I wonder how many market inefficiencies this creates. People with worse education, people who cheated their way to a job opening compared to a better candidate, etc. Basically counteracting the productivity gains AI was supposed to bring.

Many of those things were already at a “good enough” level since GPT-3.5.

There’s probably a good business usecase there for companies wanting to have smoother communication with offshore teams.

Could that be a game-changer? I wouldn’t discount it, but it does sound like something that has to operate at a very low margin and that doesn’t merit a lot more investment.

> If you can capture that market for next 20 years, it's worth $200 billion.

that's like 5% of NVIDIA's current market cap. sounds like peanuts when you lay it out like that

> Meteorology

It seems like that at first glance. But in reality, GPUs have had extremely slow adoption for real-world operational meteorology applications. Because of the fundamental design and architecture of most NWP systems, it was very difficult to leverage GPUs as compute accelerators; most efforts barely eked out any performance gains once you account for host/device memory transfers. It really wasn't until some groups started to design new weather modeling systems from the ground up that they could architect things in such a way that GPUs made a significant difference.

Obviously AI / ML weather modeling is a different story.

As someone working in a field that has used NLP for quite some time - yeah, I generally agree that those investments are worth their weight in gold... which is unfortunate because before ChatGPT came along they were viewed as niche unprofitable money-sinks. The astronomical investments we've seen lately have been in general models which can be leveraged to outperform some of our older models but had we wanted purely to improve those models there were much more efficient ways to do so.

Hopefully we can retain a lot of this value when the bubble bursts but I just haven't seen any really good success stories of converting these models into businesses. If you try and build as a middleman where you leverage a model to solve someone's problem they can always just go to the model runner and get the same results for cheaper - and the model runners seem (so far - this may change) to be unable to price model usage at a level that actually makes it sustainable.

Those older models running specialized tasks seem to be trucking along just fine for now - but I remain concerned that when the bubble bursts it's going to starve these necessary investments of capital.

You're missing the point. Those kind of narrow AI applications are not the motivation for the trillions of dollars being poured into AI. Of course AI has a variety of applications many disciplines, as it has for decades. The motivation behind the massive investment in AI is as forgetfulness said, reap the benefits from "revolutionizing the workplace"

That’s copium, as the kids say nowadays. The massive planetary investment is a 100% for AI chats. All those other things are taking the crumbs where they can.

Eh, those applications (incl. protein folding) existed for a decade-plus before LLMs came onto the scene, and there was absolutely nothing like the scale of capex that we're seeing right now. It's like literally 100-1000x larger than what GPU hosting providers were spending previously.

> In my opinion it is not a good use of your supervisory time to be babysitting an LLM versus mentoring actual juniors

Right. Because at least juniors are supposed to learn and at some point become senior and stop needing this kind of supervision. Also, interacting with people can be more rewarding (or not, depending on the people)…

They'll waste 15% more of the day.

100 billion a year, with math that’s downright delusional.

So, back of the envelope math, the US GDP is 27.72T USD, 80% (22.18) corresponds to the tertiary sector.

Let’s say that this is a 15% increase over a 10 years period, because YoY a boom like the computer revolution itself looks like 2% increases a year.[1]. This amounts to about 1.5% increases each year.

Let’s just make the huge leap that you can just scale the productivity up of all this just by making typing out reports and emails a faster activity, and summarizing information for which you’re not facing liability if the bot gets it wrong.

Yes, including the nurses, cleaners, truckers, teamsters, all of it.

1.5% of it is a cool 330 billion.

How much of a cut of that productivity increase could AI companies take? 30%? That’s 100 billion in one year there.

So with pie-in-the-sky math, they could break even if their obligations throughout the decade don’t amount to 1 trillion (since those 1.5T in bonds issued this year mature in longer periods)

1. https://www.bls.gov/opub/mlr/2021/article/the-us-productivit...

That is of immense social value - but manual translation services for commercial projects (like application localization) is already dirt cheap to do and automatic casual translation services for consumers would be incredibly difficult to monetize.