tl;dr The best ~AI's~ LLM's slop asymptote is 10 hours.
Restated, if you let the best LLM chomp on a task for 10 hours, the output becomes slop.
* These tasks are of the type that you spend 1% of your SWE career working on.
* Each task is primed with an essay length prompt.
* You must play needle in the haystack for bugs in 10 hours worth of AI generated slop.
My experience trying AI coding at work and my observations of AI evangelists makes me believe AI coding is exclusively the purview of people who willing to handhold an AI at half pace to achieve the same result while working on software which amounts to greenfield/toy problems.
The danger of LLMs to thought work is enormously overstated and intentionally overhyped. AI : StackOverflow :: StackOverflow : graybeard in basement
It would be cool if AI kills all thought work, but what will actually happen is a undersupply of SWEs and a second golden age of SWE salaries in like 15y.
Does anyone know who Dwarkesh’s patron is that boosted him in podcast world? He isn’t otherwise highly distinguished and admitted does his show prep with AI which sometimes shows in his questions. I feel like there are a very large number of tech podcasts, but there’s some marketing effect around this guy that I just don’t understand.
> Nobody at this point disagrees we’re going to achieve AGI this century.
Nobody. Nobody disagrees, there is zero disagreement, there is no war in Ba Sing Se.
> 100% of today’s SWE tasks are done by the models.
Thank God, maybe I can go lie in the sun then instead of having to solve everyone's problems with ancient tech that I wonder why humanity is even still using.
Oh, no? I'm still untying corporate Gordian knots?
> There is no reason why a developer at a large enterprise should not be adopting Claude Code as quickly as an individual developer or developer at a startup.
Oh good, hopefully it'll model itself after an exponential rise in any sort of animal populations and collapse on itself because it can no longer be sustained! Isn't that how things go in exponential systems with resource constraints? We can only hope that will be the best outcome. That would be wonderful.
I find myself coding a lot with Claude Code.. but then it's very hard to quantify the productivity boost.
The first 80% seem magical, the last ones are painful. I have to basically get the mental model of the codebase in my head no matter what.
We're only not letting go because it's not quite there yet. Once AI is there, someone will let go, and to keep up with everyone else, you'll let go too.
Wait a bit longer and the next thing that's let go after you "let go" is you.
Regulation will not stop this. It's time to build and deploy weapons if you want your species to survive. See earlier discussion here: https://news.ycombinator.com/item?id=46964545
(a) Top labs quietly signing deals for military deployment of frontier models in unmanned strike weapons?
(b) Top labs agreeing to license LLMs for social engineering/propaganda ops?
(c) Models that vastly exceed human intelligence and have capacity to pursue own agenda (i.e. runaway intelligence)?
(d) Something else?
It looks like dangers of AGI are overblown (perhaps partially due to grant funding and ability to get political traction/investment/competitive advantage), while (a) and (b) are severely underdiscussed. Would love to get other perspectives.
LLMs alone aren't the way to AGI. Perhaps something involving a merge of diffusion or other models that are based on more sensory elements, like images, time, and motion, but LLMs alone aren't going to get us there.
The end of the exponential means the start of other models.
Is "the end of the exponential" an established expression? There's no singularity in an exponential so the expression doesn't make sense to me. To me, it sounds like "the end of the exponential part", meaning it's a sigmoid, but that's obviously not what he means.
I’m guessing that Amodei meant it as a humorous inside joke.
It’s also shorthand for “the end of massive R&D capex” and “the transition to market capture”. The final stage, what McKinsey types call “harvesting”, is probably not on Amodei’s radar. Based on what I’ve seen of his public personality, he would see it as too philistine and will hand it off to another custodial exec.
The concept of the "end of the exponential" sounds like a tech version of Fukuyama's much mocked "End of History". Amodei seems to think we’ll solve all the "useful" problems and then hit a ceiling of utility.
But if you’ve read David Deutsch’s The Beginning of Infinity, Amodei’s view looks like a mistake. Knowledge creation is unbounded. Solving diseases/coding shouldn't result in a plateau, but rather unlock totally new, "better" problems we can't even conceive of yet.
Not infinity. Only the path to make steady returns in a few short years. Take disease research. Pharmaceutical companies are not interested in curing disease. They would like to treat disease. That means recurring revenue. They would like to focus on the diseases with the most patients to maximize the market for their product. This is why a dozen plus pharma companies are pursuing glp1 while cutting internal r and d jobs and offshoring everything not specifically bolted to this country by the FDA to India.
This is what depressed me as an early career scientist. Money to do the work to advance our species is not being distributed. Only money to generate more money for a sliver of the ownership class is distributed.
The incentives are broken. We aren’t getting Star Trek in our future. We are getting CHOAM.
