Yep. Such a classic Claude flavored product. Good ideas wrapped in cotton candy.
I often read as much as 1000 words thinking to myself: “this is smooth”, but then think to myself “Is this my friend Opus 4.8 or now Opus 5?.
In this case the rhetorical neatness is unmistakable especially in the beginnings and endings of paragraphs: “Here is the constructive turn, honestly ranked, with no silver bullets on offer.”
My strategy these days is to scan and look for the tells and click out when I see them. Mine was the same "honestly ranked, with no silver bullets on offer". I suspect in less than a year we won't be able to tell the difference.
I’ve thought this for a while, but why hasn’t it happened yet? At this point, OpenAI and Anthropic and friends could definitely remove the AI “smell” from writing output, or give users a first class way to specify a writing style.
So why haven’t they? My theory is they see this as a sort of fingerprint, useful to not train on later. Or something. Maybe they just don’t care. Certainly feels either intentional or a result of ambivalence.
It’s certainly true today that I probably wouldn’t know an AI written article if the author went out of their way to use one of the many prompts available to tone down the AI-isms.
It’s natural to the fact it’s the same model. Everyone has tics, and when a given model is asked to write millions of texts, they become visible. But! Some portion of the audience and user base can’t see it, so there is no benefit to fixing it. Case in point, yet another hustler felt very clever posting slop, and the likely actual audience (Google’s ranking system) probably does like it.
Because it's doing what is does best: making predictions, which works great when you're not being judged on aesthetics, such as coding or math, but the human thought process is messy or erratic. It just falls apart when you do the next token process to it. If the goal is to "convey information in readable chunks," AI does great at this.
I have started beating the hell out of Claude with stylometric analyses of writers I admire; usually technical writers like Terry Winograd, Leslie Lamport, Rodney Brooks. Then the “no or minimal rhetorical flourishes” rule; no British em-dashes, and minimize the negative phrase thesis-antithesis fun”. It helps.
I think you are right that in a year or two LLMs will be able to do a good impression of many technical styles. But not Nabokov, Kundera, or Kafka for subtlety.
If one manages to channel Edsger W. Dijkstra I will be impressed and rank it high on my leaderboard.
It has been for years. The actual audience for a great deal of text you see is Google’s ranking system. Look up any recipe and ask yourself who reads the ten paragraph story time about grandma’s cookies. It literally isn’t intended to be read.
I believe this also partly sprung up because recipes in isolation don't qualify for copyright. The flavor text gives you grounds to sue if a clone of your recipe site pops up somewhere else.
I agree and that’s why we need and indeed have an ever evolving landscape of benchmarks
> Private, refreshed test sets attack the mechanism itself, and in my view they are the only intervention that does. If the questions have never touched the public Web, they can’t be in the training data; if they rotate, memorizing this year’s set doesn’t help next year.
That’s what we have. A fresh public benchmark is also good, and teams do make efforts to decontaminate training data but there’s likely just no great way around leakage.
Btw, lots more issues in benchmarks than the ones discussed; for instance you can leak answers from the questions themselves or in the case of e.g. multiple choice formats in the actual answers. You just pass the MCQ choices themselves to the model and it may be able to guess way above chance. Coding agent benchmarks sometimes forget to delete .git. They mention e.g. a 6.9% error rate in one of the benchmark items, this seems pretty typical and I would actually be fine shipping that.
Benchmarks are very ugly, but if they didn’t exist we would need to invent them. All of the problems above and more do not explain the progress we see. There are probably 50,000 benchmarks in the literature and new ones get created frequently with varying levels of quality and usefulness.
This reminds me of many years ago when Mozilla/Firefox (I think it was) said that they stopped focusing on mainstream benchmarks because they didn’t really translate to real world browser performance gains.
I view these AI benchmarks the same. No I do not care that GPT got 1200 on FartAGIMaX-4.0-Extreme and Claude got 1350. I care about how much it costs and how correctly it does the tasks that I give it. Unfortunately the only way to know is to use them all myself and measure it myself.
At the end of the day these things are all so damn close in how they behave in whatever harness so it realy just does boil down to whatever is actually cheapest.
This is why Deepseek is great: it’s so much cheaper it doesn’t matter if I burn way more tokens because it’s still orders of magnitude cheaper than the US SotA models. If it doesn’t get it quite right immediately I just do a few more turns and then it’s fine. Barely an inconvenience.
I think Goodhart's law is just a consequence of correlation vs causation.
