So, DeepSeek wins the LLM race and then...literally nothing. It's not a race America needed to win. There's no ROI, no moat, and not much future for LLMs. This tech is too expensive and not useful or reliable enough. It's a toy that was commoditized before we figured out what it was for.
all of your claims are easy to understand, but none of them make sense. they are vague, informal claims based on vague intuitions. "there is no RoI" can be claimed about every investment if we don't specify the period. google had no RoI for years. how reliable is Meta? economic claims need to be based on data, not impressions.
There is no ROI after 3.5 years and no ROI in sight. Datacenters are not being completed, GPUs are being hoarded but unused until they break down, circular deals hide the fact that every frontier AI model is an utter economic failure. Emterprise is starting to rehire humans they thought they no longer needed. The whole thing is a heist on common sense.
If it wasn’t useful at all, people wouldn’t be using. It js certainly useful for coding. The ROI of throwing Opus at every problem probably isn’t there, but planning with it and using something cheaper for implementation. Or using a cheap-but-still-good model like GLM5.2.
I think the problem with this argument is that it assumes AGI-seeking LLM-based models are incapable of referencing, and coordinating, all of the more reliable data and computing methods of 2022 (and since) as effectively as a massive group of (also plenty fallible) humans can. Which isn’t the case: those models can already do so significantly more effectively. And in areas like cybersecurity, one of Anthropic and the US Government’s primary concerns, that dynamic is making for something that really is starting to make the “Manhattan Project” claims look a bit less hyperbolic. For an example, see the day zero findings by Mythos that caused the scare about its release.
While I agree that products like ChatGPT and Claude were initially commoditized as toys that didn’t have much real use yet, and that the doomsday/utopia claims of 2023 were largely marketing BS at that point, that isn’t the case anymore… so you’re behind the times in still claiming that. I don’t think it’s reasonable to continue calling models like the ones in question toys when they already could easily destroy every computer system and most electrical infrastructure on the planet if given open access to the internet with the instructions and hardware support to do so. We’re not going back to 2022.
You've swallowed a bunch of lies. This is all theater. None of the claims about the capabilities of models like Mythos have been corroborrated with evidence. It's pre-IPO marketing hype aided and abetted by a corrupt government that also seeks to profit from the fraud.
We’ll see. I’m a retired military officer whose background includes some cyber warfare and high-level strategy, now looking at this as an academic since 2022. To my knowledge, the cyber threat is legitimate—and based on my experience, the DoD and Three-letter support for these policies means their actual concerns are probably worse than those that have been publicly released. While I agree that there’s plenty of hype and corruption involved in what we see in the news, that doesn’t mean there isn’t also something real going on behind it all that experts in this field can’t keep denying based on outdated assumptions.
I would have 100% agreed with you in 2023-24, and appreciated your playing hardball devil’s advocate with the “AI” true believers that actually did fall for the pure theater back then. But, at this point, I wouldn’t have bothered mentioning the emergent problem with the “just a toy” argument if i’d seen your page first: I’d incorrectly assumed you just hadn’t been following how things have changed since that argument was a good one. I have no skin in the game itself, so don’t have a good reason to argue with someone so deeply committed to any specific mantra on the subject, and was just trying to help someone I thought was here for reasons similar to mine.
Are you sure you're not the one too deeply committed to believing the hype around mythos? The AISI report clearly found that the model is about as capable as a mediocre average hacker who already has easy access to a compromised system. It only completed the task 3 out of 10 times. Quote: "Mythos Preview’s success on one cyber range indicates that it is at least capable of autonomously attacking small, weakly defended and vulnerable enterprise systems where access to a network has been gained." So no different from the average security script kiddie. How is it a national security threat? It's like you're not reading what's actually written. My only crime is keeping a level head and knowing how often these oligarchs and politicians lie through their teeth to exact their desired level of control over us.
Yes, that is the concern. A mediocre hacker that can autonomously find minor vulnerabilities, then theorize and chain them into breaches
Most real world 'hackers' just use social engineering, and/or scripts that check a system against a list of known vulnerabilities/exploits
People capable of actually finding/chaining together *new* exploits are pretty rare (and a big chunk of these folks work for nation states, and do not make their findings public). Mythos acts like a mediocre version of one of *these* hackers, not like a mediocre script kiddy
If the model was made public without safeguards, then your run of the mill script kiddy would be capable of doing much, much more damage
The concern isn't about the quality of the attacks Mythos produces (the best humans do beat it). It's about the sheer volume of specific individualized attacks it could perform, in the hands of the wrong people
And my understanding is at this point it wouldn’t necessarily do that good at autonomous hacking on its own. In terms of being able to get around in a system for an extended period without detection. But that doesn’t mean that the ability to find exploits and figure out which ones are useful together to make a more serious exploit aren’t super powerful on their own. As you said, finding new usable exploits is a big deal.