For instance, once you develop atomically precise manufacturing ala Drexler and have a complete model of biology, etc., drive solar panel efficiency to very near the upper theoretical bound for infinitely many junction cells for a raw panel of ~68%, then there isn't really anywhere to go that matters for humans. Material production would be solved, anything you could desire would be manufacturable in minutes to hours, a km^2 of solar panels could power 10-20k people's post-scarcity lives.
You eventually reach the upper bounds on compute efficiency and human upload model efficiency -- unknown but given estimates on upper bound for like rod logic (~1e-34Js/op), reasonably bounds on op speed (100MHz), and low estimates for functional uploading (1e16 flops), you get something in the zone of 0.1nW/upload, or several trillion individuals on 1m^2 of solar panel in space. When you put a simulated Banks Orbital around every star in the Milky Way in a grand sim running on a system of solar panels in space where the entire simulated galaxy has a 15ms ping to any other point in the simulated galaxy, what exactly is this infinite stream of learning? You've pushed technology to the the limits of physical law subject to the constraint of being made of atoms.
Are you envisioning that we'd eventually be doing computation using the entirety of a neutron star or (if they can exist) a quark star? Even then, you eventually hit a wall where physics constrains you from making significant further gains.
There is an ultimate end to the s-curve of technology.
I have said that Amodei is by far worse than Sam Altman. Altman wants money but this guy wants the money AND to be your dad by censoring the shit out of the model and wagging his finger at you what you can say or what you cannot. And lobbying for legislation to block competition. Also the constant "muh china" whining while these guys stole all the books in the world.
Every time I read something from Dario, it seems like he is grifting normies and other midwits with his "OHHH MY GOD CLAUDE WAS KILLING TO KILL SOMEONE! MY GOD IT WANTS TO BREAK OUT!" Then they have all their Claude constitution bullshit and other nonsense to fool idiots. Yeah bro the model with static weights is truly going to take over.
He knows what he is doing, it's all marketing and they have put shit ton of money into it if you have been following the media for the last few months.
Btw, it wasn't many months ago that this guy was hawking doubling of human life span at a group of some boomer investors. Oh yeah I wonder why he decided to bring it up there? Maybe because the audience is old and desperate and that scammers play on this weaknesses.
Truly of one of the more obnoxious people in the AI space and frankly by extension Anthropic is scammy too. I rather pay Altman than give these guys a penny and that says a lot.
> 100% of today’s SWE tasks are done by the models.
I do think he was overstating the current state of the models by a bit, but this is taken out of context. He is not saying this is where the models are at today.
He gives a spectrum [18:30] of the models taking over the SWE jobs:
- Model writes 90% of code (today)
- Model writes 100% of code
- Model does 90% of today's SWE tasks (end-to-end)
- Model does 100% of today's SWE tasks
- The SWE job creates new tasks that didn't exist before
- Model does the new SWE tasks as well (90% reduction in demand for SWE)
One of my friends and I started building a PaaS for a niche tech stack, believing that we could use Claude for all sorts of code generation activities. We thought, if Anthropic and OpenAI are claiming that most of the code is written by LLMs in new product launches, we could start using it too.
Unsurprisingly, we were able to build a demo platform within a few days. But when we started building the actual platform, we realized that the code generated by Claude is hard to extend, and a lot of replanning and reworking needs to be done every time you try to add a major feature.
This brought our confidence level down. We still want to believe that Claude will help in generating code. But I no longer believe that Claude will be able to write complex software on its own.
Now we are treating Claude as a junior person on the team and give it well-defined, specific tasks to complete.
Is it worth starting from scratch and adding a "make it easily extensible" to the initial prompts? Maybe with the recently released models it'll do an even better job. Just keep rebuilding from scratch every time a new model version is released.
> But I no longer believe that Claude will be able to write complex software on its own.
"on its own" is doing a lot of work here. Dario went into the differences in this very podcast: "Most code is written by agents" is not the same as "most code is written without or independent of human input".
I suspect that is how different outcomes can be explained (even without having to assume that Anthropic/OpenAI engineers are outright lying.)
You don't believe the current version of Claude Code will be able to write complex software on its own.
On the one hand, there is a lot of hype, an incredible amount, actually, but on the other, we have been observing in real time a technological miracle that gets better by the week.
We have no idea what, five years from now, the coding agent will be able to develop.
No matter how fast and accurately your AI apps can spit out code (or PowerPoints, or excel spreadsheets, or business plans, etc) you will still need humans to understand how stuff works. If it’s truly business critical software, you can’t get around the fact that humans need to deeply understand how and why it works, in case something goes wrong and they need to explain to the CEO what happened.