It is very easy to find a metric that is correlated with what you want. But once you start trying to influence a system, you quickly push it out of the range where the correlation holds.
In order to optimize for something, you need to maximize the actual causative variable. This is much harder.
That is correct. It's just that in most cases in the real world, there is a complex casual network, and many of those variables are not even measurable. So you have to pick a proxy for one or more of the variables and make a metric out of this.
The other problem is that in the real world, we want to make decisions, and the easiest way to make decisions is to have a single metric to judge everything by. With multiple metrics, you get into these debates about subjectivity.
You can get around Goodhart's law if you are able to pick multiple proxy variables and demand that the user optimize them all. And you pick these variables in a way that it's really hard to cheat (i.e. deoptimize the actual intended variable while optimizing the proxy variables). Game designers do this all the time for example, because the system is clean and simple enough to do it.
I think the difference is that Goodhart's law describes how the causal chain _changes_ as a result of management behavior, and in particular the incentives they design for the labor they manage. Incentives are a causal variable for outcomes, and what happens is that people find much easier ways to produce the outcomes you thought you wanted.
Like if you manage a call center and set up KPIs around average call time, reps will start hanging up on customers. Employees could always have done that, and the causal link was always there, there was just no reason to.
IMO the problem is executives want (and perhaps need) their directs to report and track one big number month over month. If you give them five metrics they'll never know if you're making progress or just oscillating between a few local minima. And if each of their ten directs has five metrics, you now have 50 numbers and no idea what time it is[1].
Good benchmarks are costly to build even for mid-large corporations. And once the benchmark is used on models you really can’t tell if the problems would be scrapped for training
Had to stop reading when the article devolved into Claude spam. "defensible in isolation," "honestly ranked," ugh. Please write your own blog post.
Yep. Such a classic Claude flavored product. Good ideas wrapped in cotton candy.
I often read as much as 1000 words thinking to myself: “this is smooth”, but then think to myself “Is this my friend Opus 4.8 or now Opus 5?.
In this case the rhetorical neatness is unmistakable especially in the beginnings and endings of paragraphs: “Here is the constructive turn, honestly ranked, with no silver bullets on offer.”
Yes: and that is actually a smoking gun.
My strategy these days is to scan and look for the tells and click out when I see them. Mine was the same "honestly ranked, with no silver bullets on offer". I suspect in less than a year we won't be able to tell the difference.
I’ve thought this for a while, but why hasn’t it happened yet? At this point, OpenAI and Anthropic and friends could definitely remove the AI “smell” from writing output, or give users a first class way to specify a writing style.
So why haven’t they? My theory is they see this as a sort of fingerprint, useful to not train on later. Or something. Maybe they just don’t care. Certainly feels either intentional or a result of ambivalence.
It’s certainly true today that I probably wouldn’t know an AI written article if the author went out of their way to use one of the many prompts available to tone down the AI-isms.
It’s natural to the fact it’s the same model. Everyone has tics, and when a given model is asked to write millions of texts, they become visible. But! Some portion of the audience and user base can’t see it, so there is no benefit to fixing it. Case in point, yet another hustler felt very clever posting slop, and the likely actual audience (Google’s ranking system) probably does like it.
Because it's doing what is does best: making predictions, which works great when you're not being judged on aesthetics, such as coding or math, but the human thought process is messy or erratic. It just falls apart when you do the next token process to it. If the goal is to "convey information in readable chunks," AI does great at this.
I have started beating the hell out of Claude with stylometric analyses of writers I admire; usually technical writers like Terry Winograd, Leslie Lamport, Rodney Brooks. Then the “no or minimal rhetorical flourishes” rule; no British em-dashes, and minimize the negative phrase thesis-antithesis fun”. It helps.
I think you are right that in a year or two LLMs will be able to do a good impression of many technical styles. But not Nabokov, Kundera, or Kafka for subtlety.
If one manages to channel Edsger W. Dijkstra I will be impressed and rank it high on my leaderboard.
In a year or more, the audience will likely be an agent instead of a person. It'll be interesting to see how that shifts language and article formats.
It has been for years. The actual audience for a great deal of text you see is Google’s ranking system. Look up any recipe and ask yourself who reads the ten paragraph story time about grandma’s cookies. It literally isn’t intended to be read.
I believe this also partly sprung up because recipes in isolation don't qualify for copyright. The flavor text gives you grounds to sue if a clone of your recipe site pops up somewhere else.
