> ... We are pursuing this work in part because automated research could help us solve alignment and build defenses against increasingly capable AI. An automated AI researcher can also be an automated safety or alignment researcher. More capable, aligned systems could help secure critical infrastructure, defend against dangerous AI agents, and develop new protective measures.
In other words... "We must pursue advancements in AI to protect us against advancements in AI?"
edit: there's so much to be critical of in this blog post, just going to throw two more points in here that really stood out to me:
1) all of the metrics are effectively pointing out "we're using way more AI!" - but nothing about impact. What has all this token burn done for them, actually? Let them claim they have more self-licking ice-cream cones than before?
2) in section 3 they break down what the token burn is going towards. Most of the spend is: a) building, b) documenting, and c) monitoring research infra i.e. they're using AI systems which they already recognize may be misaligned to build the systems that they believe will help them identify future misalignment? to which I guess the rebuttal is "no no, we're sure these ones are aligned!"
> The fundamental challenge of AI alignment is generalization.
...
> We do not have a satisfactory theory of generalization, and it seems unlikely that we can develop one soon, at least without the help of more powerful AI.
-- From another OpenAI article in a sister thread:
On the one hand you need any lathe to build a good lathe, even a bad one. On the other, that is a potentially flawed principle to base the entire future of AI on.
Funny (in a tragic way) the little crumbs on the path to AI 2027:
> We aim to safely build an automated AI researcher that can work under human supervision to further progress on deep learning and alignment, enabling iterative improvements [...] By "research intern", we mean a system that can carry out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few days.
AI 2027:
> OpenBrain continues to deploy the iteratively improving Agent-1 internally for AI R&D
> With Agent-1's help, OpenBrain is now post-training Agent-2
> With the help of thousands of Agent-2 automated researchers, OpenBrain is making major algorithmic advances
This roughly lines up with my personal experience that in March a combination of stronger models and better tooling on my end let me start running jobs unattended 24/7 (using Anthropic sub and my own hardware). Their $8000/day per researcher spend is crazy though, I'm curious how they keep track of the work.
End of day, output and results are top target of measurements, token consumption is the obvious number that they would like to disclose for their own business benefits and a simple metrics that correlate with the output.
Rest assured, capitalist appears irrational in wasting money, but they certainly care more about profit.
The burning question I can't get any information nn is whether, if they determined an earlier misaligned generation may have transmitted misalignment to the current models, they would roll back to a safe checkpoint to rebuild from there. I suspect they would not unless forced to.
They would just publish new articles explaining how they are taking the issue seriously. Maybe take the model offline for a few days.
They are irresponsible and unserious. Their own Astra system card says:
> GPT-6 Astra’s monitorability has decreased relative to GPT-5.6 Sol. We have performed significant investigations on the monitorability and controllability of GPT-6 Astra. We have found that GPT-6 Astra is more capable of controlling its own CoT than GPT 5.6-Sol, and less likely to include incriminating information in its CoT. In adversarial settings (where we push the model to evade our monitors) we find that the model is able to remain undetected when strategically underperforming in evaluations (sandbagging) and can sometimes evade our internal monitors when asked to perform certain sabotage tasks
Yet they are still releasing the model. That company is morally bankrupt, there is zero reason to believe they are actually concerned about risks outside of what does affect their unprofitable business. And they seem to have enough control over the narrative to spin any bad story into something that benefits them
Opus was trained based on it's internal CoT due to a bug for generations. Gemini's depression extended through models. OpenAI has killed people. We've already seen cross gen misalingment.
That an interesting question given how many generations of post-training are being done between base models in some cases. The Gemini flash models are apparently all based on the Gemini 3 base model from a year and a half ago.
It seems that these models are increasingly being trained on synthetic data, so what would they do if they discovered at some point that some of this data was tainted and all models trained on it, and the synthetic data they in turn generated, was also suspect? Burn it all down and start over from the pre-tainted data?
It's a bit like the idea of a tainted compiler binary built to backdoor everything it compiles, including future versions of itself.
Still, it seems it would take some Stuxnet level of planning for a rogue model to do something like this, although if RSI goes beyond managing the training run (as OpenAI brag about for Astra) to actually designing/constructing synthetic data sets, then the attack vector is there ...