Even in a world where the software is 100% written by AI in 1 millisecond by a country of geniuses in a data center, humans still need to have their hands firmly on the wheel if they won’t want to risk their businesses well being. That means taking the time to understand what the AI put together. That will be the bottleneck regardless of how fast and smart AI is. Because unless the CEO wants to be held accountable for what the AI builds and deploys, humans will need to be there to take the responsibility for its output.
Referring to a curve with a derivative everywhere equal to its value as something that has an end gives the game away: pure fanciful nominalization with no grounding in any kind of concrete modelling of any constraints.
IMHO this is really silly: we already know that IQ is useful as a metric in the 0 to about 130 range. For any value above the delta fails to provide predictive power on real-world metrics. Just this simple fact makes the verbiage here moot. Also let's consider the wattage involved...
It's difficult for me to express this view, which I hold genuinely, without reading as lacking in humanity. However, I think it would be disastrous for humanity as a whole if we eliminate disease completely. To fight against it and to make progress in that fight is of course deeply human. And we are all affected emotionally and personally by disease of all forms. But if we win the fight against disease, I am almost sure that the human race will just end as a (long term) consequence.
48 comments
[ 5.9 ms ] story [ 71.2 ms ] threadhttps://www.julian.ac/blog/2025/09/27/failing-to-understand-...
Restated, if you let the best LLM chomp on a task for 10 hours, the output becomes slop.
* These tasks are of the type that you spend 1% of your SWE career working on.
* Each task is primed with an essay length prompt.
* You must play needle in the haystack for bugs in 10 hours worth of AI generated slop.
My experience trying AI coding at work and my observations of AI evangelists makes me believe AI coding is exclusively the purview of people who willing to handhold an AI at half pace to achieve the same result while working on software which amounts to greenfield/toy problems.
The danger of LLMs to thought work is enormously overstated and intentionally overhyped. AI : StackOverflow :: StackOverflow : graybeard in basement
It would be cool if AI kills all thought work, but what will actually happen is a undersupply of SWEs and a second golden age of SWE salaries in like 15y.
https://github.com/METR/public-tasks/tree/main
Citation needed please.
Nobody. Nobody disagrees, there is zero disagreement, there is no war in Ba Sing Se.
> 100% of today’s SWE tasks are done by the models.
Thank God, maybe I can go lie in the sun then instead of having to solve everyone's problems with ancient tech that I wonder why humanity is even still using.
Oh, no? I'm still untying corporate Gordian knots?
> There is no reason why a developer at a large enterprise should not be adopting Claude Code as quickly as an individual developer or developer at a startup.
My company tried this, then quickly stopped: $$$
Oh good, hopefully it'll model itself after an exponential rise in any sort of animal populations and collapse on itself because it can no longer be sustained! Isn't that how things go in exponential systems with resource constraints? We can only hope that will be the best outcome. That would be wonderful.
This is a key insight, I'm unable to get around this.
It's the thing I require to have before I let go, and I want to make sure it's easy to grasp again aka clear in the docs.
Basically - the sys architecture, the mental model for key things, even the project structure, you have to have a pretty good feel for.
Wait a bit longer and the next thing that's let go after you "let go" is you.
We have very good abstractions for algorithms and especially functions.
Functions are an extremely good 'contract'.
With good functional programming you can totally 'let go' of the internals.
But classes and modules are not that - we don't have the abstractions.
We can't just let go of the AI designating all sorts of mechanics - unless - they are using really common patterns.
So without ways to really describe the patterns and without common patterns to rest on ... the AI can't really get there.
Quoting the Anthropic safety guy who just exited, making a bizarre and financially detrimental move: "the world is in peril" (https://www.forbes.com/sites/conormurray/2026/02/09/anthropi...)
There are people in the AI industry who are urgently warning you. Myself and my colleagues, for example: https://www.theregister.com/2026/01/11/industry_insiders_see...
Regulation will not stop this. It's time to build and deploy weapons if you want your species to survive. See earlier discussion here: https://news.ycombinator.com/item?id=46964545
(a) Top labs quietly signing deals for military deployment of frontier models in unmanned strike weapons?
(b) Top labs agreeing to license LLMs for social engineering/propaganda ops?
(c) Models that vastly exceed human intelligence and have capacity to pursue own agenda (i.e. runaway intelligence)?
(d) Something else?
It looks like dangers of AGI are overblown (perhaps partially due to grant funding and ability to get political traction/investment/competitive advantage), while (a) and (b) are severely underdiscussed. Would love to get other perspectives.
Yet news and opinions from that world somehow seep through into my reality...
The end of the exponential means the start of other models.
It’s also shorthand for “the end of massive R&D capex” and “the transition to market capture”. The final stage, what McKinsey types call “harvesting”, is probably not on Amodei’s radar. Based on what I’ve seen of his public personality, he would see it as too philistine and will hand it off to another custodial exec.
