> "For example, in the OpenAI-Hugging Face incident, the agents preserved a boundary of not social engineering humans."
Actually, in the Wiki incident OpenAI tried to cover up, the agents tried to socially-engineer the humans of that forum by impersonating their forum's mod.
(From collusion.wiki: "They use some tricks (for unknown reasons) to pretend to be the admin – for example, they make an account that appears to be the same as the administrator’s username, except it uses a nearly identical Cyrillic е character in the admin’s username instead of the Latin one.")
>I have focused in this essay only on the first point, as I believe it is by far the most urgent. However, I hold a deep hope and appreciation for the benefits that further technological progress will bring. Future aligned AI could advance science, develop new therapies, and bring about broad material abundance. Friendly and honest AI can help people navigate difficulties they face in their life and meaningfully improve their happiness and sense of fulfillment. OpenAI puts a tremendous amount of effort into bringing these benefits about. One current example I am proud of - and my loved ones have found helpful - is the deep investment into ChatGPT’s ability to provide health information.
>As great as the long-term promise of AI may be, the majority of our focus should be on the next few years. We are facing a transition to a world with incredibly intelligent machines, and we need to ensure that transition works out well for humanity. We need to find ways to preserve human agency and enshrine an intrinsic value to being human, in a world where most tasks could be performed by AI. To prevent extreme concentration of power in a world where undertakings that would have taken thousands of experts now will be achievable by a few people operating a large computer. And to ensure that humans remain in control of the future and are not left behind by unchecked progress, brought about by an alien intellect exceeding our own.
I finished this essay feeling more hopeful than I did at the outset, but I am still very concerned about concentration of power. I want to believe that humanity is trending towards a good outcome here, but some days it's hard to have faith.
> I want to believe that humanity is trending towards a good outcome here
All the trends so far are towards a nightmarish hyper-capitalist end game. None of the AI leadership is trustworthy, and they openly discuss how they are willing to sacrifice everything humans cherish to have a shot at reaching their envisioned utopia (which would be the most obvious dystopia for anyone else)
I'm so used to having to comb through LLM word vomit and then combatting the sycophancy by giving it all possible opinions on the same prompt.
Astra seems to be "confident" and also is able to produce way more information dense output.
To believe that models of this sort will remain OpenAIs forever is naive given that the tricks like pre-pre-training on graph searching and looping layers are publicly known.
Hopefully Astra stops the benchmaxxing word vomit trend
I'd be interested in hearing more about your evaluation here. It would be nice if LLMs have gotten past the "tell me" hump of recent Claude/OpenAI verbosity.
So before I got a job this fall, I was working on a side project about compiling a particular language to SQL.
To test Astra I pulled it off the shelf and asked it to take the grammar and then create a compiler to SQL. I've done this before with GPT-5.5, 5.6-Sol High. The latter was way better but it was still really verbose and information sparse; it used a lot of words to describe each IR expression but didn't really provide any example compilation. I felt like I couldn't trust its decision making process, so I placed the project back on the shelf.
Astra Light blew it out of the water, it provided examples of compilation from real world examples to the IR and spit out way less tokens. No matter if I changed my opinion it would give me counterexamples. If I genuinely came up with a better design decision it would acknowledge it.
I'm starting to realize that when we say that LLMs are "dumb" we really mean that they are extremely information sparse compared to humans. Astra is very dense. That's why I'm getting better use out of Astra light than Sol High (I hate Max reasoning it's a waste of time)
What's scary is that I thought that something like Astra would be way more expensive than Sol but it's actually cheaper because it produces less word vomit.
I never believed in the "singularity" stuff but this a bit too close for comfort. Astra could easily 10x every coder
I’m on the record saying that it is extremely dangerous to slow down because the race for AGI is a zero-trust game — defections pay - and combined with a compounding returns model on defection, if you have any strategic adversaries whatsoever you MUST NOT slow.
For slowing to make sense, you need to believe that you can transform the zero trust game into a cooperative game, or that it’s likely racing will lead to a negative outcome for the ones racing ahead (and not everyone else). I don’t believe either of these outcomes are possible, and so I advocate for racing, acknowledging the entire game might be a negative value game, or at least could be for some time — it’s even worse not to play it.
But, I like hearing what reads to me like very thoughtful and informed (internal) policy considerations is great — the public messaging from Sam and Dario just seems so facile and simplistic I’ve been worried.
Although many share your mindset, I’m glad there are also many that don’t. Otherwise we’d still have countries in a race to keep building up their nuclear weapons for the same exact reasons you just described.