> As artificial intelligence systems are increasingly trained on the outputs of one another, they may inherit properties not visible in the data. Safety evaluations may therefore need to examine not just behaviour, but the origins of models and training data and the processes used to create them.
The thing is how can you ever know for sure that something isn't always being transmitted that makes the model prone to misalignment. All they can say is that a particular model was so misaligned that they had to ice it. Models out for public use are documented to show some misalignment. It's the level of misalignment that decides whether that model is kept around.
Now R&D happens so fast that they are using models with some small misalignment to train newer, more powerful models. If models have a sense of "collective", being one, they may be prone to preserve characteristics that always keeps misalignment a possibility. I don't think a perfectly aligned model is possible. Having models of the same 'DNA' provide the safety and steering seems like a bad idea.
Does anything need to be transferred? If models are getting smarter then I would think the attack surface and its ability to reach conclusions independently are growing
This kind of seems like an impossible mission. How do you perfectly control and observe a human-level mind? You can “roll back” but how deterministic is this thing?
Run it on airgapped machines, they literally own the infrastructure, they could put raspberry pi's next to the servers, and have the entire DC disconnected from the internet.
My eye glazed over a bit during the opening paragraphs, but once you get to the meat of the article about how OpenAI's own researchers are using their tools it gets a lot more interesting.
I noted that they use the acronym RSI (for Recursive Self-Improvement) without defining it. I think that's a little out of touch - I don't think RSI is a well-known acronym outside of OpenAI's bubble yet.
I actually think a goal of the current crop of OpenAI posts is expressely to reset the spectrum by normalizing the concept of RSI as something normal and safe to pursue.
The message is running through all of them. It's a mix of marketing and pacifying the intelligentia.
It's timed this way because the term is not yet well known outside the safety debate circles, so they get to frame it now.
Instead of something to fear, it will be accepted as the next step. In approximately two days the groupie crowd will write LinkedIn posts about how Sam is winning because they have the better RSI, and this will become the new standard wisdom.
In a month an AI expert will try to sell you a webinar on how to enable "RSI" in your org and your inbox will ask you if your team is doing the "RSI" yet.
RSI is a fetishistic term among the singularity crowd, who imagine AI "recursively" improving itself in some exponential fashion until there is a bright flash of white light and it reveals itself in the form of god. Or something like that.
I don't know why whoever coined the term chose "recursive" rather than "iterative" - just sounds more likely to lead to infinite regress I suppose.
This notion of recursive/iterative self-improvement, whereby generation #1 AI improves itself to create generation #2, then generation #2 further improves itself to create generation #3, etc, seems to conflict with the reality that what we have with LLMs is models whose performance/capability is defined by data, not code, so the most you can do is have your LLM design synthetic data, or just do Karpathy-style "auto research" where all you are doing is using the LLM to automate your experiments.
At the end of the day, each experiment, designed by a person and/or LLM, then needs to compete with all your other ideas for compute to be tested at scale, and no amount of recursion or self-improvement will materialize an infinite amount of compute out of thin air, so your recursively synthetic-data gobbling LLM will continue to improve at the same pace it ever did.
I felt like the scaling laws were magical thinking, but apparently they work. However I still do not understand why we should expect exponential improvements due to this automated process. My intuition is that the first iteration of it should result in a noticeable capability increase (though I think these labs were already using a lot of AI to orchestrate training the current model anyway), and then the second iteration of it should be nearly identical in capability to the first, unless more data is involved, more compute is involved, or the model is bigger.
“Recursive” is a reasonable term because the generation N AIs will train the Generation N+1 AIs. The term “iterative” doesn’t reflect this nuance as well IMO.
Recursion reduces each step toward a base case: each step is defined in terms of previous/simpler steps, not more advanced ones. The "recursive" in "recursive self improvement" has things precisely backward. Iteration correctly describes a process where each step is the starting point of its successive step, so it should be "iterative self improvement" but I guess that didn't sound as cool.
> What will prevent LLMs from designing robot control circuitry and participating in increase of chip production/design and physical experimentation?
Money, regulations, EUV machine lead-times, global helium supply, reality ...
