Remembering when I lost people to offshoring in the mid 00s. I made the case, with metrics, that the quality was nowhere near as good and the spec overhead was non-trivial.

Management's response? It doesn't matter - for the price they could get cheap offshore workers to do it over again 5 times and still save money.

It's better to work for companies that understand time/opportunity cost, and so won't penny pinch on something which will make them lots of money.

The "it saves money" mindset is bad for everyone, distorts everything away from valuable work and investments that generate the most revenue.

At least offshoring was seemingly stable. Here we're all waiting for the rug pull.

Why would there be a rug pull? Open models do seem to be giving companies like OpenAI and Anthropic some decent-ish competition, so if they try to increase prices too much then people can just leave them.

For the same reason that most businesses are on AWS/GCloud rather than having a server room in the office.

Self hosting is a complex investment. Competitors need an stupidly expensive amount of hardware to enter the scene if they're going to serve SaaS. Chinese backed models can be regulated out of competition. Openai and Anthropic will soon be semi exclusive sources of training data as development happens on private conversations.

Success is not guaranteed, but it is crystal clear that the goal is replacing us and then pricing the replacement as the largest possible percentage of our salaries possible.

It's the oldest trick in the book. Place a heavily discounted supermarket in town, wait for all local businesses to close, raise prices.

That's not the rug pull.

The rug pull is every software engineer losing their job.

What about people who are running open models on their own GPUs? They are being subsidised too?

That's the point gp was making. If the big labs all increase prices 10x, the cost of running your own open models stays the same

Sure, open models will definitely be an alternative once the frontier models hike prices to match costs. But that will be a minority, and people will start to switch, but people use the frontier models for a reason and these AI labs will extort that, once the easy investor money dries up.

My company pay a copilot subscription to everyone, plus my team have a bedrock subscription that cost us 7k last month. In the meantime, some guys created a new internal team and are testing open source models (for our red team) on our own hardware (we didn't sold our 'old', 3 year old servers RAM/CPU, so the only investment was GPUs). My guess is that within the next 6 months our red team will use open source models exclusively, and my team will probably follow within two years. Especially if the GPU prices go down

> Especially if the GPU prices go down

I rather expect them to go up. I bought a bunch three years ago and a lot more recently. All of them I could sell for 2x what I paid.

How long are those models going to be safe to operate with internet access and tool use without ever being updated? Imagine if "do it for my grandma" is still as effective today because the models haven't changed and that every open-source maintainer upset with AI started polluting their libraries' logs with prompts that have been proven to work with no known reliable fix ever coming in this hypothetical future.

Evidence ?

Who are these investors subsidizing random inference providers on https://openrouter.ai/ ?

Also: who are the investors subsidizing electricity bills for the people who run LLMs on their own GPUs? I'm still waiting for my subsidy cheque over here, so if you can plz send it to me ASAP, thanks.

Question: did you train that LLM yourself on your own GPU?

Question back: how do you imagine them "raising the prices" of the tokens generated on my own GPU? That's the implication we're discussing here. Please explain how that works.

as much as training might be subsidized, I don't buy that inference is, if there was no training required and labs could run their current models forever it'd be a gravy train even at current prices

I think the crux of the issue is that frontier labs do not make any financial sense. They are spending investor dollars in order to sell tokens at pennies, and then a Chinese lab can build a competing model on those pennies, and the frontier labs can't then start selling the tokens at a price where they'd actually even break even because most of the volume will then shift to those cheaper models.

Serious question, what sort of FAFO timelines did you see?

(And I assume that there's a sliver of people who never went this route in the first place, and another sliver who never stopped).

That’s the thing - the FAFO is significantly dampened by the remaining grey beards nudging things back on track over time.

See this over and over on hn and Reddit - CEOs are stupid and don’t understand how much pain offshore/AI causes. No…they’re taking a calculated risk on the organisations resilience and know the people whining about CEOs being stupid will be the ones patching over the cracks

I could see a republican administration banning this entirely, though they would just find cheaper workers in the U.S.

I wonder when this comes from other white class jobs— process engineering, autocad drafter, electronics design.

Republicans were, mostly, the ones responsible for encouraging this in the 80's and 90's with manufacturing.

Now that the unbelievable damage is done to families, towns, and cultures of entire regions of the US, they can swoop in and fix it for white collar workers.

Amazing.

Arguably, it would have happened regardless of who was in power. I do not think Clinton did anything to stop it.

You are right though, the damage is unbelievable and it is heartbreaking.

Agreed, AI is of-shoring at home.

Although obviously I wouldn't like it happening to me, if the offshoring quality was as good as the original -- and I'm sure sometimes it was -- I can't find any valid objection to it. It is, after all, just a different bunch of people being paid to do some work. Why must it be "us" who get that work? Why can't it be "them"?

Being replaced by a machine is different -- but still, not in a way that I can authoritatively say ought not to happen. If you could get your shirt made by a woman, or the same shirt made by a machine that costs 5% of what it would cost to pay the woman, you will pick the machine 99% of the time. Almost anyone would.

Being replaced by other people, or by a machine, is a very sad thing for the person being replaced. But I can't persuade myself that all by itself it is morally wrong in any sense.

The bottleneck in software engineering has always been communication, whether that is engineers communicating amongst themselves or clients communicating what they want, or client's customers communicating to clients who communicate to management who communicate to middle management who communicate to engineers.

Every step introduces misunderstanding and miscommunication, and minds change, and when you go down the wrong path it is hard to alter course or start over while avoiding blame and ill feeling.

Offshoring quality was not as good as the original, because the original was being produced by developers who were close to the clients and steeped in the same assumptions.

With offshoring, all of the weight was borne by the specification - if something was not captured, it did not exist. It doesn't matter how much it was just a basic assumption that something would work that way - if you didn't say precisely what you wanted, you didn't get anything even close.

So someone close to a product could implement something in a week for x money. Management would see that "week" as a fixed unit of work that could be given to someone who would work for less - but that other person would require an order of magnitude more "prompting" in order to produce what you really wanted, because they didn't have the same knowledge and assumptions.

So their calculus was, well, it might actually take 5 weeks to get the same output, but it was still cheaper. But that completely overlooks the local overhead of 5 failed iterations, and the cost to client relations, or the cost to internal staff burning out in frustration at what was once simple becoming multiple needless failures, or the additional cost of documenting the spec to such a degree nothing gets left out. And the latter, you can't do until you've failed enough to know both sides different assumptions and knowledge gaps.

(Edit: also my point wasn't don't do it or that this can never work - it was that in this case management simply didn't care about the demonstrable quality problem because the cost of reiterating was so low they could afford to write off the first couple of failed attempts, and this is exactly the attitude we'll get with AI. If you can get something that's superficially most of the way there in an hour with a prompt all by yourself, there's no cost to just changing your mind, retrying, trying other things, all with no pesky engineers slowing things down)

In the end, money always wins. It’s the way the machine works.

This is true in every online space, whether explicitly or implicitly.

I would be 100% behind a version of Infinite Jest where the footnotes are in a separate volume.