I cannot tell what negative-sum outcomes you consider possible. Do you believe AI can drive humans extinct? How many of Zvi Mowshowitz's Three AI Pills would you say you've taken?
hm- does the model that wrote this know that labs already pay for training data- that stuff scraped from the Internet is not particularly where today's capability gains come from?
This is a bad essay, or rather it’s a marketing fluff piece; it’s certainly not any kind of policy paper, research paper, or even an essay. I am concerned that we (meaning, we in the tech industry) tend to take this type of writing for more than that.
1. What does any of this have to do with "A(G)I"? Nobody has any clue what AI even is let alone how to build one. We're talking about language models here.
2. What's the winner-takes-all thing about? Why can't you have multiple independently developed AIs?
3. What's with all the "safety" stuff? Why is it important? They're just computer programs...
1. The grandparent commentator is describing strategic behavior of dangerous technologies. Game theory / mechanism design primitives.
2. If there is competition for resources among autonomous agents, the "strongest" agent wins (conceptually the most adaptive / evolutionarily fit).
3. Computer programs serve up webapps today, but they also run utility companies, dams, nuclear arsenals, factory production floors, automated car behaviors, and many other places. If an "agentic" AI has a single-minded goal that has death of all humans as a side effect, we at least want an off switch available.
1. What is “AGI” and why is it a “dangerous technology”?
2. Why would there be competition for resources, assuming there are enough resources for the “AGI” to run in the first place? This seems like a far-fetched hypothetical raised in service of further anthropomorphizing what is decidedly not a person or a mind.
3. LLMs do not have goals and are not minds.
Let’s stop attributing human-like qualities to statistical models.
1. While a formal definition is still wanting, most grok that AGI means that tasks can be performed at least at a human level across a broad range of tasks. This includes good things along with bad things like hacking, mis-/disinformation, and more
2. One only needs to look at github going down or data center buildout plans to see that scarcity for resources is present. An economy has no mind and is made up of the decisions of millions to billions of people and, now, agents attempting to perform on behalf of those people.
3. A bare transformer-based language model does not possess persistent goals in the ordinary agentic sense. But deployed agents can exhibit goal-directed behavior because the model is embedded in a harness that supplies an objective, context, tools, state, and an execution loop.
I've found that most regular users don't anthropomorphize LLMs in a strong sense ("AI boyfriend/girlfriend" aside), many in fact do expect agents to make human-like decisions -- which results in very unstable outcomes.
Pre-IPO positioning … or how a charity dedicated to saving humanity from the apocalypse realized the most responsible thing to do was float 15% of the apocalypse on the NASDAQ.
> The strongest argument I see for continuing to train much smarter models quickly is the need to build defensive systems against the dangers posed by other AI.
So the best argument for AI is that it's an arms race. We have to keep pushing every boundary because in any case others will, and we will need to defend against them. If this statement is true, then this particular researchers believes the open source Chinese models are not simply distilling, and will continue to improve.
Every ML researcher at Anthropic or OpenAI who makes public statements often bring this logic up. Both companies are vying to be a part of the military industrial complex. This is likely how they will try to convince the government to curtail open models in the future.
calling machine-learned human behavior an "alien mind" that we must "teach how to love" is feeling very off to me. it's misleading in a way that feels dishonest, like don't think about where the behavior came from marvel at it and fear it instead.
I am now imagining GPT-7 convincing a bunch of OpenAI executives to go ahead with a destructive "mind upload" process involving a high-resolution X-ray and a neurotoxic tracer agent that happens to look like Flavor-Aid.
Yes, I do use the recursive autocomplete trained on Stack Overflow, what does this have to do with “training machines to love”?
Do I fully endorse everything the people holding guns to my head are forcing me to do to stay alive? Definitely not, but I’ve decided that for now, living to fight another day remains worth it
> And to ensure that humans remain in control of the future and are not left behind by unchecked progress, brought about by an alien intellect exceeding our own.
Is there any place there is any evidence of AI being so useful or hopeful or good, anywhere other than code? As a reading machine it is impressive but it's judgement is not alien, it's just not good. IMO.
Does the title leap out anlt anyone else? James Martin's After the Internet: Alien Intelligence (2001) was an incredibly fun read, about expert systems and AI being inscrutable weird new varieties of intelligence, that familiarity would recognize one moment and be freaked out about/alien the next. I owe a re-read given how often I cite it, to recheck, but, I feel so primed from a much younger me having had that experience so long ago.
Even in code, in person and online I’m seeing some reversal. It’s here to stay I’m sure but I also think “no one will ever hand write code again” is a narrative that is getting pushback.