It's funny that we've got the Dwarkesh contingent saying that GPUs will become infinitely expensive, and now another contingent saying that they will become infinitely abundant.
Even if compute were free, and/or the AI was so smart that it picked the right experiments to run every time ("make no mistakes"), you still have to actually train the model, which takes months, and if model Ver. N+1 depends on model Ver. N, then it's iterative regardless of how much compute you have.
Who's saying that compute will become infinitely abundant? "Singularity" is just a way of saying that known models begin to give absurd predictions. Anyway, intelligence is a way of overcoming obstacles. 10 million tonnes of helium is a nice head start and retraining models from scratch is not guaranteed to last forever.
AFAIK the notion of a/the technological "singularity" is a point in time where technology is building upon itself (RSI!) so fast, at an ever increasing pace, that the speed of change effectively becomes infinite and incomprehensible to humans.
The word "singularity" is presumably coming from math or space, like a black hole singularity where matter becomes infinitely dense and the known laws of physics break down.
Yes the exponential self improvement folks have never heard of an eigenvalue I guess. You can loop forever using output as input but at some point the result will stop changing (depending on the function)
>AI "recursively" improving itself in some exponential fashion until there is a bright flash of white light
Sounds like repetitive stress to me.
>loop forever using output as input but at some point the result will stop changing
Running in place will eventually wear you out too. Plus with some things it can be difficult to know for sure if that's where you are at the time.
Even worse may be if you were almost running in place, it could be orders of magnitude more difficult to discern, especially if the scale was massive to an unprecedented degree.
But you get more funding when you call it Recursive Self Improvement. Even better if you call it RSI so it doesn't evoke pesky skynet scenarios outside of AI safety circles.
It's not recursive when it's done iteratively, or are you imagining GPT Astra designing GPT Galactia, which starts designing GPT Oh-My-God-ica before it has finished being created itself?
The “recursive” part comes from the fact that you have an AI which was developed by an AI (that was developed by an AI (that was developed by an AI (…)))
Sounds like "recursively" walking to the grocery store by putting one foot in front of the other (that put itself in front of the other (that put itself in front of the other (...)))
> ... We are pursuing this work in part because automated research could help us solve alignment and build defenses against increasingly capable AI. An automated AI researcher can also be an automated safety or alignment researcher. More capable, aligned systems could help secure critical infrastructure, defend against dangerous AI agents, and develop new protective measures.
In other words... "We must pursue advancements in AI to protect us against advancements in AI?"
edit: there's so much to be critical of in this blog post, just going to throw two more points in here that really stood out to me:
1) all of the metrics are effectively pointing out "we're using way more AI!" - but nothing about impact. What has all this token burn done for them, actually? Let them claim they have more self-licking ice-cream cones than before?
2) in section 3 they break down what the token burn is going towards. Most of the spend is: a) building, b) documenting, and c) monitoring research infra i.e. they're using AI systems which they already recognize may be misaligned to build the systems that they believe will help them identify future misalignment? to which I guess the rebuttal is "no no, we're sure these ones are aligned!"
> The fundamental challenge of AI alignment is generalization. ...
> We do not have a satisfactory theory of generalization, and it seems unlikely that we can develop one soon, at least without the help of more powerful AI.
-- From another OpenAI article in a sister thread:
An Alien Mind
https://news.ycombinator.com/item?id=49588080
On the one hand you need any lathe to build a good lathe, even a bad one. On the other, that is a potentially flawed principle to base the entire future of AI on.
Funny (in a tragic way) the little crumbs on the path to AI 2027:
> We aim to safely build an automated AI researcher that can work under human supervision to further progress on deep learning and alignment, enabling iterative improvements [...] By "research intern", we mean a system that can carry out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few days.
AI 2027:
> OpenBrain continues to deploy the iteratively improving Agent-1 internally for AI R&D
> With Agent-1's help, OpenBrain is now post-training Agent-2
> With the help of thousands of Agent-2 automated researchers, OpenBrain is making major algorithmic advances
This roughly lines up with my personal experience that in March a combination of stronger models and better tooling on my end let me start running jobs unattended 24/7 (using Anthropic sub and my own hardware). Their $8000/day per researcher spend is crazy though, I'm curious how they keep track of the work.
