[lnsru]

Actual intelligence is useless when decision makers send new weekly AI rules to be better employees. It’s race to the bottom. Race to an endless technical debt. Some companies will implode when codebases stop being manageable. The small minority will thrive. But majority not. I see it used in hardware world. Clever dudes without prior experience with software craft working Python scripts, automate tests, control hardware from rudimentary GUIs. That’s awesome. I see software companies sending internal memo requiring all code to be produced from prompts… It’s like steroids - cleverly used they bring more advantages, though one shouldn’t take double dose with every meal.

[827a]

This too shall pass. Among my software engineering friend group bubble: Every single individual (~12 of us) are actively and seriously tokenmaxing. We have middle-managers who have been given an AI mandate, upper-managers saying "uhh...maybe that brush stroke was too broad" when they look at the bill every month, and zero people in that chain have the authority or even ability to roll it back. This week one of my friends cobbled together an agent that runs in an infinite loop, grabs whatever song they're actively listening to on Spotify, writes it in a file, then instructs the agent to emit tokens for 2-3 minutes on what that song and previous songs that day might mean for that person's mental state, like a little music-based diary. Repeat, run all day, 24/7. Kinda cool. But its just a way to use tokens, because the first thing all these AI labs built was a good coding model, and the second thing they built was a dashboard for admins to track how much their users are using the good coding model.

A TON of companies are getting looted by the AI labs and AI users. Many will not survive. I think Meta will be one of them (a shell of their former selves by 2030). The ones who survive to thrive in the 2030s will be the ones that are relentlessly focused on their customers and products, not the process. If you don't regularly hear both "AI would be awesome for that" and "actually AI probably won't be good for that", your company won't make it. You'll either get lapped by the companies who find the strong use-cases, or you'll get looted by infinite and aimless tokenmaxing. The path through the middle is far more narrow than most companies realize, and some major, major companies are waking up to that harsh reality; for some, too late.

[RJIb8RBYxzAMX9u]

In case this anecdote is not made up, I would implore you or your friend be a bit more subtle at tokenmaxing (ugh). At $JOB, I'm under the same mandate, and it turns out that every prompt is logged and aggregated. When someone else at $JOB asked the team PM who's in charge of the logging, s/he replied that the log is only used to correlate with commits, and nothing else, trust us (wink). I doubt this is unique to my $JOB.

Therefore, sigh burn those tokens, but make sure your prompts are at least superficially defensible, in the unlikely event that you get audited. Use multiple models for the same prompt / task, for instance. It's well know that LLMs are prone hallucinations, so it's only prudent to double / triple cross-check the results with multiple models.

[827a]

The situation is insane at the company my friend works at. There's no central oversight, because, as I understand it, someone in leadership had the idea that they didn't want to prescribe which AI tools their engineers should use. So they just let their engineers expensify any tools they want. Afaik he hasn't yet hit the upper-bound; last month he said he reimbursed a ~$500 anthropic bill. The same company also makes their engineers go to the Apple store to buy laptops, with their own money, then reimburse it. The credit card points must go crazy over there.

But definitely yeah, normally: be careful about these things. In his case when I said "admin dashboard" i moreso meant the general idea of admin oversight; he's said he's been complimented internally about how much he's using AI.

[ungovernableCat]

The same type of executive who will instil a tokenmaxxing mandate will be the same type of executive who decides layoffs are the best way to improve margins when the token bill comes due with no revenue to back it up.

[Imustaskforhelp]

> This too shall pass. Among my software engineering friend group bubble: Every single individual (~12 of us) are actively and seriously tokenmaxing. We have middle-managers who have been given an AI mandate, upper-managers saying "uhh...maybe that brush stroke was too broad" when they look at the bill every month, and zero people in that chain have the authority or even ability to roll it back. This week one of my friends cobbled together an agent that runs in an infinite loop, grabs whatever song they're actively listening to on Spotify, writes it in a file, then instructs the agent to emit tokens for 2-3 minutes on what that song and previous songs that day might mean for that person's mental state, like a little music-based diary. Repeat, run all day, 24/7. Kinda cool. But its just a way to use tokens, because the first thing all these AI labs built was a good coding model, and the second thing they built was a dashboard for admins to track how much their users are using the good coding model.

What the actual quack, are we being really serious? What is even happening at this point and what have I even just read. This might be worse than burning money in a fire pit.

The sad part is the perverse incentives have made it so that these datacenters and gpu's and ram prices and energy costs are going up in price for this...

Why are people tokenmaxxing? Why are these companies literally burning (actually worse) of their money and their investors money? AI psychosis, what exact reason is the cause behind such things?

[ryukoposting]

Is this an ad?

[blowscum]

> Many will not survive.

Thankfully the people responsible have already prepared a golden parachute to land safely to destroy something else.

