Comment by drzaiusx11
16 days ago
I have deep concerns surrounding LLM-based systems in general, which you can see discussed in my other threads and comments. However in this particular article's case, I feel the same fears outlined largely predate mass LLM adoption.
If you substitute "artificial intelligence" with offshored labor ("actually indo-asians" meme moniker) you have some parallels: cheap spaghetti code that "mostly works", just written by farms of humans instead of farms of GPUs. The result is largely the same. The primary difference is that we've now subsidized (through massive, unsustainable private investment) the cost of "offshoring" to basically zero. Obviously that has its own set of problems, but the piper will need to be paid eventually...
Instead of money flowing to lower income countries (by virtue of their cheaper labour), which helped those countries grow, money is now flowing to the already richest economy on earth. That's a big difference.
EDIT: it has been rightfully pointed out that my above comment can easily be read as a racially charged overgeneralization of overseas workers of indo-asian descent. The "cheap spaghetti code" was meant as shorthand for "wildly variable code output with respect to quality and consistency, with no overarching architecture or plan", and was intended to target offshoring _agency_ output along with cheap labor systems that US companies created. These systems attempt to exploit workers at these agencies to avoid paying US salaries and are NOT a reflection on the actual individuals working in the aforementioned cube farms themselves. These workers are already subjected to a number of dehumanizing labor issues entirely outside of their control and I did not intend to further dehumanize. I apologize for my terse, careless wording.
I've worked with, trained and lived alongside workers overseas for months at a time and can say that there's no meaningful difference across racial divides, save for some variation on cultural norms. I would have assumed a more charitable interpretation of my words, but we live in uncharitable times. I'll do better going forward.
Cheers
Interesting how your "structural critique of AI" requires you to characterize an entire workforce of engineers as producing "cheap spaghetti code" from "farms of humans" with a racial meme thrown in for flavor. Code quality tracks with investment and management, not ethnicity. You're not making the sophisticated point you think you're making
I'm specifically speaking of the "race to the bottom" offshore consultancies that exploit cheap labor in foreign, largely asian countries, for export to the US to bypass paying US wages. The preexisting meme I referenced is around corporate lies where their "AI" is largely backed by offshore labor. Think the latest Waymo news, etc. Regardless of those controversies, within the US at least, we've been offshoring technical labor overseas for decades.
I didn't mean to imply that anyone of asian descent is inherently generating "spaghetti code". If that's how it read, I apologize, that was not my intention.
To further clarity, I've dealt with a number of these offshoring agencies (the really inexpensive ones specifically), and their output is very similar to what AI produces today. They have extreme turnover rates, and team assignments change at random so lost context and variable output is common. They do operate cube farms just like US workers, though I'm not sure why that's pertinent to call out though.
I agree, however, that I'm not saying anything sophisticated or complex, merely stating an observation.
Cost of offshoring to ai isn’t zero. Chatgpt and such are businesses. They charge subscriptions. In fact whatever cost you’d pay offshoring to india is probably where chatgpt is hoping to price its subscriptions eventually. Anything less is just leaving money on the table for chatgpt.
I agree it's greater than zero, however, like a good drug dealer you get the first few hits for free (or at cost in the most charitable interpretation of current subscriptions.)
I get the feeling that we're not even close to paying for the _actual_ costs of our frontier GenAI models at current usage levels, with or without subscriptions in the picture. AFAICT we're all using a highly subsidized product, made possible by private capital on the promise of future returns that may or may not materialize.
Outside of a few vertically integrated companies (Google with their custom TPUs, possibly AWS with theirs) LLM companies like OpenAI have to rely on massive data centers via MSFT, Oracle and Nvidia deals to train their frontier models to stay competitive. Theres a lot to pay for when wielding 20 Gigawatts of compute on other folks' machines. For OpenAI we're talking 4+ trillion USD so far with no signs of slowing. That's a hell of a lot of subscriptions to make up for that spend and they have a long climb ahead of them to get there. Maybe their "killer app" will be their new "erotica" models, who knows (porn has lead several tech initiatives in the past.) But I wouldnt bet money on it working out for them.
It's estimated that OpenAI spends 3 USD for every 1 it makes. Obviously that will have to change to make them an actually viable company in the long term. In the end, I see the most likely scenario is we're left with the few large players like Google. They're the ones that have any hope on "winning" the GenAI race, as they're in the best position to not rely on someone else's shovels.
All that said offshoring started out with similar promises to GenAI and some things panned out with offshoring and others didn't. Only time will tell what shakes out of all this mess. I just hope we get a sane readjustment of expectations for GenAI before our next economic collapse (the massive GenAI investment has helped prop up our economy to an extent, at least in the US)
In short, a business exists to turn a profit and OpenAI has yet to do so. Perhaps they eventually will and be the new "offshore" solution going forward as you imply, but just like actually moving your technical talent overseas it comes with a significant amount of tradeoffs to consider (tradeoffs already outlined in parent and other posts on this thread.)
The thing with data center build out is it isn’t just lighting money on fire. You are building out infrastructure that you can then lease out to other users who no longer have to pay to build out their own infrastructure since yours exists and is for lease.