But if you’ve read David Deutsch’s The Beginning of Infinity, Amodei’s view looks like a mistake. Knowledge creation is unbounded. Solving diseases/coding shouldn't result in a plateau, but rather unlock totally new, "better" problems we can't even conceive of yet.
It's the begining of Inifinity, no end in sight!
This is what depressed me as an early career scientist. Money to do the work to advance our species is not being distributed. Only money to generate more money for a sliver of the ownership class is distributed.
The incentives are broken. We aren’t getting Star Trek in our future. We are getting CHOAM.
For instance, once you develop atomically precise manufacturing ala Drexler and have a complete model of biology, etc., drive solar panel efficiency to very near the upper theoretical bound for infinitely many junction cells for a raw panel of ~68%, then there isn't really anywhere to go that matters for humans. Material production would be solved, anything you could desire would be manufacturable in minutes to hours, a km^2 of solar panels could power 10-20k people's post-scarcity lives.
You eventually reach the upper bounds on compute efficiency and human upload model efficiency -- unknown but given estimates on upper bound for like rod logic (~1e-34Js/op), reasonably bounds on op speed (100MHz), and low estimates for functional uploading (1e16 flops), you get something in the zone of 0.1nW/upload, or several trillion individuals on 1m^2 of solar panel in space. When you put a simulated Banks Orbital around every star in the Milky Way in a grand sim running on a system of solar panels in space where the entire simulated galaxy has a 15ms ping to any other point in the simulated galaxy, what exactly is this infinite stream of learning? You've pushed technology to the the limits of physical law subject to the constraint of being made of atoms.
Are you envisioning that we'd eventually be doing computation using the entirety of a neutron star or (if they can exist) a quark star? Even then, you eventually hit a wall where physics constrains you from making significant further gains.
There is an ultimate end to the s-curve of technology.
Every time I read something from Dario, it seems like he is grifting normies and other midwits with his "OHHH MY GOD CLAUDE WAS KILLING TO KILL SOMEONE! MY GOD IT WANTS TO BREAK OUT!" Then they have all their Claude constitution bullshit and other nonsense to fool idiots. Yeah bro the model with static weights is truly going to take over.
He knows what he is doing, it's all marketing and they have put shit ton of money into it if you have been following the media for the last few months.
Btw, it wasn't many months ago that this guy was hawking doubling of human life span at a group of some boomer investors. Oh yeah I wonder why he decided to bring it up there? Maybe because the audience is old and desperate and that scammers play on this weaknesses.
Truly of one of the more obnoxious people in the AI space and frankly by extension Anthropic is scammy too. I rather pay Altman than give these guys a penny and that says a lot.
> 100% of today’s SWE tasks are done by the models.
Maybe that’s why the software is so shitty nowadays.
I do think he was overstating the current state of the models by a bit, but this is taken out of context. He is not saying this is where the models are at today.
He gives a spectrum [18:30] of the models taking over the SWE jobs:
- Model writes 90% of code (today)
- Model writes 100% of code
- Model does 90% of today's SWE tasks (end-to-end)
- Model does 100% of today's SWE tasks
- The SWE job creates new tasks that didn't exist before
- Model does the new SWE tasks as well (90% reduction in demand for SWE)
Unsurprisingly, we were able to build a demo platform within a few days. But when we started building the actual platform, we realized that the code generated by Claude is hard to extend, and a lot of replanning and reworking needs to be done every time you try to add a major feature.
This brought our confidence level down. We still want to believe that Claude will help in generating code. But I no longer believe that Claude will be able to write complex software on its own.
Now we are treating Claude as a junior person on the team and give it well-defined, specific tasks to complete.
"on its own" is doing a lot of work here. Dario went into the differences in this very podcast: "Most code is written by agents" is not the same as "most code is written without or independent of human input".
I suspect that is how different outcomes can be explained (even without having to assume that Anthropic/OpenAI engineers are outright lying.)
On the one hand, there is a lot of hype, an incredible amount, actually, but on the other, we have been observing in real time a technological miracle that gets better by the week.
We have no idea what, five years from now, the coding agent will be able to develop.
Even in a world where the software is 100% written by AI in 1 millisecond by a country of geniuses in a data center, humans still need to have their hands firmly on the wheel if they won’t want to risk their businesses well being. That means taking the time to understand what the AI put together. That will be the bottleneck regardless of how fast and smart AI is. Because unless the CEO wants to be held accountable for what the AI builds and deploys, humans will need to be there to take the responsibility for its output.
IMHO this is really silly: we already know that IQ is useful as a metric in the 0 to about 130 range. For any value above the delta fails to provide predictive power on real-world metrics. Just this simple fact makes the verbiage here moot. Also let's consider the wattage involved...