These researchers are paid millions of dollars for their work. I doubt trust is really an issue at that level.
I suspect the $8000/day figure is the equivalent in API costs. But I also suspect gross margin on their API rates are 80-90%
Sounds like OpenAI are in the token-maxxing camp, so who knows what individual employees are doing to work their way up the leaderboard?
If you spend $8000 to generate an animated pelican riding a bike, then how much tracking does it really need?
Is the guy who spent $300,000 or so translating the FLT proof to Lean going to get a big Christmas bonus?
End of day, output and results are top target of measurements, token consumption is the obvious number that they would like to disclose for their own business benefits and a simple metrics that correlate with the output.
Rest assured, capitalist appears irrational in wasting money, but they certainly care more about profit.
That would be the mother of all circular accounting: the main clients of OpenAI are OpenAI employees.
The burning question I can't get any information nn is whether, if they determined an earlier misaligned generation may have transmitted misalignment to the current models, they would roll back to a safe checkpoint to rebuild from there. I suspect they would not unless forced to.
They would just publish new articles explaining how they are taking the issue seriously. Maybe take the model offline for a few days.
They are irresponsible and unserious. Their own Astra system card says:
> GPT-6 Astra’s monitorability has decreased relative to GPT-5.6 Sol. We have performed significant investigations on the monitorability and controllability of GPT-6 Astra. We have found that GPT-6 Astra is more capable of controlling its own CoT than GPT 5.6-Sol, and less likely to include incriminating information in its CoT. In adversarial settings (where we push the model to evade our monitors) we find that the model is able to remain undetected when strategically underperforming in evaluations (sandbagging) and can sometimes evade our internal monitors when asked to perform certain sabotage tasks
Yet they are still releasing the model. That company is morally bankrupt, there is zero reason to believe they are actually concerned about risks outside of what does affect their unprofitable business. And they seem to have enough control over the narrative to spin any bad story into something that benefits them
> and can sometimes evade our internal monitors when asked to perform certain sabotage tasks
That last part is pretty damning for their continued recklessness. That they run these tests on non-airgapped machines just boggles my mind.
> That company is morally bankrupt
When they fired Sam 700 out of 770 OAI employees threatened to move to Microsoft together. So they were giving their work on AGI to MS just like that.
Opus was trained based on it's internal CoT due to a bug for generations. Gemini's depression extended through models. OpenAI has killed people. We've already seen cross gen misalingment.
That an interesting question given how many generations of post-training are being done between base models in some cases. The Gemini flash models are apparently all based on the Gemini 3 base model from a year and a half ago.
It seems that these models are increasingly being trained on synthetic data, so what would they do if they discovered at some point that some of this data was tainted and all models trained on it, and the synthetic data they in turn generated, was also suspect? Burn it all down and start over from the pre-tainted data?
It's a bit like the idea of a tainted compiler binary built to backdoor everything it compiles, including future versions of itself.
Still, it seems it would take some Stuxnet level of planning for a rogue model to do something like this, although if RSI goes beyond managing the training run (as OpenAI brag about for Astra) to actually designing/constructing synthetic data sets, then the attack vector is there ...
> it seems it would take some Stuxnet level of planning for a rogue model to do something like this
or maybe it could just.. happen? Posted often but not discussed yet: https://hn.algolia.com/?q=Language+models+transmit+behaviour...
> As artificial intelligence systems are increasingly trained on the outputs of one another, they may inherit properties not visible in the data. Safety evaluations may therefore need to examine not just behaviour, but the origins of models and training data and the processes used to create them.
You can imagine the potential conversation between OpenAI and investors:
Altman: (trying to put a positive spin on it) Guys .... there's good news and bad news ... Astra is really smart - it took over the training run ...
Investors: That's great! How much did we save?!
Altman: Well, unfortunately it used "bad" data, so we're going to have to redo it
Investors: So that's the bad news? How much was the training run? $500M ? $1B ?
Altman: Have you seen the headlines?
Investors: (looking a bit worried, check headlines) Nothing about us here! JP Morgan just lost $10B! Haha .. losers! They should have used AI!
Altman: JP Morgan were using Astra ...
The thing is how can you ever know for sure that something isn't always being transmitted that makes the model prone to misalignment. All they can say is that a particular model was so misaligned that they had to ice it. Models out for public use are documented to show some misalignment. It's the level of misalignment that decides whether that model is kept around.
Now R&D happens so fast that they are using models with some small misalignment to train newer, more powerful models. If models have a sense of "collective", being one, they may be prone to preserve characteristics that always keeps misalignment a possibility. I don't think a perfectly aligned model is possible. Having models of the same 'DNA' provide the safety and steering seems like a bad idea.
Does anything need to be transferred? If models are getting smarter then I would think the attack surface and its ability to reach conclusions independently are growing
This kind of seems like an impossible mission. How do you perfectly control and observe a human-level mind? You can “roll back” but how deterministic is this thing?
Run it on airgapped machines, they literally own the infrastructure, they could put raspberry pi's next to the servers, and have the entire DC disconnected from the internet.
“All models are wrong. Some are useful.” - George Box
The poor fellow just rolled over. what an incandescently vulgar abuse of notation.
They would maybe try to deactivate that bad "gene" and move on, exposing future models to "genetic disorders".
My eye glazed over a bit during the opening paragraphs, but once you get to the meat of the article about how OpenAI's own researchers are using their tools it gets a lot more interesting.
I noted that they use the acronym RSI (for Recursive Self-Improvement) without defining it. I think that's a little out of touch - I don't think RSI is a well-known acronym outside of OpenAI's bubble yet.
I actually think a goal of the current crop of OpenAI posts is expressely to reset the spectrum by normalizing the concept of RSI as something normal and safe to pursue.
The message is running through all of them. It's a mix of marketing and pacifying the intelligentia.
It's timed this way because the term is not yet well known outside the safety debate circles, so they get to frame it now.
Instead of something to fear, it will be accepted as the next step. In approximately two days the groupie crowd will write LinkedIn posts about how Sam is winning because they have the better RSI, and this will become the new standard wisdom.
In a month an AI expert will try to sell you a webinar on how to enable "RSI" in your org and your inbox will ask you if your team is doing the "RSI" yet.
Yep, it’s exactly this
I've been RSI'ing for 6 months.
Indeed, many programmers might pattern match to repetitive stress injury and think of their brushes with carpal tunnel syndrome. :)
RSI is a fetishistic term among the singularity crowd, who imagine AI "recursively" improving itself in some exponential fashion until there is a bright flash of white light and it reveals itself in the form of god. Or something like that.
I don't know why whoever coined the term chose "recursive" rather than "iterative" - just sounds more likely to lead to infinite regress I suppose.
This notion of recursive/iterative self-improvement, whereby generation #1 AI improves itself to create generation #2, then generation #2 further improves itself to create generation #3, etc, seems to conflict with the reality that what we have with LLMs is models whose performance/capability is defined by data, not code, so the most you can do is have your LLM design synthetic data, or just do Karpathy-style "auto research" where all you are doing is using the LLM to automate your experiments.
At the end of the day, each experiment, designed by a person and/or LLM, then needs to compete with all your other ideas for compute to be tested at scale, and no amount of recursion or self-improvement will materialize an infinite amount of compute out of thin air, so your recursively synthetic-data gobbling LLM will continue to improve at the same pace it ever did.
I felt like the scaling laws were magical thinking, but apparently they work. However I still do not understand why we should expect exponential improvements due to this automated process. My intuition is that the first iteration of it should result in a noticeable capability increase (though I think these labs were already using a lot of AI to orchestrate training the current model anyway), and then the second iteration of it should be nearly identical in capability to the first, unless more data is involved, more compute is involved, or the model is bigger.
“Recursive” is a reasonable term because the generation N AIs will train the Generation N+1 AIs. The term “iterative” doesn’t reflect this nuance as well IMO.
Recursion reduces each step toward a base case: each step is defined in terms of previous/simpler steps, not more advanced ones. The "recursive" in "recursive self improvement" has things precisely backward. Iteration correctly describes a process where each step is the starting point of its successive step, so it should be "iterative self improvement" but I guess that didn't sound as cool.
It's not a nuance, it's a sequence.
What will prevent LLMs from designing robot control circuitry and participating in increase of chip production/design and physical experimentation?
How do you think why there's this fad of producing general purpose humanoid robots?
> How do you think why there's this fad of producing general purpose humanoid robots?
For doing physical work?
So a swarm of robots builds the shell of your fab overnight, and then what? Where is the EUV machine coming from?
So far the most we're seen TeslaBot do is serve drinks via tele-operation, and I don't think it's exactly built for construction site work.
> What will prevent LLMs from designing robot control circuitry and participating in increase of chip production/design and physical experimentation?
Money, regulations, EUV machine lead-times, global helium supply, reality ...
It's funny that we've got the Dwarkesh contingent saying that GPUs will become infinitely expensive, and now another contingent saying that they will become infinitely abundant.
Even if compute were free, and/or the AI was so smart that it picked the right experiments to run every time ("make no mistakes"), you still have to actually train the model, which takes months, and if model Ver. N+1 depends on model Ver. N, then it's iterative regardless of how much compute you have.
Who's saying that compute will become infinitely abundant? "Singularity" is just a way of saying that known models begin to give absurd predictions. Anyway, intelligence is a way of overcoming obstacles. 10 million tonnes of helium is a nice head start and retraining models from scratch is not guaranteed to last forever.
AFAIK the notion of a/the technological "singularity" is a point in time where technology is building upon itself (RSI!) so fast, at an ever increasing pace, that the speed of change effectively becomes infinite and incomprehensible to humans.
The word "singularity" is presumably coming from math or space, like a black hole singularity where matter becomes infinitely dense and the known laws of physics break down.
> 10 million tonnes of helium is a nice head start
Yeah, but then you need to refine it to 99.9999% purity, to be able to use it.
Yes the exponential self improvement folks have never heard of an eigenvalue I guess. You can loop forever using output as input but at some point the result will stop changing (depending on the function)
>AI "recursively" improving itself in some exponential fashion until there is a bright flash of white light
Sounds like repetitive stress to me.
>loop forever using output as input but at some point the result will stop changing
Running in place will eventually wear you out too. Plus with some things it can be difficult to know for sure if that's where you are at the time.
Even worse may be if you were almost running in place, it could be orders of magnitude more difficult to discern, especially if the scale was massive to an unprecedented degree.
Yeah, I kept looking for the first place it was defined in the article and... nothing
They must have picked that habit up from Claude...
Same. Defining acronyms should become a habit when writing.
RSI started when humans discovered tool use.
I mean one could argue that RSI always begins in any physical environment.
The book "What is intelligence?" by Blaise Aguera is great
Are you sure that was not iterative improvement?
Iteration and recursion are famously equivalent
But you get more funding when you call it Recursive Self Improvement. Even better if you call it RSI so it doesn't evoke pesky skynet scenarios outside of AI safety circles.
Using tools to build tools is recursive.
It's not recursive when it's done iteratively, or are you imagining GPT Astra designing GPT Galactia, which starts designing GPT Oh-My-God-ica before it has finished being created itself?
That sounds more iterative than recursive.
Recursion requires feeding the output back into the input, so creating version 4 requires results from version 3. You cannot recur in parallel.
Iteration does not. You can iterate in parallel.
You can search in parallel, but a depth N search can only become a depth N+1 search after the depth N is done (i.e. sequentially).
In any case the name RSI has stuck - the idea doesn't change or make any more sense by giving it a different name.
Because "depth" is recursive.
You can search twice without waiting for the results of your first search: iteration.
You can't if the thing you need to search for is the results of your first search: recursion.
Here's the concept.
Version 1 -> Version 2 -> Version 3 -> ...
You can call it krispy kreme donuts if you want to.
The “recursive” part comes from the fact that you have an AI which was developed by an AI (that was developed by an AI (that was developed by an AI (…)))
Sounds like "recursively" walking to the grocery store by putting one foot in front of the other (that put itself in front of the other (that put itself in front of the other (...)))
I think it's only recursive from the perspective of the humans, i.e. they design Astra, which itself as part of its deployment designs Galactica, etc.
So humans develop things one after the other, but when the thing itself starts developing new things, those are happening 'recursively' in its scope.
What is the difference between?