Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Tokenmaxxing is dead, long live tokenmaxxing

12gramsofcarbon.com

Post reply on HN
Print view View original

251–260 of 315 posts

Re: Tokenmaxxing is dead, long live tokenmaxxing

#251

Post by hwillis »

RugnirViking wrote: post #227
jdiff wrote: post #118

Earlier quoted context omitted.

Even modern frontier models benefit so hugely from careful context pruning, maintenance, and rewriting to erase mistakes that it's astonishing to me that there are no tools centered around it. The one tool that used to have such a feature, Zed and its retroactively-named Text Threads, has now stripped itself of it.

this! the back-and-forth chat interface where you can edit only your own messages, and only then to get a new response, is a terrible one, but I think favored by vendors because it helps them fight in vain against prompt injection. Custom harnesses and stuff are nice but incredibly time consuming to set up when all I want to do is like, see what the agent is reading, and editing out some irelevant nonsense side quest…

"some caching gains" is a pretty huge understatement- snipping something out of the middle of the window requires rebuilding the entire context. Thats a shitload of tokens.

Afaik messing with the context also pretty reliably degrades performance still. The model responses reference things that no longer exist to it and it becomes more chaotic.

The real usefulness of parallel or sub-agents is not that they run at the same time, its that they isolate noisy or self-contained context away from the main window.

Re: Tokenmaxxing is dead, long live tokenmaxxing

#253

Post by kaicianflone »

ambicapter wrote: post #131

> That’s no longer true. We’ve entered a different regime, where spending more tokens generally results in better results. We call this “compounding correctness” — the more tokens you spend on getting a task correct, the more likely you’ll get a good outcome. We talked about this a bit at the last in person Agentics meetup: Have we? Is it generally the case that the more tokens you spend, you better results you get?…

Looks like it based on Anthropic’s own multi-agent orchestration research:

https://www.anthropic.com/engineering/multi-agent-research-s...

Their findings suggest multi-agent systems result in better performance attributed mostly to token usage (80% of variance).

Re: Tokenmaxxing is dead, long live tokenmaxxing

#254

Post by deadbabe »

liquicity wrote: post #225
et1337 wrote: post #77

Folks have been saying “things are different now, the agents are now compounding success instead of error” for at least a year now, but I just don’t see it. I was lucky enough to receive a weeklong $50k per head AI training from the people saying these things, and one of their few helpful concrete recommendations was to constantly clear context all the time, to avoid things going off the rails. However, I think findi…

50k per head training and the largest takeaway was to clear context.. that is the "hello world" of using agents, insane.

$50k a head is cheap compared to the productivity gains, probably can push it to $75k

Re: Tokenmaxxing is dead, long live tokenmaxxing

#255

Post by manicennui »

noodletheworld wrote: post #250
manicennui wrote: post #240

Earlier quoted context omitted.

Did you read the article? "Most teams haven’t yet figured out how to build their own Ramp Inspect or Stripe Minions (if that’s you, reach out — we can help!) but basically everyone is at least using cursor in the side bar."

? What is your point? That the OP is obviously finically motived to encourage tokenmaxing? Here’s what they said, $$$ aside: > That’s no longer true. We’ve entered a different regime, where spending more tokens generally results in better results. We call this “compounding correctness” — the more tokens you spend on getting a task correct > Compounding correctness flips the calculus. If more token spend leads to bett…

"I suspect author somehow financially benefits from tokenmaxxing"

Yes, because they are selling AI services. The article is an ad.

Re: Tokenmaxxing is dead, long live tokenmaxxing

#256

Post by aurareturn »

latexr wrote: post #216
aurareturn wrote: post #121

Earlier quoted context omitted.

I'm a +26 on my post so far so it seems like there are a lot of people who agree with me but most replies disagree with me. I suppose this is the nature of online forums - that those who disagree will take the time to reply but those who agree rarely do.

> I suppose this is the nature of online forums - that those who disagree will take the time to reply but those who agree rarely do. Why would those who agree “take the time to reply”? To say what? “This”? “Agreed”? “This guy knows it”? Those comments don’t add anything of value. When you agree, it only makes sense to reply if you have something to say which wasn’t covered by the original argument.

I don't disagree with you.

I'm just pointing out that there are equally, if not more, people who agree with me than what the replies seem to suggest.

Re: Tokenmaxxing is dead, long live tokenmaxxing

#257

Post by aurareturn »

SecretDreams wrote: post #36
witx wrote: post #11

Earlier quoted context omitted.

You're naive, uninformed or turfing if you think companies are still not tokenmaxxing. Also tokenmaxxing was never an intentional and smart strategy employed by companies like you say. It was a mix of fear of missing out, signaling to investors they were in on the hype and recouping investmenets in data centers

CEOs are just as, if not moreso, susceptibility to fomo than everyone else!

I don't disagree. They talk amongst each other, get advice from expensive consultants, but often lack the knowledge that their on the ground workers have. That said, I still think this was done to get employees to adopt AI faster and see what is possible rather than a long-term incentive.

Re: Tokenmaxxing is dead, long live tokenmaxxing

#258

Post by jamiecurle »

cyphar wrote: post #50
SR2Z wrote: post #46

Earlier quoted context omitted.

I really don't understand this take. If you're a carpentry shop that just bought power tools for the first time and you're worried that your employees are sticking with hand tools because that's what they know, then you look for sawdust. The goal isn't to have people work at converting wood into sawdust, the point is that if you wanna see if the tools are working you wanna see proof they're actually being used. I'm s…

Why would a carpentry shop buy hundreds of thousands of dollars of power tools without consulting with their employees to see what they actually need to get their job done more effectively? The logic of buying the tools then forcing the employees to use them "or else" is completely backwards in any sane world. (Of course, we've all had bosses that went to some marketing seminar and come back having been tricked^Wsold…

I see the point but, I'm not really sure the analogy holds up here. If i was in a cabinet shop and had to joint, plane and resaw and cross cut a pile of timber fresh from the saw mill for the next job I'd be very grateful for the jointer, the planer, the bandsaw and the table saw. I'd also be very grateful for the dust extraction.

In in total agreement with you though, forcing tools on employees is very dumb and is terrible leadership. Ask your people what they need to be optimally exceptional and go get them it. Then let them get on with it.

Re: Tokenmaxxing is dead, long live tokenmaxxing

#259

Post by walt_grata »

TeMPOraL wrote: post #202
oblio wrote: post #189

Earlier quoted context omitted.

> the worthiness of the adoption was blindingly obvious at that point Everything is so "blindingly obvious" yet nobody can point to ANY serious peer reviewed studies that prove it. I'm patient, I'll wait.

You don't need peer reviewed studies to tell you water is wet. Peer review is a technique to get evidence from data when SNR is low. It's not "science", it's just a technique. So is "throwing shit at a wall and seeing what sticks". Don't turn techniques into rituals, and science into religion.

Vibes are not evidence, neither is a curated demo. You need actual measured evidence that has an adversarial review to actually prove something without falling to confirmation bias.

Re: Tokenmaxxing is dead, long live tokenmaxxing

#260

Post by aurareturn »

socialcommenter wrote: post #181
aurareturn wrote: post #29

Earlier quoted context omitted.

the big tech companies needing to pump demand for compute. Demand is already so large that OpenAI, Anthropic, Meta, Google could not fill it. Tokenmaxxing for these companies strictly to pump fake demand is just plain wrong. The inference demand for these companies internally must be a drop in a bucket in overall inference demand. This reminds me of the popular opinion on HN for return to office mandates as executive…

Out of $13Bln of 2025 revenue, OpenAI received $867 million from one customer (less charitably, one bankroller), SoftBank. And $300 million from Microsoft[0]. That's more than a drop in the bucket, especially given that they're not the only players complicit in being both an investor and a customer. Also are we sure it's all at arm's length? Barring a full audit, it's not possible to guarantee that there's no round-t…

Are we not all getting timeout issues from Claude Code and Codex frequently due to too much demand?
Post reply on HN