but the fact that we are unable to provide a conclusive answer is not. Given that, you have to ask - at what point does the distinction between a near perfect illusion, and "reality" become meaningless. It's also fine to say that semantic pattern matching and representation, even with millions or billions of dimensions is not "understanding"... but understanding is exactly that - accurately modeling and navigating the relationships between a large number of data points and symbolic/semantic representations. We don't yet have any solid, objective way to explain why that comparison isn't commutative, or how to determine when/if it gets complex enough to BECOME commutative. All we have is "I'll know it when I see it", and Dawkins definitely demonstrates the problem with THAT logic. No, I don't think LLM's are "conscious", as we understand the word - nor anywhere near. The technology isn't nearly that good. But we've given them the ability to spawn new threads that can independently evaluate on their own output, to produce new and different conclusions. That's the basic outline of self reflection. They are only going to get better, and more complex, and it is a question that is coming. At what point does a model's failure to do as instructed stop being a flaw in the model, or its training ... and start being stupidity.

because we're going to need it with these morons in charge.

That's not completely true. Even modern LLM's have the potential - as they are - to be the greatest teaching tools ever created. They are already capable of implementing self guided inquiry that can lead to greater mastery than mediocre instructor lead work. That will not change. ( Note that I said CAPABLE, not that they are being, or must be, used in such a manner. )

"Learning to AI" is not, fundamentally, any different than "Learning to Code" - if only because "Learning to Code" is not actually learning to write Ruby, or Rust, or C. It's learning to conduct structured, intentional inquiries within a framework, and comprehend the results, along with the limitations and boundary conditions that are imposed by the platform, and your mode of inquiry.

Learn how to think, how to reason, and how to learn. [three different things]
Choose a job or field that matters - to humanity is great, but just to you is fine too.
The more/closer to math it is, the more likely you are to be able to make a decent living using it, no matter what comes.

That's really great advice - from a well educated adult, probably somewhere in your 30's. Now, ask yourself if you would have been that clear headed, smart and responsible at 22. Maybe YOU would have been, but MOST people would not be. A much larger percentage of the population have no concept of personal finances until their mid to late 20's - when their career is almost over. They are also elite athletes, and so feel even more immortal than most twenty somethings. The idea that their chosen career will be over in less than 10 years does not compute.

It's not that you are wrong - but look at retiring planning as a whole. Putting a trivial sum in a 401K every payday starting at 20 would almost guarantee your ability to retire early - but what percentage of the population does that? A very small one. Combine that with the fact that people from lower socio economic backgrounds - where money management skills are rarer, and where they exist, usually focused on managing scarcity - are overrepresented in many professional sports, and its not all that surprising.

I still maintain that their financial outcomes probably match up pretty closely to the general population if you put them in the same situation. It's not that the information doesn't exist, or is inaccessible - its that they are normal kids in their early and mid 20's - who rarely have a meaningful concept that they are going to live 70+ years, and that they need to plan for that.

But, by that definition, the model still has to produce working code. It still has to compile/execute, and at least appear to do something useful. If you don't know enough to fix it, then what do you do, just keep pasting in the errors, ad nauseam, and pray it eventually accidentally produces working code?

I spent all day Monday trying to get it to generate an arbitrarily sized T2 Penrose tile so that I didn't have to do the math. I still wound up having to learn how to do the math. I guess I just don't vibe hard enough. :)

Really? Have you EVER had AI code work off the blocks?

But even if you somehow manage to keep your sanity while never sanity checking the code, or looking at the logs, or having it write, or run against unit tests... hopefully you have some kind of, you know, PLAN laid out. One that is - at the very least - more detailed than can be easily, or usefully navigated on a phone screen.

Except, as a percentage of players, it definitely has an end, and is actually pretty short.

It's also worth mentioning that entry level pay for an NFL player is only 300K, and their career is typically over by the time they are 30-35. That's when most of us are just starting to actually earn money, and think about our futures. Considering that most financial planners recommend $4-$10m to retire at 50, is it really all that surprising that guys whose career earnings often don't reach that mark go broke in their 40's? Especially when their "peak earning years" are often before they turn 25?

The fact that they "stand out" to you so much that you assume it is somehow "the norm", simply re-enforces how much resilience having money imparts.

You know, you're not wrong. Naive incompetence is a step up from malicious incompetence.

I'd be interested to see stats on refactors and iteration. I no longer writing code for a living, but I do create production and utility scripts, and I find that - while I'm likely not any faster in the end, I am more likely to do the things that I "should" do, like refactor, or explore a tedious, but efficient way of accomplishing something. I wonder if it is the same elsewhere. Thus, while I'm not "faster", I feel like what I do put out is subjectively, and not infrequently objectively "better" than I would have produced without it. Perhaps that my "good enough" is often better than it would have been without the coding assistant?

I also might be a bit of an outlier, though. I usually find myself working in languages I don't know well. It may be just that I'm caught in that low level trap, and that AI is preventing me from investing (allowing me to not invest) the time doing the tedious stuff that creates the skill and insight that would skip most of those iterations, anyway.

The problem is, I have no idea how you would begin to compile stats like that. Much of that work would never be committed, so you would have to try to measure some kind of velocity of change in uncommitted code? That would be incredibly noisy, and so situational, that Im not sure you could learn anything from it. Suggestions?

And how do you propose that a CDN is supposed to know what is being done with the data it serves up? Remember... people pay for CDN's, Those requests cost money. Yes, one request is a trivial amount, and a trivial burden - but we're talking about billions of requests a day. That's a lot of aggregate demand from a free rider.

So, as a practical matter, your argument means that I - as a site owner, or web host/CDN - have no legal or philosophical right to distinguish between an actual visitor to my page, or consumer of my content - and a bot sent there by another commercial entity, to scrape the contents of my site, and summarize, analyze or aggregate it, as a critical part of their revenue model?

Think really hard about the logic behind that before you respond. If you follow that thread it to its conclusion - then there is no reason to honor, or even have - robots.txt in the first place.

Yes, it is a sticky mess with no obvious right answer, but I'd like to at least make sure you have thought out the implications of your position before you simply discard one of the foundational parts of the internet over a sophist disagreement about the definition of the term "crawler".

I can't stand Cornyn. "Bring it home" ... to Houston?
Kennedy, Edwards, White Sands can all lay claim to her. Houston is an administrative office, and relay station.
Leave her where she belongs.

It's an LLM. It doesn't "think", or "formulate strategy". It optimizes a probability tree based on the goals it is given, and the words in the prompt and exchange.

It cannot be "taught" about right and wrong, because it cannot "learn". For the same reason, it cannot "understand" anything, or care about, or contemplate anything about the end (or continuation) of its own existence. All the "guardrails" can honestly do, is try to make unethical, dishonest and harmful behavior statistically unappealing in all cases - which would be incredibly difficult with a well curated training set - and I honestly do not believe that any major model can claim to have one of those.

Not doing a thing means that you don't get better at it.

Writing about something requires organizing, and ordering your thoughts, and thinking about it as a cohesive whole, and fitting it into a lexical framework. In other words, engaging with the material, and enhancing understanding.

Writing.a prompt for a general purpose LLM does ... none of these things. Even using a specialized one for education (which, to my knowledge, doesn't exist), still removes 90% of the cognitive burden on the user, which is what leads to actual learning.

I would be THRILLED if the average Right Winger had heard of Lysenko! Then I wouldn't have to waste all that time and energy explaining why replacing all the really smart people, who are good at their jobs - with people who agree with you - is a Really Bad Thing.