[eloisius]

It may be useful outside the current tech rat race. One possibility is that a decade of openly user-hostile business decisions will reach their logical conclusion even faster, and those that haven’t fried our brains with CC may be in a position to pick up customers from these behemoths as they disintegrate.

[kuerbel]

I work in infrastructure (backups, networking etc) and no longer in software. I just don't see llms being that useful right now. If I have a problem and ask an LLM the answer is either fabricated or useless, rarely does it know what it's talking about. And yes I know how to describe the problem so that it has a chance to give an useful answer.

Also even with agents, you can't just try and error your way out of some (most) of the problems I encounter without doing harm if the solution fails.

Might be different if used for infrastructure as code or ansible or some such. That I can see.

[jve]

Well Coding agents are being tackled. Infrastructure agents that would read your host event logs, device configuration, ilo, etc, etc - that is probably the missing piece.

Having a chat with chatgpt may give you clues or ideas when you have gone throught your own checklist of what could have went wrong, but can go only as far.

Agent on the other side will decompile .dll to find out issues if needed to go deep enought.

[ratorx]

Providing access to the data is easy. It is just an MCP or equivalent, and coding such CRUD is cheap now.

Applying the actions is unsolved. Unless you YOLO the LLMs, taking stateful actions automatically requires a lot of protective infrastructure, solid testing infra etc.

It’s all just more code, but a “create me a shopping website” LLM is likely not going to be doing the infrastructure level thinking required to handle it for now.

[kuerbel]

Might be but I just can't imagine a customer being fine with a loose cannon agent in their environment. E.g. coding agents are ignoring instructions. Who is to say that Claudes solution to a, say, slow backup isn't deleting the backup?

[foobar10000]

Imagine an agent shadowing all your terminals, providing ideas and asking to run commands that will let it verify the hypotheses it comes up with, while at the same time doing research on vendor docs, etc...

Quite safe, and already a force multiplier - this would be a harness. Maybe have it be able to write to a shadow system with similar (ideally same) hardware to verify it's hypothesis on how the system works, etc...

[locopati]

It is also possible to walk away from tech. To stop chasing the demands of anything for a buck and focus on something real.

[seanclayton]

Some people live paycheck to paycheck in tech. Where do they walk away to that isn't or won't be impacted by AI? Or are you assuming they have the financial support for such a risky switch?

[liotier]

Yes - it is easier than ever thanks to AI !

[ponector]

You don't even need to do anything: layoffs will hit you anyway.

[liotier]

Thanks to AI too - wow, it really is versatile !

[blowscum]

Indeed, here’s a prompt snippet to help you afterwards”.

“Create me a resume for [newjob]. Ensure that it is properly embellished so that my two years of superficial, directionless AI-driven learning seem equivalent to the multi-decade experience and domain expertise the company is actually hiring for”.

[archagon]

If you run your own company — even a tiny one — you don't have to do any of that shit (unless you want to).

[Jtarii]

Companies that use AI well will replace the companies that use AI badly. There is no world in which AI is not used extensively in all employment going forward.

[datsci_est_2015]

I agree, with the caveat that I don’t think any company is using AI well at the moment, specifically because I think our tooling around AI is woefully inadequate and immature.

Right now the AI marketing paradigm is to create rockstar superusers who can (supposedly) do the job of hundreds of individuals at the speed of light! Which bleeds into the design paradigm, which is trash. I’m bullish on AI that can be used more cooperatively and collectively by a company.

[ungreased0675]

Right now LLMs are heavily subsidized. When that ends, the actual cost of the service may exceed its usefulness for many use cases.

[almostdeadguy]

I'm less sure of the fact that ending subsidized token consumption (in isolation) will happen and change this. I think we've seen this play out before with other tech companies where discounting early use ends up entrenching demand and allowing the company to build larger and more efficient infrastructure.

I'm slightly _more_ convinced (still not all that strongly) that the rising cost of memory and chips, data center construction that gets outpaced by computing demand, increasing energy costs, and low switching costs for customers will force the model labs to make changes that increase the barrier to entry (either via higher pricing, more restrictive rate limiting, etc.). or force their customers into longer term commitments.

[foobarian]

> I think we've seen this play out before with other tech companies where discounting early use ends up entrenching demand and allowing the company to build larger and more efficient infrastructure.

We've also seen failures who were convinced "they would make it up in volume." I guess the bet is that infra will get that much more efficient, but it's not clear how much slack there is.

[foobar10000]

A lot - and over the coming 2 years, even more. Utilization rates are under 50% across the board, and special and cheaper chips are coming out all the time for inference. And a truckload of research - TurboQuant, HC (deepseek), etc, etc..

