Hacker Newsnew | past | comments | ask | show | jobs | submit | danielmarkbruce's commentslogin

You have picked one dimension, incentives, and say the analogy doesn't work because on that dimension it's different. Analogies aren't supposed to be exactly the same.

The point of the analogy stands - most players using paid advertising can't win. They don't realize the entire structure of the market makes it close to impossible for them to win. Google (and a small number of others) have an extremely strong position in what is effectively the distribution channel for a lot of businesses, and a pricing mechanism which enables them to take a giant chunk of economics of the transaction.


Sure, we can make some analogy between ads and gambling, and transitively apply that to ad platforms and casinos. But this article frames ad platforms as exploitative entities engineering advertisers to lose money, when the reality is that the ad platforms provide a marketplace, and via the consequences of supply and demand this marketplace can end up being negative ROI for some advertisers.

Ads have a positive EV while gambling has a negative EV, which completely changes how the entities hosting these products behave. Casinos try to get you to stay on by tricking you into thinking you have a positive EV, while ad platforms get you to stay on by genuinely trying to give you a positive EV. If you aren't getting positive results on your ads, you genuinely have the agency to adjust your ads to try to get positive EV. Casinos are explicitly engineered to give you negative EV regardless of how you behave, and kick you out if you try to change that (e.g. card counting).

If we say that you shouldn't buy the products of an ads platform because the platform tries to get you to buy more, we may as well say you shouldn't buy the products of a grocery store because the store tries to get you to buy more.


>> Ads have a positive EV.

For some, yes. For many, no. That's basically the point of the article. Not only are they negative EV for many, the competitive position of Google and the pricing mechanism make it difficult for any temporary positive EV to stay that way.

All you have done is make the same argument again - the incentives aren't the same. Agreed, they are not. That doesn't change the outcome.


> But this article frames ad platforms as exploitative entities engineering advertisers to lose money, when the reality is that the ad platforms provide a marketplace, and via the consequences of supply and demand this marketplace can end up being negative ROI for some advertisers.

I can agree that prices are not engineered this way consciously. But the article doesn't state they are. It says that ad platform ended at that equilibrium, full stop. Kind of accidentally. Well, the article doesn't say "accidentally" either, the point of the article is: ad platforms extract as much money as they can, and when they do that they become a slot machine, because human nature is fallible to gambling. Gamblers drive prices higher than it is rational. And (the article says) it happens because of the auction model. I'm not sure if it is enough to state that the outcome is a deliberate one or not, but it is beyond the point of the article.

NB. I don't know about ads business and I can't judge if the article is right or wrong.


The article gives absolutely no evidence to support their claim that ads platforms are at an equilibrium. I don't work on the part of our ads business that deals with this and I couldn't reveal confidential data even if I did know it, but there exist ad products that are explicitly designed to give you positive ROI (keyword ROAS if you want to google it). A commonly quoted figure for this type of ad is 2x return, which would strongly imply that the market is nowhere near an equilibrium.

I'm not sure that I understand it (my grasp of economics is pretty slim). You are saying that the current state of affairs is not a stable point and that system is still moving to one. Right?

If so, I can't agree or disagree, because I'm not competent enough for it, but still your words can't be an argument against the article, because the use of "equilibrium" is my addition, and I didn't mean it in a sense you use it (if I understand correctly ofc). What I wanted to say, those prices are not engineered but a result of competing forces shifting them in both direction. This mechanical interpretation I hope explain my use of "equilibrium".


Sorry, slightly misread your comment, so I misused the word equilibrium in response.

The article is arguing that advertisers tend to get neutral or worse ROI, and I think it's saying that's where the equilibrium point is and therefore advertisers will always get neutral or worse ROI on average. But it provides zero evidence that that is truly the equilibrium point.


I'm also unclear here whether this is applicable, I didn't think they meant all ads ever are a slot machine, it sounded like a pretty pointed complaint with one ad system (for which it didn't sound like there was an alternative).

The other products being mentioned, how do they relate to the (slot) machine being discussed in the article?


The article describes a fairly generic ads system; it could plausibly be many different ones. There are many alternatives and I don't think it's particularly difficult to switch to a different ad product within a platform or switch to a new platform entirely.

I don't have much insight into how advertisers behave, but I feel like if an advertiser has a low enough budget that conversions (i.e. real return) are a slot machine should not be targeting conversions and should instead target higher volume actions (e.g. clicks).


> The other products being mentioned, how do they relate to the (slot) machine being discussed in the article?

idk. I have no clue how ads industry is working. I've read the article, I've got the idea what they are talking about, and now I'm reading comments trying to figure out how far/close to the reality the article is.

I'm just frustrated a bit by the fact, that people ignore the gist of the article and talking with voices in their head instead.


