That's a model specific problem, to a large degree.

The "AI assumed something stupid and destructive" problem went away when I started using better planning skills and models which were not anthropic's.

I've not had any of those "wild destructive hallucination reasoning" problems with GLM or Qwen models, even small versions, which I fought with regularly with opus.

Sure, you can do that. You can do that with buggy code that compiles, or python, too.

For AI, you can also have another agent and/or model review and fix it. You can ask it adversarial questions until you get it right. You can ask it to explain the architecture and design (or: you can plan it out ahead of time so that you know the design and then verify it).

It's no different than starting on an existing code base... just faster.

"a bunch of AI bots doing the work of a whole team, is just a pipe dream."

You're not doing things right, then. Or very well. It's pretty low intensity to author/implement 5 substantial features a day, test, fix the bugs, run QA, and do 10-20 odd commits. Give or take, given the complexity of the plans/features and how much they let the LLM fill in the blanks and how thorough they were in scoping the implementation. Maybe 100M-200M tokens, depending on your model/skill/skills/etc. More, if it's a language like python or typescript (noisy, lots of bugs).

That kind of thing used to take a team a day. A good dev might close 2-3 bugs a day if they already have a thorough RCA done by someone else (which, itself, usually takes a 4-8 hours, depending on the issue).

Wait... you mean there were actually people trying to use as many tokens as possible?

It's one thing to try to exhaust a usage quota (which has never been difficult) because it's a subsidized commodity. It's another thing entirely to just burn tokens.

How can a CEO say stupid shit like this? Oh right. CEO brain.

The fact we've all gotten used to workarounds for laptops' abysmal battery life isn't a palliating factor, it's a condemnation.

There is no part of this which even aligns partially with reality.

Let's assume there's a catastrophic blow up with the AI market. Anthropic and OpenAI are gone, or crippled. Resold for parts. NVIDIA closes up shop or goes back to making video game cards after a restructure.

What then?

There's still incredible demand for inference. Every business is scrambling to get more, cheaper, better inference. Almost all of them - whether they're successful with it or just getting in the ropes. That demand isn't just going to disappear, and people/companies able to leverage AI will continue to do so.

People are currently still running - and buying - 2018-era GPUs for inference locally. In fact, most cards cost markedly more (and continue to cost more) than they did new. They're also more capable at running inference than they were even a short while ago: you can do things on a 12GB 3060 today which weren't even possible on the 24GB 3090 6 months ago ago. For the time being, these cards are still increasing in value.

Short of electrical failure, the forefront GPUs will be capable for both inference and training for a long time to come. "Depreciating assets" aside, the actual value will continue to increase short of systems with more capabilities being released.

Which, if the AI bubble pops hard, is likely to slow dramatically.

"The main fall will be the idiots that 'invested' in AI companies."

That's so far from the truth. This has the possibility of completely cratering the IT industry as a whole, because these companies are not going to give up on AI. They'll cut other things first.

The blast radius is huge: OpenAI/Anthropic are heavily invested in Google, Amazon, Oracle, and NVIDIA. Those companies in turn, are invested in myriad AI research efforts (of which OpenAI/Anthropic are in no small part).

Anthropic is likely the lynch pin here. In the last couple months, they've bunged things up great:

* “Ai is dangerous!” Fearmongering to push an agenda about open weight models, and had the entire industry rally against them (including their investors).
* A serious high profile "escape" which impacted HuggingFace
* Constant drama with nerfed models and downgrading Fable users
* Covert quota manipulation on the plans
* Frequent and severe outage/downtime since the beginning of the year
* Opus 5 being what appears to be a fairly crippled model: it can one shot things, but don't ask it to actually plan or follow directions. It won't.
* Heavy handed "safety" alignment which makes their models refuse to do... almost anything, really.
* Kimi K3 coming out and eating their lunch in many regards + new Deepseek, GLM and Qwen models right the horizon (amongst others) - nevermind Llama, Mistral and others which could well be another surprise.

Anthropic's escape hatch of an IPO is all but sealed shut, given the above. Their one hope was regulatory capture of the mote, and now they don't even have the mote, and the entire industry is against them for regulatory capture purposes.

If you need a permit from the government to criticize the government...

The limiting factor for top models is primarily hardware.

You're right that you're not going to get Kimi K3 performance, requiring roughly 600GB of VRAM to run, from a small number of 24GB consumer GPUs. You're also not going to get closed-top-model performance, because you simply don't have any way to get the weights.

If we're talking about K2.7, Deepseek R1 Distilled, Qwen 2.5, Llama 3.2 though - All of which can be run locally -The question really becomes what you expect from them. They're all strong models that know more than we (individually) do about 99% of topics. A "high-school level education" is more than enough for most humans to get by in the world. Now imagine a "high-school level" understanding across every topic. It's not doing calculus, but it can tell you what patterns to use. It's not writing Shakespeare, but it can likely recite most of his works to you from memory. It's not going to come up with a clean-room reimplementation of Windows 11, but it can check your syntax and tell you the function prototypes of every publicly-available library in every programming language without breaking a sweat.

More importantly, to the GP's point - Is a model "smart" enough to understand you're trying to abuse it necessarily better than one that just carries out your orders successfully? Case in point, lately I've found Gemini's new "watcher" blocks a good 90% of my totally legit questions before the question ever makes it to the real model. Umm... No thanks? The best AI is the one that works, not the one that gives the deepest philosophical reason for saying no.

I have to wonder: how much of this is due to how college is designed? You take maybe 1-2 classes a semester related to your field of study, until 3rd or 4th year. History is not a common major, especially after 1-2 year.

I would think that someone bothering to major in a particular field would at least take the time to proofread their own papers.

At 32 out of 35, with only 3 not cheating, that makes me wonder a couple things (besides whether there are fewer history majors now that college is so expensive, and job outcomes in academia are increasingly uncertain): Is this a 100-200 level class? His video seems to suggest so - only part of the midterm was on the Industrial Revolution, and these answers suggest broad 100-level "required" class.

This isn't a story about AI cheating and ruining education. It's about a much deeper, systemic problem with both higher education as an institution, how it's financed, why it's even required, and our cultural valuation (or lack thereof) of the value of a holistic education.

Fair point.

I also don't see them paying hundreds+ if not thousands of dollars for unique books, per book.

You realize they banned the founder of Wikipedia, from Wikipedia, for having badthink ideas - right?

That's one model. There are many.

IOW I need to now run local (or cloud) Chinese models to make sure Claude isn't designing back doors into my software.