One of the most obnoxious things about AI advocates is that they are awfully opinionated on my specific work even if they don't know what my specific work even is. So they see it working well in some areas and when I express a nuanced experience that isn't a slam dunk except in certain situations, they are disdainful that I am not all-in.

One concrete example, some guy who has never written or managed code (not even vibe coding), asked if I used git and I said sure, and he informed me that I should change to something that was more AI native based on what he had read online. People reading random tech articles are once again trying to "help" people that get to read the same articles, but without actual experience one way or another. At least the ones that have successful vibe coding have a leg to stand on, but still oversimplifying a very nuanced domain to a boolean 'ai can do all of it it: yes/no' is obnoxious

I doubt that. They may have wondered if it was *worth* the price of the computers at the time, but they were not debating the capabilities.

That facet certainly exists here about token pricing and how it will evolve and in a larger sense about whether it is worth the footprint, and there's every possibility that those concerns will ease with time like they did for computers doing calculations.

The issue here is that folks are treating software development as a monolithic thing. Even very milquetoast software development demanded skill and tedium, but GenAI can chew through the milquetoast scenarios adequately if you are careful and know what you are doing. People who live in that realm then seem to get the mindset that reality applies to the entire domain of software development equally. We have a much more weird subjective results than computers being able to calculate.

Broadly speaking, I find agentic coding to be even more exhausting (and expensive) than AI augmented completion. It's babysitting a mechanism that can screw up at any time, so it involves a *lot* more exhaustive work to make sure you have a harness to let the model try and retry with objective feedback very quickly. A continuing issue being that it fails to recognize an inability to do something but tries to do something anyway, which means any result that is not objectively measured in strict terms is at risk of weird drift.

Today I took a swing and it did a few weird things in my domain. Issue being here there's documentation of a standard but the very first and reference implementation deviated from the documentation, so every implementation was based on the reality rather than documentation. This particular scenario was very specific and you would have had to know very specifically about this risk to warn the model about that sort of mistake. It had a variety of legitimate inputs to process but dropped them without error because it didn't make sense and so it assumed malformed data.

It really depends on your inclinations and problem domain. If you are making a straightforward web ui interface to a fairly straightforward data set, then you can get pretty far. If you have requirements that aren't very picky and you can tolerate some drift, it can get pretty far. If you work in pretty esoteric spaces, then you have to either use it to accelerate and correct code by code or you have to spend a maddening amount of time describing all sorts of stuff to *not* do.

Completions are useful especially for boilerplate and such.

Seems like this pretty much, though perhaps the more interesting piece is that 'generate a bunch of tokens consistent with human output associated with the environment without more specific direction' influenced the models to have what seemed to be better tokens generated later when they had to now be consistent with specific scenarios.

Though I'm too lazy to read the paper and the 'generally beat' phrasing gives me some wondering about how definitive this observation is versus how much of it might be finding what they wanted to find in the results.

As much as it can apply to online communities in general including slashdot, Reddit and reddit-likes *really* take the cake for creating circle-jerk communities. Creating niche communities with moderators that ban at any comment they didn't want to see. Saw a reply to one of my comments that I thought to be a bit dumb (basically, that we should replace TLS certificates with blockchain), but not horribly unforgiveable or anything, and the comment got removed and the user banned from the community. Mod saw 'blockchain' and while I understand a negative gut reaction, nothing that should have rated a moderator going that far.

Alternatively, some people in that subreddit can't stand that anyone could express valid concerns about Tesla and decides that it must be some irrational, inexplicable blind anti-Tesla rather than informed and real concerns.

Not to worry, recall that Musk *also* has ambitions of X being a finance company, so voila, every child can have an 'X brokerage account' waiting for them with a share in it. Now he gets every child defaulting to his network, no problem.

I don't think a grant of one share of a specific stock will be a particularly educational event. I think generally in school a teach will, at some point, do a more educational exercise where students are actually doing mock trading and, critically, trying to evaluate different companies and make predictions on where things will go.

Yeah, that latter one basically focused on the fact the output looked verbatim like the 'training' material, so the judgement was that AI was more an excuse than a demonstrable transformative thing.

Note in some other case I recall the AI company arguing that the ingest and training of material is 'transformative' but the engines could be used to violate copyright, but then the *operator* would be on the hook for infringement, as the operator supplied the prompt that resulted in infringement.

There was also that OpenAI deal with Disney that was going to let Disney characters be used in OpenAI output but that got dropped. Lots of mixed assumptions in play with the AI companies asserting that they don't need permission, getting permission or forcing permission anyway, and shifting responsibility to the users.

Even if it exactly works as advertised, if you thought a rock kicked up by a tire was bad, just wait....

And thanks to hiring 1,000 new people to patch it, now they have N * 1.5 vulnerabilities!

I meant infested, not invested

*You* would not want this.

However on endpoints vaguely related to your work, your employer *loves* excuses to give Microsoft money to invest as many endpoints as possible.

Nowhere in this do I see how maintaining the history in this manner even benefits AI. I haven't seen evidence that it can benefit from an insanely over-granular change history.

This is not debating the volume of code, it's the granularity of delivering it as tiny 'changesets'. Just because you *can* create a system that can partition changes into 3 second intervals does not mean there's any value for machine nor human in doing so.

To start, the condescension is hardly helpful.

The point stands, you have to consider the audience for the change tracking. How much granularity is really useful even for the AI models themselves to digest? Certainly no human will digest, but I also don't think the AI is going to be prioritizing putting hundreds of thousands of change history entries into the context window on the off chance that it is relevant to the task at hand.

It seems like a solution to pretend around change management while abandoning any actual use of change management, by human or machines.