abm53

I’ve seen people recommend writing “failure is an option” into AGENTS.md as a non-training based crutch.

You’re trying to make a logical argument from first principles about a complex, dynamic and ultimately social system that admits no such argument.

I wouldn’t say the pessimists fall into that category.

In my experience they are mostly the subset of engineers who enjoyed coding in and of itself and ——in some cases—— without concern for the end product.

I think GP was a joke about the ability of a typical programmer.

I certainly read it as one and found it funny.

The article doesn’t present any hypotheses regarding this, and I suspect we simply don’t know yet.

But if true presumably it’s one of the usual reasons for observing data with low likelihood according to a model: misspecification or statistical bias/variance.

The rebuttal to this would be that you can do many such tasks in parallel.

I’m not sure it’s really true in practice yet, but that would certainly be the claim.

I’m unsure exactly in what way you believe it has gone “down the hill” so this isn’t aimed at you specifically but more a general pattern I see.

That pattern is people complaining that a particular model has degraded in quality of its responses over time or that it has been “nerfed” etc.

Although the models may evolve, and the tools calling them may change, I suspect a huge amount of this is simply confirmation bias.

> If you're the type of programmer who thinks of yourself as just a programmer, and take pride in your secure code, ability to optimize functions and algorithms, you're exactly the kind of programmer AI will replace.

The most successful engineers are the ones who can accurately assess the trade-offs regarding those things. The things you list still may be critical for many applications and worth obsessing over.

The question becomes can we still achieve the same trade-offs without writing code by hand in those cases.

That’s an open question.

More to the point: is randomness of representation or implementation an inherent issue if the desired semantics of a program are still obeyed?

This is not really a point about whether LLMs can currently be used as English compilers, but more questioning whether determinism of the final machine code output is a critical property of a build system.

Correct… reading code is a much more difficult and ultimately, productive, task.

I suspect those using the tools in the best way are thinking harder than ever for this reason.

My advice: keep it on a tight leash.

In the happy case where I have a good idea of the changes necessary, I will ask it to do small things, step by step, and examine what it does and commit.

In the unhappy case where one is faced with a massive codebase and no idea where to start, I find asking it to just “do the thing” generates slop, but enough for me to use as inspiration for the above.

> They are prone to nervous breakdown, social withdrawal, and anxiety if anyone within earshot goes outside of the guard rails for acceptable speech.

I say this with sincerity: I have met precisely zero young people who I think come anywhere close to this description over the last decade.

I’ve seen it in the online world, yes, but this tends to amplify the very very small minority who (on the surface) appear to fit your description. And I see it across all age ranges and political persuasions.

I’m trying to figure out if you’re asking a leading question and if so, in which direction…

We could make the distinction between price discovery, i.e. what price are people currently willing to buy and sell at (short-term) vs value discovery (long-term).

Perhaps the best source would now be the statistics of LLM queries, if they were available.

Edit: I see they raise this point at length themselves in TFA.

I’d go further than the other reply: not only do those first two things definitely exist, they probably represent the plurality of programming tasks.

I assume the original reply was addressing the “never” in this specific point:

“The fact is most ordinary mortals never get access to a fraction of that kind of power”

The most obvious difference (and one worth much more than $10 to me) is that one is native and the other is not.

I am also constantly astonished.

That said, observing attempts by skeptics to “unsuccessfully” prompt an LLM have been illuminating.

My reaction is usually either:

- I would never have asked that kind of question in the first place.

- The output you claim is useless looks very useful to me.

Perhaps you could fill in a few of the details for us?

I’m not sure how this relates to the point raised.

I agree.

In my view many of these small regions (that blend into one another) could be combined to give a much more useful map with more sharply distinct accents.

Such a map may be less precise, but far more useful to most.

I assume PaulDavidThe1st was asking for one which actually supported the assertion.

In the U.K. I remember it being a novelty flavour for a brief period in the 90s.

I’ve not seen or heard of the idea since then, although this may reflect my own consumer preferences.

100% correct, but there are ways to push Bayesian inference back a step to justify this sort of thing.

It of course makes the problem even more complex and likely requires further approximations to computing the posterior (or even the MAP solution).

This stretches the notion that you are still doing Bayesian reasoning but can still lead to useful insights.

I did raise an eyebrow at it too, but I doubt his contribution was entirely “management of resources”.

I think one must also give him the credit for the vision, risk taking and drive to apply the resources at his disposal, and RL, to these particular problems.

Without that push this research would never have been done, but there may have been many fungible people willing to iron out the details (and, to be fair, contribute some important ideas along the way).

I’m not a proponent of the “great man” theory of history, but based on the above I can see that this could be fair (although I have no way of telling if internally this is actually how it played out).

I think this illustrates the biggest failure mode when people start using LLMs: asking it to do too much in one step.

It’s a very useful tool, not magic.

And I would go further… these “common tasks” cover 80% of the work in even the most demanding engineering or research positions.

> The hedge funds (usually) discount the expected value by a ton when deciding how much to pay.

That’s called a risk premium.

It’s often a dangerous kind of answer.

I once saw somebody say something to the effect that in every Hacker News thread, there is always a highly upvoted comment that sounds completely plausible, well argued, made by somebody who appears highly qualified to answer, and that is completely incorrect.

I don’t think it’s always true, but it often is.