The vague situation described in the article is probably close to reality for some advertisers, particularly smaller ones. However it's very anecdotal, and doesn't provide any evidence that what it's saying is broadly true. The general consensus I am aware of (not based on my company's internal data, which I've never analyzed) is that well-designed and strategic ad campaigns consistently get >100% return, especially when accounting for awareness and recognition as a positive benefit.

There is 0 reason to engage with the article, it is completely divorced from reality. Even giving it an ounce of credit is too much. That's why you don't see anyone who has experience working in ads giving it any benefit of the doubt.

ads have -ev for everyone but the advertiser.

100% this. Lots of people want cheaper. The demand for same performance but cheaper is off the charts.

Agree. I was talking to a coworker the other day and remarked that I would be pretty content if we had five years of Claude stagnating in quality/capability but reducing in price/environmental/electrical impact.

It reminds me of when Apple switched to their own silicon in the iPhones and we had year after year of hardware with 1.5-2.5x performance compared to the previous year. We very rapidly got to the point where the processor in my phone was more power than I would ever need (unless apps got more and more bloated and inefficient, which they obviously did).


The performance is really fuzzy too, but the price you do know for certain.

Price is quite fuzzy too though when it's pricing tokens, because a model that's more token efficient but more expensive per token can be cheaper for a given task. And token efficiency is as fuzzy as performance so you're back where you started.

> a model that's more token efficient but more expensive per token can be cheaper for a given task.

True, but I'm skeptical that this is now overly used to ignore cost per token. Your comment inspired me to do a quick investigation. Looks like in my data, 92% of cost per task can be predicted by cost per token https://micjm.com/blog/cost-per-task-vs-token/


Definitely. Cost per task is nice in theory, pretty absurd in practice.

Well, it depends on the data source. Cost per task for benchmarks is absurd. Cost per task for real world usage, where the model identity is randomized (Like Arena or Agent Duel) seems better, though the tasks still are not going to be representative if you're not doing the same stuff as the battle participants.

Or even 90% of the performance but much cheaper. Many business tasks to not require frontier intelligence and were solvable two years ago.

Yeah if you're talking 85-90% of the performance at 10-15% the cost, you're going to win most of the business use cases over. That has been the broad theoretical appeal of the Chinese models, that they might provide that kind of value. Most businesses could not care less about open weights vs not, that does not enter into their discussions very heavily. They care about cost to performance. Fable 5.2 is amazing, and it's as good at eating dollars as it is at producing code.

Risk.

Can i have more explanation please? I'm not in risk management but i would like to understand the situation better. Just risk that they change it under you? Or that they're training on your data, or what kinds of risks are we most worried about?

They can make your model disappear. If you have a system and a set of prompts and evals and everything is working well, it's no fun for that model to just disappear. Some will reply "do a better job building your system so it's not model dependent" and they are sort of right, but that isn't always economically feasible and/or just isn't the reality of how a lot of systems get built.

Models hosted as a service can be subtly tweaked without any announcement, or suddenly be 'upgraded' to a new one with a ticking clock on how long the old one sticks around. In both cases, the flows people establish around them are disrupted or broken.

I've lost track of the anguished cries of those unhappy when either of these has happened.


That's a problem with decision models, for example, but coding models have no such risk. I switch between multiple coding models from multiple providers each day.

Yes, we do need more. I would like significantly lower price for same performance. I'd like higher throughput.

There is nothing wrong with this on it's face.

This is an SEC filing.

They aren't obliged to report it, and it's actively misleading. Companies stop reporting various numbers for various reasons.

> Companies stop reporting various numbers for various reasons

You're getting there.


yep.

China stopped reporting youth unemployment numbers. When stats are not reported, but a reasonable reason not given, then there must be some motive behind it. It might be benign, but i would always err on the side of nefarious.


Which would make you correct more often than not. For the simple reason of how powerful the status quo is. I.e. it generally requires a forceful decision from the top for any change to happen in a large bureaucratic system, which means that there is some strong motivation behind removing it.

Ask anyone who reads a lot of SEC filings, it's common and usually innocuous. Not always, but usually.

It’s conspicuous timing to stop reporting when your CEO gets caught in bed with the Chinese government. Could be nothing, but suspicion is warranted.

The first line of the article.

>Nvidia stopped disclosing the scale of its business in Singapore last year


So.. approximately when someone at the company realized something fishy was going on with Jensen, but before they got caught? Obviously this is conjecture but your reply isn’t the slam dunk you seem to think it is.

I think it's a stretch to say 'Gets caught' was conspicuously timed with this action.

I quoted a line from an article that seemed relevant. I would appreciate you not projecting your opinion of what I thought about it onto me.