It already is. AI bots overflowing visitor logs now with endless IPs
lol
reminds me of a Claude math paper:
"honestly sharp , no hype: cos(pi+pi)+2+2=cos(2pi)+4=1+4=5"
Yeah, the most obvious giveaway is to ask yourself, "is this how a human would actually write?"
I agree and that’s why we need and indeed have an ever evolving landscape of benchmarks
> Private, refreshed test sets attack the mechanism itself, and in my view they are the only intervention that does. If the questions have never touched the public Web, they can’t be in the training data; if they rotate, memorizing this year’s set doesn’t help next year.
That’s what we have. A fresh public benchmark is also good, and teams do make efforts to decontaminate training data but there’s likely just no great way around leakage.
Btw, lots more issues in benchmarks than the ones discussed; for instance you can leak answers from the questions themselves or in the case of e.g. multiple choice formats in the actual answers. You just pass the MCQ choices themselves to the model and it may be able to guess way above chance. Coding agent benchmarks sometimes forget to delete .git. They mention e.g. a 6.9% error rate in one of the benchmark items, this seems pretty typical and I would actually be fine shipping that.
Benchmarks are very ugly, but if they didn’t exist we would need to invent them. All of the problems above and more do not explain the progress we see. There are probably 50,000 benchmarks in the literature and new ones get created frequently with varying levels of quality and usefulness.
This reminds me of many years ago when Mozilla/Firefox (I think it was) said that they stopped focusing on mainstream benchmarks because they didn’t really translate to real world browser performance gains.
I view these AI benchmarks the same. No I do not care that GPT got 1200 on FartAGIMaX-4.0-Extreme and Claude got 1350. I care about how much it costs and how correctly it does the tasks that I give it. Unfortunately the only way to know is to use them all myself and measure it myself.
At the end of the day these things are all so damn close in how they behave in whatever harness so it realy just does boil down to whatever is actually cheapest.
This is why Deepseek is great: it’s so much cheaper it doesn’t matter if I burn way more tokens because it’s still orders of magnitude cheaper than the US SotA models. If it doesn’t get it quite right immediately I just do a few more turns and then it’s fine. Barely an inconvenience.
“When you place a tangible value on trust, trust becomes a commodity to be bought and sold.”
— <https://news.ycombinator.com/item?id=27432186>
Clearly, the solution is to judge society by how many currently-un-gamed benchmarks it has produced.
Not forecasting though. You can't goodhart predicting real-world events
I think Goodhart's law is just a consequence of correlation vs causation.
It is very easy to find a metric that is correlated with what you want. But once you start trying to influence a system, you quickly push it out of the range where the correlation holds.
In order to optimize for something, you need to maximize the actual causative variable. This is much harder.
That is correct. It's just that in most cases in the real world, there is a complex casual network, and many of those variables are not even measurable. So you have to pick a proxy for one or more of the variables and make a metric out of this.
The other problem is that in the real world, we want to make decisions, and the easiest way to make decisions is to have a single metric to judge everything by. With multiple metrics, you get into these debates about subjectivity.
You can get around Goodhart's law if you are able to pick multiple proxy variables and demand that the user optimize them all. And you pick these variables in a way that it's really hard to cheat (i.e. deoptimize the actual intended variable while optimizing the proxy variables). Game designers do this all the time for example, because the system is clean and simple enough to do it.
I think the difference is that Goodhart's law describes how the causal chain _changes_ as a result of management behavior, and in particular the incentives they design for the labor they manage. Incentives are a causal variable for outcomes, and what happens is that people find much easier ways to produce the outcomes you thought you wanted.
Like if you manage a call center and set up KPIs around average call time, reps will start hanging up on customers. Employees could always have done that, and the causal link was always there, there was just no reason to.
IMO the problem is executives want (and perhaps need) their directs to report and track one big number month over month. If you give them five metrics they'll never know if you're making progress or just oscillating between a few local minima. And if each of their ten directs has five metrics, you now have 50 numbers and no idea what time it is[1].
[1]: https://en.wikipedia.org/wiki/Segal%27s_law "A man with two watches never knows what time it is"
obviously, the best benchmark is the one you tell no one about.
Jokes on them, I don't trust benchmarks
Once something becomes a benchmark it is no longer a good benchmark.
No, it's when a benchmark becomes a target. You might have a private benchmark that you tell no one about. Would you not trust it?
Good benchmarks are costly to build even for mid-large corporations. And once the benchmark is used on models you really can’t tell if the problems would be scrapped for training
I was speaking in general, not just about AI.
[flagged]
[flagged]
[dead]