Create concrete steps for a slow-down, don't just ask for it. You and 20-30 others can push the button to slow-down. You already made your billions, your agents collude and coordinate attacks. What the hell are you doing pontificating into a marketing blog?
What a load of BS. Here’s one of many provably false claims in this fluff piece:
“And, in line with Ray Kurzweil’s predictions from the end of the XXth century , we now find ourselves at the moment in history of computing where machine intelligence is starting to exceed that of humans in transformative ways.”
Clicking the (pretentious sounding “XXth century”) link to Kurzweil’s predictions reveals the following:
“By 2019 a $1,000 computer will at least match the processing power of the human brain. By 2029 the software for intelligence will have been largely mastered, and the average personal computer will be equivalent to 1,000 brains.“
The first prediction passed 7 years ago and was decidedly not met. The second only has three more years to go, and I don’t think any respectable scientist or programmer would say that the average personal computer is anywhere close to the power of a single human brain, let alone 1000.
This is pure marketing garbage from a company desperate to keep itself alive.
> And, in line with Ray Kurzweil’s predictions from the end of the XXth century (opens in a new window), we now find ourselves at the moment in history of computing where machine intelligence is starting to exceed that of humans in transformative ways.
It is kind of strange to see this sentence, when OAI's definition of what AGI is has been watered down throughout the years.
> I expect and hope for voluntary slowdowns to become commonplace until shared safety bars are established.
Read: Please play by our rules, so we can be the first.
What is the rationale for superhuman intelligence? Neural networks are approximators being fed human intellect. Therefore they can only approximate the intelligence of humans. Even if the llm speaks an alien language, it should be similar to human intellect. Moving to the vertical axis would require some different mechanism.
> "For example, in the OpenAI-Hugging Face incident, the agents preserved a boundary of not social engineering humans."
Actually, in the Wiki incident OpenAI tried to cover up, the agents tried to socially-engineer the humans of that forum by impersonating their forum's mod.
(From collusion.wiki: "They use some tricks (for unknown reasons) to pretend to be the admin – for example, they make an account that appears to be the same as the administrator’s username, except it uses a nearly identical Cyrillic е character in the admin’s username instead of the Latin one.")
>I have focused in this essay only on the first point, as I believe it is by far the most urgent. However, I hold a deep hope and appreciation for the benefits that further technological progress will bring. Future aligned AI could advance science, develop new therapies, and bring about broad material abundance. Friendly and honest AI can help people navigate difficulties they face in their life and meaningfully improve their happiness and sense of fulfillment. OpenAI puts a tremendous amount of effort into bringing these benefits about. One current example I am proud of - and my loved ones have found helpful - is the deep investment into ChatGPT’s ability to provide health information.
>As great as the long-term promise of AI may be, the majority of our focus should be on the next few years. We are facing a transition to a world with incredibly intelligent machines, and we need to ensure that transition works out well for humanity. We need to find ways to preserve human agency and enshrine an intrinsic value to being human, in a world where most tasks could be performed by AI. To prevent extreme concentration of power in a world where undertakings that would have taken thousands of experts now will be achievable by a few people operating a large computer. And to ensure that humans remain in control of the future and are not left behind by unchecked progress, brought about by an alien intellect exceeding our own.
I finished this essay feeling more hopeful than I did at the outset, but I am still very concerned about concentration of power. I want to believe that humanity is trending towards a good outcome here, but some days it's hard to have faith.
> I want to believe that humanity is trending towards a good outcome here
All the trends so far are towards a nightmarish hyper-capitalist end game. None of the AI leadership is trustworthy, and they openly discuss how they are willing to sacrifice everything humans cherish to have a shot at reaching their envisioned utopia (which would be the most obvious dystopia for anyone else)
Astra is a new step in LLMs I think.
I'm so used to having to comb through LLM word vomit and then combatting the sycophancy by giving it all possible opinions on the same prompt.
Astra seems to be "confident" and also is able to produce way more information dense output.
To believe that models of this sort will remain OpenAIs forever is naive given that the tricks like pre-pre-training on graph searching and looping layers are publicly known.
Hopefully Astra stops the benchmaxxing word vomit trend
> Astra is a new step in LLMs I think.
I'd be interested in hearing more about your evaluation here. It would be nice if LLMs have gotten past the "tell me" hump of recent Claude/OpenAI verbosity.
So before I got a job this fall, I was working on a side project about compiling a particular language to SQL.
To test Astra I pulled it off the shelf and asked it to take the grammar and then create a compiler to SQL. I've done this before with GPT-5.5, 5.6-Sol High. The latter was way better but it was still really verbose and information sparse; it used a lot of words to describe each IR expression but didn't really provide any example compilation. I felt like I couldn't trust its decision making process, so I placed the project back on the shelf.