> it's a stretch to say 'Gets caught' was conspicuously timed with this action

It could be timed with when Jensen "got caught" internally, when his missteps became too conspicuous to ignore by other high level internals.

Didn't look like GP projected that opinion on you. Just a plausible observation that someone inside Nvidia tried to mitigate the risk put by Jensen's actions by implementing this change. If they were lucky this could have stayed hidden forever. But it came out in the public and now we can all see some conspicuous connections.


It could be that Hugh Heffner ordered the assassination of JFK. Something that is possible but lacks evidence is not reasonable grounds for suspicion.

>your reply isn’t the slam dunk you seem to think it is.

This is what I was referring to in the latter part of my comment.


More likely: the CEO drama has meant any news related to NVIDIA is being highlighted

Of course it's the same. It's just on a larger scale. Business has been done on credit since the beginning of time.

Selling on credit is fine as long as the underlying demand is real. And, in this case, it appears to be real.


And who tells you that a youtube video has taught you well?

My common sense tells me that a college professor is likely to have taught me better than an AI

Common sense around a lot of topics can be misleading. Especially when human bias is taken into account. It's not at all clear to me that a properly trained LLM would not out perform a professor. Even we are talking about known great teaching professors, not average teachers, the LLM has several advantages. It can tailor curriculum to each individual. It can learn and adapt curriculum on the fly which a prof is unlikely to do. It can adjust testing and grading so that the student is maximally engaged. It never misses a day or has a bad day. It can answer questions thoroughly and with as much detail as the student wishes, taking as much time as the student desires. What are the disadvantages?

> It can adjust testing and grading so that the student is maximally engaged

This part is not reality.

In order to do this, the AI would need a way of measuring student engagement.

As soon as you have a measure like that, it is going to be gamed.

Somebody needs to set the standards. You can’t rely on an LLM for that because the LLM doesn’t make up a goal.

Even in a very basic non-LLM example like learning vocabulary, the acceptable rate of correct answers (a day or a week after studying) needs to be set externally.


That it will make up stuff up out of thin air???

And what does properly trained mean, do we know how to "properly train" something with your characteristics?


My Stats I and II professor (don’t ask me how) spend the first 5 weeks explaining his take on best-practice US train travel, and sharing his experiences with us.

5 weeks!

They’re often just humans.


General common sense tells everyone that not all teachers teach well, and that not all students can learn well from all teachers who may teach well because style preferences may collide. One may easily calibrate an AI to teach them in their preferred style; not so much a college professor.

It's more that you can continue prodding chatgpt/whatever. If you are a naturally skeptical person and try to poke holes in ideas by nature, you don't really get to do to your hearts content with a professor.

But, some people aren't skeptical. A professor says X and they just accept it and feel like they learned from a trusted source or authority or whatever.


That's probably wrong, if you know how to use AI well.

Sure, just tell it "don't be wrong"

You appear to have a naive level of trust in the competence of humans.

You appear to have an extremely naive level of trust in LLMs.

No, one can just keep prodding, in a way that couldn't be done with a professor. It requires no trust.

So, prodding an LLM makes it not lie?

No....just like that doesn't work on humans. But with prodding you can work out when it's wrong or at least when you should be skeptical. People usually call this "critical thinking".

If you go around assuming authorities on subjects are just right, you will be in for a difficult time. If you endlessly prod and show skepticism towards human experts/authorities, you'll piss a lot of people off. Most of the time they won't even engage.

If you are genuinely interested in learning things and have critical thinking skills, sota LLMs are as good as it gets.


No, just no.

You seem the kind of guy who distrusts professors and experts and used to trust YouTube idiots, and now turned to LLMs.

Professors will occasionally teach something wrong, but you'll also be studying from books, and if you really want you're free to eventually also use LLMs.

Beginning from them, though, is not too smart, and relying exclusively on them is just plain stupid.

LLMs are going to make stuff out of thin air, and can be more and more confident of it the more you "prod" them.


lol, these videos are on Youtube...

You seem to be struggling to read the thread. Nobody (except you) said trust, just use. Critical thinking is always required. And no one said rely exclusively on LLMs. And "prod" doesn't mean what you seem to think it means. You prod to find holes.


> lol, these videos are on Youtube

not from idiots

> And no one said rely exclusively on LLMs

It's what you seemed to be arguing; or, at least to use them as the primary education method

> And "prod" doesn't mean what you seem to think it means

Do explain


Make no mistakes.

You can ask ChatGPT for a detailed listing of all the areas within a discipline and it will give a totally decent answer, better than you'll likely find anywhere else. You can ask it to list disciplines that might be useful, that are taught at university, or whatever.

You can even just ask it "help me know what areas/subjects/topics are useful that I wouldn't even know exist"


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: