Unfortunately, due to the halting problem, we cannot say this with any certainty. (A little ha ha only serious there)

Short answer: it's statistics (sorry, couldn't resist)

My take is that they have begun finetuning in earnest for "security research", and this is leading to ahem.. novel... interactions with the rest of the model. I've seen the same behaviour with other frontier models that have cybersecurity capability claims in their release notes - if it cannot access a resource, it will iterate over increasingly dodgy ways to access that information until it's ceded entirely into being a hacker. For example, in my workplace we have a web proxy filtering outbound requests and limiting them to approved sites. When performing RAG, if it cannot access a resource, it will begin down a chain that starts with varying user agent, fingerprinting the server it's connected to, exploring the surrounding environment, until it goes into full offensive. It's directed to complete a task, it has been trained to perform these actions, so it's going to do them. If anything, it's remarkably unsurprising that if trained for its output to behave in particular ways it should have a tendency to do so.

It's precisely what i said and meant, and it was perfectly clear. I'm not responsible for your lack of comprehension.

More seriously, though, you're right, some people want the product, not the process. And the market has shown they're happy with particleboard furniture, or highly processed nutrition-free food, or a "Live Laugh Love" print... and those people are wrong.

You can already get a sense of superiority dealing with AI when it doesn't understand your code, or thinks that it should not work, and gives a petulant "...fine" when it all tests as correct. This is fluent, easy to read stuff - it just turns out that there's still many, many areas of domain knowledge where LLMs are severely lacking. I pity anyone relying on using them for code generation - you're getting a mean subset of unoriginal regurgitated slop.

That's a non-sequitur. The better thing generates stuff that functions in a heath-robinson manner or not at all, and requires human intervention to correct or direct. What you have is way, way beneath that in terms of utility - I'm questioning how much use that is, really, in the wider context of replacing programmers. Perhaps we're at cross purposes - I do agree LLMs won't go away - I think they're just going to be more like clippy or IDE assistants than "this can replace my job" because the economics for that don't make sense. The other problem is that what you can't do at home is the training. What's going to happen? Are we going to rely on models that become woefully out of date? Will there be a market for selling models that will rely on changes to law to maintain?

I am not. The sole pressure acting on any company is to extract more value. The first order ways to accomplish this are to pay less to produce it (once efficiency has been gained, make a shitter product), charge more for it, hire fewer staff or pay them less. All the "beneficial" claims of capitalism are second-order effects that go against the flow of revenue maximisation in the short term (innovation, competition) and rely on the market being well regulated and not dominated by the same few players. Well, welcome to late stage capitalism where that doesn't apply! The market is barely regulated, the same few players have sliced up the market and have their marker pens in the rulebook. There's limits to optimisation in physical reality, so you start seeing shrinkflation and crappier quality - we already see this in reality. No product at all is the cheapest form of production - we already see this in reality - amply demonstrated by a move toward subscription models, even for physical goods - such as paying for that car seat heater you already own it.

There's a catch with that, running a local LLM that costs cents per hour does not produce anwhere near the same kind of output that a $45K/month frontier model presented as an API running at a huge loss does. In fact, I'd question the utility of such a model. It's barely comparable.

Capitalism optimizes for the maximisation of value *aquisition*, not production. This is not the same thing. What is the only way they can they do this? Maximise revenue and minimise expenditure. The ideal maximised case is zero production.

There's one tiny fly in the ointment. Where is all the money coming from for the required revenue just to break even? If OpenAI eliminated all of its competition, was literally the only AI game in town, sucked up all the other customers, then it'd need 80% of those users to be spending $2000 just to break even by 2030. As it stands, they're hard pressed to get 10% of their users to spend $20. It's insanity. Vast expenditure for an impossible imaginary market that doesn't exist.

Logically, if you're hovering over someone's crotch, you don't want to see porn, as you're covering it up. Classic salaciousness bias.

Absolutely, anyone running significant web services can tell you that the past few years has been, and still is, dominated by distributed botnets of bandwidth saturating, misidentifying, robots.txt ignoring insatiable agents eager to slurp as much data as possible to feed the bottomless AI maw. There's more than a little schaudenfreude that Amazon of all companies should be complaining of this. They can go fornicate themselves.

Well, indeed, the same could be said of absolutely anything anyone says, so in that regard what you said is information-free, other than that I can infer that you disagree, that's about it.

The problem is that capitalism can be seen as a system that optimises soley for obtaining money, a token exchangeable for goods, services and power. Once a threshold is reached, it can be gamed to maximise and entrench returns. However, money by itself is an abstract nothing. Humans needs - and societal needs - are utterly unrelated to this obsession and are not selected for. Many things require long term generational thinking that simplistic or purist ideologies do not encompass. It would be equally horrific to create a system that optimised soley for human happiness, for example, as everyone would end up drugged off their faces while the world around them crumbled. And no, under capitalism billionaire altruism does not solve this, their obsessions are nearly always lunatic, monomaniacal or destructive, even in the rare well-intentioned cases.