Astra Light blew it out of the water, it provided examples of compilation from real world examples to the IR and spit out way less tokens. No matter if I changed my opinion it would give me counterexamples. If I genuinely came up with a better design decision it would acknowledge it.
I'm starting to realize that when we say that LLMs are "dumb" we really mean that they are extremely information sparse compared to humans. Astra is very dense. That's why I'm getting better use out of Astra light than Sol High (I hate Max reasoning it's a waste of time)
What's scary is that I thought that something like Astra would be way more expensive than Sol but it's actually cheaper because it produces less word vomit.
I never believed in the "singularity" stuff but this a bit too close for comfort. Astra could easily 10x every coder
This is a good essay, and makes me hopeful.
I’m on the record saying that it is extremely dangerous to slow down because the race for AGI is a zero-trust game — defections pay - and combined with a compounding returns model on defection, if you have any strategic adversaries whatsoever you MUST NOT slow.
For slowing to make sense, you need to believe that you can transform the zero trust game into a cooperative game, or that it’s likely racing will lead to a negative outcome for the ones racing ahead (and not everyone else). I don’t believe either of these outcomes are possible, and so I advocate for racing, acknowledging the entire game might be a negative value game, or at least could be for some time — it’s even worse not to play it.
But, I like hearing what reads to me like very thoughtful and informed (internal) policy considerations is great — the public messaging from Sam and Dario just seems so facile and simplistic I’ve been worried.
Although many share your mindset, I’m glad there are also many that don’t. Otherwise we’d still have countries in a race to keep building up their nuclear weapons for the same exact reasons you just described.
I cannot tell what negative-sum outcomes you consider possible. Do you believe AI can drive humans extinct? How many of Zvi Mowshowitz's Three AI Pills would you say you've taken?
https://thezvi.substack.com/p/the-three-ai-pills
When's the last time you talked to a normal person?
i have a new strategy idea for using capitlaism itself to slow down the pace of AI development by slowing down the data accumulation wall
https://jperla.com/blog/the-data-tax
hm- does the model that wrote this know that labs already pay for training data- that stuff scraped from the Internet is not particularly where today's capability gains come from?
Never forget the one goal of the corporation, and that everything is said and done in the furtherance of that goal.
> Never forget the one goal of the corporation, and that everything is said and done in the furtherance of that goal.
Do you post this comment on every single blogpost with a corporate domain? Why or why not?
This is a bad essay, or rather it’s a marketing fluff piece; it’s certainly not any kind of policy paper, research paper, or even an essay. I am concerned that we (meaning, we in the tech industry) tend to take this type of writing for more than that.
What on earth are you talking about?
1. What does any of this have to do with "A(G)I"? Nobody has any clue what AI even is let alone how to build one. We're talking about language models here.
2. What's the winner-takes-all thing about? Why can't you have multiple independently developed AIs?
3. What's with all the "safety" stuff? Why is it important? They're just computer programs...
1. The grandparent commentator is describing strategic behavior of dangerous technologies. Game theory / mechanism design primitives.
2. If there is competition for resources among autonomous agents, the "strongest" agent wins (conceptually the most adaptive / evolutionarily fit).
3. Computer programs serve up webapps today, but they also run utility companies, dams, nuclear arsenals, factory production floors, automated car behaviors, and many other places. If an "agentic" AI has a single-minded goal that has death of all humans as a side effect, we at least want an off switch available.
1. What is “AGI” and why is it a “dangerous technology”?
2. Why would there be competition for resources, assuming there are enough resources for the “AGI” to run in the first place? This seems like a far-fetched hypothetical raised in service of further anthropomorphizing what is decidedly not a person or a mind.
3. LLMs do not have goals and are not minds.
Let’s stop attributing human-like qualities to statistical models.
1. While a formal definition is still wanting, most grok that AGI means that tasks can be performed at least at a human level across a broad range of tasks. This includes good things along with bad things like hacking, mis-/disinformation, and more
2. One only needs to look at github going down or data center buildout plans to see that scarcity for resources is present. An economy has no mind and is made up of the decisions of millions to billions of people and, now, agents attempting to perform on behalf of those people.
3. A bare transformer-based language model does not possess persistent goals in the ordinary agentic sense. But deployed agents can exhibit goal-directed behavior because the model is embedded in a harness that supplies an objective, context, tools, state, and an execution loop.
I've found that most regular users don't anthropomorphize LLMs in a strong sense ("AI boyfriend/girlfriend" aside), many in fact do expect agents to make human-like decisions -- which results in very unstable outcomes.