[hellohello2]

Computation halves in price every ~2 years so maybe in the short term but not in the long term

[dawnerd]

How is that possible when the cost of memory and hard drives have gone up 3x+ in the last six months? Maybe cheaper if you're OAI or one of the lucky companies Nvidia is propping up. Everyone else is getting screwed.

[hellohello2]

Sorry but I don't really see how this contradicts what I said in context i.e. both our statements are compatible in the context of what I was replying to

[ungreased0675]

Even so, frontier models get bigger and more complicated, and agentic workflows consume exponentially more tokens than the simple chat windows of two years ago.

[hellohello2]

I agree that the amount people pay for these services is very unlikely to decrease (i.e. Blinn's Law but for tokens). Still, the current level of "intelligence" will eventually become available for a very low price almost surely. Really I simply don't see how you can disagree with the parent's comment "There is no world in which AI is not used extensively in all employment going forward." Honestly I'd like to understand the mindset, is it mainly that you dislike working with these tools/hope they don't get imposed on you or did you actually find them harmful in your work in some way and think they are overvalued?

[contravariant]

You've somehow confused using AI well with using it extensively.

Sometimes using something well involves not using it at all.

[bdangubic]

not using it at all is no longer an option, companies that are not using it at all will die slow/fast death but death nonetheless.

[contravariant]

That's just a baseless assumption. To use AI well you should do the things that allow you to use stuff well. You shouldn't just use it any way you can because you assume that 'not using it at all' is not the best option.

This is literally the same with every single technological development.

[bdangubic]

> This is literally the same with every single technological development.

yup, there are a lot of successful companies today not using the internet :)

[contravariant]

What's the LLM equivalent to email and a basic webpage?

[SecretDreams]

Ironically, companies overusing it will probably die at a similar speed. Maybe faster, even, depending whether cash burn or technical debt catches up to them first.

[bdangubic]

100%

[dawnerd]

That's like saying farmers that don't use pesticides will die out. There's whole industries around doing things not the way big companies say you have to. Human-centric firms will pop up and proposer.

[throwatdem12311]

Maybe. But I used to have unlimited Claude Code usage but now I’m relegated to whatever the subscription happens to give me and when I run out of tokens I need to trad code until my limits reset. My manager saw the bill and nearly fell out of his chair. Small companies just can’t afford the added cost of AI at the real price (and we aren’t even in the real price territory yet).

Hell, even Microsoft is having trouble paying Anthropic’s API rates.

There is a ceiling to how much people are willing to pay for work slop. Just look at the backlash to GitHub Copilot’s token based billing changes.

I don’t want to live in a world where the barrier to entry on entrepreneurship is how much you can pay Anthropic or OpenAI.

[dawnerd]

The copilot token pricing is going to wake a lot of companies up. Even with our smaller company only using around 1/3 of our allotted requests, next month the bill will easily twice as much.

[LtWorf]

If by "using AI well" you mean kill off customer service; maybe customers will want to switch to other companies that are more expensive but have customer service.

[eloisius]

Say the line, Bart!

[paganel]

> internal memo requiring all code to be produced from prompts

That is absolutely insane. Thing is I can honestly believe that it happens, which makes it even more insane.

[pjmlp]

This is basically the next step of all the AI trainings and hacktons that many of us are now required to take part into, with KPI metrics on how each one is using their tokens.

[vips7L]

https://github.com/dtnewman/burn-baby-burn

[baal80spam]

Oh it happens all right.

[Oras]

It’s not like code base written by developers before AI were manageable. The term tech debt was there way before AI coding, and was mainly due to changes made by developers.

I see the point of your argument when this is done by inexperienced developers, as they wouldn’t know what’s happening but for those who knows and guide what has to be done, I don’t see much difference. It’s about understanding the outcome, and evaluating the risk.

[throwatdem12311]

Technical Debt is not a developer skill issue. It’s a management planning, capacity and budget issue. It’s a bet that the cost of servicing the debt will be less than the cost of paying for it outright with cash. I’ve been in the industry for decades and 95% of the dysfunction in an engineering organization is always management.

AI doesn’t really fix that or is really even that suited for it. In many cases it makes it worse.

That’s why you see software quality going down. Developers aren’t told to make better quality software even though AI does really make that easier. Instead they’re told to make more software faster for cheaper.

Cheap, Fast, Quality. Pick two. Business will pick cheap (short term) and fast every single time.

[acdha]

It’s a question of degree: technical debt has a carrying cost trying to balance features against your ability to support the codebase. LLMs change both sides of that equation but I think most companies are going to struggle with maintaining a balance when it’s so easy to push past concerns and get something which seems to work.

[bayindirh]

Tech debt is a debt taken to reduce development time. It's a time debt actually. Patching something that would work until the team has the time to do it correctly.

...and that time never comes in most cases. Because monies are earned in exchanged for that debt and, management cares about monies. They don't see that debt as important, or as debt at all.