And if you know how few of the real problems QA normally detects before a product is being shipped to the customer you know the shit show that those products will cause.

Hand me the popcorn, that show has a lot of potential.

Well, they're right.

"If we stop testing right now, we'd have very few cases, if any," Trump asserted, "Without testing, or weak testing, we would be showing almost no cases."

The more things change the more they stay the same.

If shit rolls downhill QA is at the bottom and I'm saying this as someone who comes from Ops. I deal with the shit when it hits production.

Truth of the matter is in any org it goes like this

Devs - literally walk on water. Give us new and shiny things

Ops - support the half baked shit that devs promote to prod on lies that "it's production ready we will fix in future sprints"

QA - magical role that is never filled because "we can just get devs to review their own code"

Some of the devs in my team do this occasionally and it ALWAYS eventually causes issues

If QA hadn’t been replaced with a chatbot, that is.

I've never known QA evaluations to be design-based, only results-based. They would test each of the cases and reject the code if it failed any of them, but they wouldn't check for or reject the code just because it was handling them in stupid, non-generalized ways.

Checking for bad design is more of a team lead code reviewer thing (if it happens at all).

I'd say it's never been as good as Musk and Altman especially make it sound.

If you drill into the nuance of their proof points that their companies bother to share, it's generally a bit more clear that things aren't exactly 'asterisk free' even then. Most recently I read how one of them proudly proclaimed they were 100% manual code free, but for the task described they still had 200 developers, which is significantly more developers than I have traditionally seen. To get things vaguely working, you need to be relentless in providing a super comprehensive set of requirements and criteria that are impeccable at fully capturing what is needed and could reasonably go wrong. Also a fanatical amount of detail to audit how the solution was reached and correct missteps, like when it decides to delete a failing test case instead of fixing the correlated problem. The 'purist' approach is, in my opinion, worse for the sake of just signaling how 'AI-friendly' you are.

I can at least buy that from a pragmatic perspective, the codegen can accelerate things if used under a more balanced approach, being able to be in the code and grasp in detail where it might have gone wrong and correct it before it gets lost in a mess of garbage. It will do things that a human might skip that they shouldn't, like hooking various unlikely errors and providing a nicer treatment. It is often useful to get a general idea of a good choice of library to do something and how it fits in your code, though you still have to beware that it may invent details that would be nice and convenient, but are wrong.

The problem is that narrative is totally not what they need to justify the spend. They needed to take randos and prompt up things without full understanding. They needed AI codegen to be really valued on the same level as a full headcount. This trickles down to little folk like the team lead in my situation, where he thinks this is his only ticket to the 'big leagues' if he manages to drive a success story for 'AI native'. He doesn't understand any of it and his team doesn't understand any of it and so of course the AI can't bring what actually understanding the scenario would.

Even more rough for that guy: he imagines a world of slopping up code and then selling the result, without ever having unique skill or insight. Assuming that can work, then why would any company buy your slop instead of just slopping it up themselves?

Is AI imploding within itself through poisoning and no new content?

Model collapse doesn't retroactively affect models that have already been trained, so this option seems like a pretty nonrigorous belief

Are they putting in their own guardrails to take away the really good innovations because of greed?

This would be difficult and likely counterproductive, but they are known to adjust the amount of compute they allow models to use, so they could be throttling that

AI was never as good as they claimed, even the super advanced models?

Ding ding ding, we have a winner! AI has improved, but those who are selling it will always and forever oversell it, like they would with any product.

Well, so far, it's the "I've bet my entire career on AI" leader who has not yet gotten traction. He had a large benefit of the doubt but I feel like leadership is getting more and more skeptical about him and his band of merry sloppers.

Moving QA to the customer base is straight Microsoft.

As someone who went through the dot.com bust at a very unfortunate time in my career, I'm always looking for the escape hatch.

But I feel like the prevaling opinion in the leadership guy is increasingly dubious of this guy, so I'm hopeful sanity prevails.

That's not the case with anthropic, though. They are simply burning money while not making a profit.

That's not how shares work at all!

You can't turn them into the company at market value, you have to sell it to other people to cash out. The only time a company buys its shares is when they do a buyback, but that wouldn't be a debt on the balance sheet.

The shareholders are just the company owners.

A lot of people are getting addicted, though. I have the sinking feeling they're trying to keep this going for as long as they can so there's enough addicts who will then sink their savings to get their fix once the bubble pops.

Why not a local model on owned hardware?

Only for the hyper scale centralized compute llms. Real machine learning tools predate and will persist past the bubble pop, but data center cloud based llms will go the way of the dodo one way or the other. Which is a disaster for the big companies that want to sell you compute as a utility, but really totally fine for most everyone else.

Idk, the cat is kind of out of the bag here with local open-weights models. Even if all the providers shut down, someone who really wanted an AI could use one of those with not too much effort. A good model would require some impressive hardware, but a bad one is "good enough" for some people.

They are profitable on inference (at least anthropic). Inference is cheap, it’s not susbsidized. It’s training that’s expensive, but if the bubble really does pop then current open source models (which are pretty much sold at inference rate + small margin) will actually go down in price since the cost of hardware will have gone down (assuming less demand for GPUs). Considering that the current best open model is roughly one month behind, we won’t revert to pre-ai anywhere. It’s just too cheap for that.

Still, it’s lovely seeing people spreading misinformation based on something they saw on tiktok years ago. Keep up with the times or don’t spread outdated info.

Oh musk did that one already. That’s why he has helpers ;)

Robots, flying cars, quantumAI, Epstein Files...