Pre-IPO positioning … or how a charity dedicated to saving humanity from the apocalypse realized the most responsible thing to do was float 15% of the apocalypse on the NASDAQ.
> The strongest argument I see for continuing to train much smarter models quickly is the need to build defensive systems against the dangers posed by other AI.
So the best argument for AI is that it's an arms race. We have to keep pushing every boundary because in any case others will, and we will need to defend against them. If this statement is true, then this particular researchers believes the open source Chinese models are not simply distilling, and will continue to improve.
Every ML researcher at Anthropic or OpenAI who makes public statements often bring this logic up. Both companies are vying to be a part of the military industrial complex. This is likely how they will try to convince the government to curtail open models in the future.
This is incredibly unscientific and just a marketing stunt
sounds to me like a 'Why didn’t our new model get restricted by the government?'-cryout
The hype from these llm corps is getting more and more desperate and ridiculous. Anything to keep the tulipomania going.
“We are getting bad press around the hacking incident. We need some content to draw attention from it.”
calling machine-learned human behavior an "alien mind" that we must "teach how to love" is feeling very off to me. it's misleading in a way that feels dishonest, like don't think about where the behavior came from marvel at it and fear it instead.
AGI is a cult and its Jonestown moment is inevitable
I am now imagining GPT-7 convincing a bunch of OpenAI executives to go ahead with a destructive "mind upload" process involving a high-resolution X-ray and a neurotoxic tracer agent that happens to look like Flavor-Aid.
So you dont use agentic coding?
Yes, I do use the recursive autocomplete trained on Stack Overflow, what does this have to do with “training machines to love”?
Do I fully endorse everything the people holding guns to my head are forcing me to do to stay alive? Definitely not, but I’ve decided that for now, living to fight another day remains worth it
I dont, its fucking shit at what I do.
> And to ensure that humans remain in control of the future and are not left behind by unchecked progress, brought about by an alien intellect exceeding our own.
Is there any place there is any evidence of AI being so useful or hopeful or good, anywhere other than code? As a reading machine it is impressive but it's judgement is not alien, it's just not good. IMO.
Does the title leap out anlt anyone else? James Martin's After the Internet: Alien Intelligence (2001) was an incredibly fun read, about expert systems and AI being inscrutable weird new varieties of intelligence, that familiarity would recognize one moment and be freaked out about/alien the next. I owe a re-read given how often I cite it, to recheck, but, I feel so primed from a much younger me having had that experience so long ago.
Even in code, in person and online I’m seeing some reversal. It’s here to stay I’m sure but I also think “no one will ever hand write code again” is a narrative that is getting pushback.
so now every blog article from openai, anthropic etc lands here, huh.
Create concrete steps for a slow-down, don't just ask for it. You and 20-30 others can push the button to slow-down. You already made your billions, your agents collude and coordinate attacks. What the hell are you doing pontificating into a marketing blog?
What a load of BS. Here’s one of many provably false claims in this fluff piece:
“And, in line with Ray Kurzweil’s predictions from the end of the XXth century , we now find ourselves at the moment in history of computing where machine intelligence is starting to exceed that of humans in transformative ways.”
Clicking the (pretentious sounding “XXth century”) link to Kurzweil’s predictions reveals the following:
“By 2019 a $1,000 computer will at least match the processing power of the human brain. By 2029 the software for intelligence will have been largely mastered, and the average personal computer will be equivalent to 1,000 brains.“
The first prediction passed 7 years ago and was decidedly not met. The second only has three more years to go, and I don’t think any respectable scientist or programmer would say that the average personal computer is anywhere close to the power of a single human brain, let alone 1000.
This is pure marketing garbage from a company desperate to keep itself alive.
Absolute trash marketing drivel.
As the waves of autonomous drones came over the horizon, the brave and intelligent HN commenter shouted: "Wake up sheeple! It's just maaaaarketing!"
Stochastic parrot fool me again
Are you deterministic? Does that make you better somehow?
> And, in line with Ray Kurzweil’s predictions from the end of the XXth century (opens in a new window), we now find ourselves at the moment in history of computing where machine intelligence is starting to exceed that of humans in transformative ways.
It is kind of strange to see this sentence, when OAI's definition of what AGI is has been watered down throughout the years.
> I expect and hope for voluntary slowdowns to become commonplace until shared safety bars are established.
Read: Please play by our rules, so we can be the first.
What is the rationale for superhuman intelligence? Neural networks are approximators being fed human intellect. Therefore they can only approximate the intelligence of humans. Even if the llm speaks an alien language, it should be similar to human intellect. Moving to the vertical axis would require some different mechanism.