Thank you for this informative update. I can see that both of the competing AI frontier (Anthropic & OpenAi), have serialized or cloned in parallel each of their model development side by side with advancement of nuclear development,therefore it is indeed a "highly dangerous" instrument. While the technology is further developing at a rapid pace & phase, and by leaving it into a level of uncobtrollable momentum, in an alarming pressure point, which can probably be unleashed or use as an alternative very lethal pseudo kind of authority to replace the government as they think it fit a legitimate entity & authority next in power to government, in which case it can become a probable condition which theory and assumption refers it back from previous paragraph.
I fear that Alberto is right about this. Choice has been withdrawn from the market. The best model(s) are to be withheld from average citizens. The ultimate elitist scheme is coming to pass. Possession of Fable or Mythos will become punishable and offenders will be deprived of their right ever to own the best models because they are deemed unworthy. And it was accomplished without the nasty, unending fight that is being waged over gun rights.
I ask you, which is more dangerous to human happiness, guns or Fable?
I’m afraid to inform you - and you should know, that AI is genuinely existential as a risk and therefore, regulating it is something that should have happened, not now, but months or years ago.
Yup, what's shocking is not that it's happening, it's that it took so long to be this blatant. I wrote this on Medium almost 2 years ago:
AI makes me a better programmer. AI makes engineers better engineers, doctors better doctors, automobile technicians better automobile technicians. Hell, it would even make politicians better politicians if any of them gave a crap about the quality of their policy.
But AI also makes scammers better scammers, astroturfers better astroturfers, propagandists better propagandists, misinformation peddlers better misinformation peddlers, sociopathic power brokers better sociopathic power brokers, etc.
The point is that, yeah maybe someday the agency of an AI will actually be a problem. Yeah we should think about that. But look at the major players who, right now, are promoting the AI agency fear. If you exclude anyone who is personally benefitting from the fear mongering (because conflict of interest), you're left with tech corporations, their representatives, and politicians. These are the very people who want to control or facilitate the control of AI into the hands of the few.
Your incident report not withstanding, Gen-AI is not even on the horizon. What we have today, right now, is the very real threat of people, not AI, subverting our public institutions and threatening our future. If we don't solve the problem of people with too much power and no accountability, then we won't ever need to face the problem of a gen-AI run amuck.
The Manhattan Project parallel keeps showing up across every major AI thinker for a reason. It's not a metaphor anymore. It's a description. Staggered release, government-approved access lists, trusted partner tiers, capability classified as strategic. Every one of those steps has a direct precedent in how nuclear technology moved from commercial research to national security asset in the 1940s, and the transition happened the same way: gradually, then all at once, and by the time anyone outside the programme realised the rules had changed, the rules had already been in place for months.
The part that should keep every developer outside the US awake tonight is the asymmetry. China responds to this by releasing open-weight models under MIT licence with no regional restrictions, which means the restriction doesn't contain the capability. It relocates who distributes it. The US built the best models and gated them. China built good-enough models and gave them away. A year from now the global AI infrastructure outside America will run on Chinese foundations not because those foundations are better but because they were the ones that showed up without conditions. The commercial era of American AI didn't end because the technology got too dangerous. It ended because the government decided the technology was too important to share, and the rest of the world went shopping elsewhere.
For starters, it would upset a lot of investors in Anthropic. They just filed for IPO.
You would need to believe a) there is no path to profitability, and b) they view nationalization as an off ramp. Is that what you believe is happening here?
Or do you genuinely believe Anthropic believes their technology is so scary they desperately want to hand over all responsibility?
That doesn’t really line up with any of their public policy proposals or advocacy, does it? I just find it hard to believe that what we are seeing is the result of some covert agenda pushed on the part of Anthropic. The order of events seem to be largely driven by incompetence.
Agree with your comment. For me the idea does not sum up:
1) a nationalized AI would eventually not be competitive with other AIs that evolve in the free market, including chinese ones. Governments are not good entrepreneurs. The Atomic bomb had no private competitors, but AI has extremely strong ones.
Lets assume, Anthropic reaches a sort of "AGI" and this is then under gov. control. But what we see, as that the competition is behind a couple of months: what is the performance of Anthropics frontier model today is reached by cheaper competitors a little later. There is no way for any government in the world to project this moat.
Last but not least I have a hard time to believe, that its the goal of Dario Amodei to have Trump and Hegseth as his Bosses tomorrow.
--> That said, I still value a lot Albertos inspiring writing about this stuff - even if I dont agree with any of his conclusions !
Thank you for engaging Jochen. Let me clarify a few things that I may have not conveyed well.
- AFAIK, the largest/most important projects humanity has achieved are all government-based or government-adjacent and I don't think AI will be different.
- This doesn't mean the best people in the world stop doing what they're doing. Commercial pressures were always a race to the bottom and essentially no one in the safety space wanted that. They've always advocated for tighter control.
- The moats are there but only for those with eyes to see. The government has more moats than anyone else has (eg intelligence agencies, military, fed, etc., which are extremely useful for ambitious projects)
- This is not about the Trump administration (who can trust then?). It's about a national project for AI being at the frontier. This is not about any political party but about taking AI seriously. Those who claim we're going to have a "country of geniuses in a datacenter" are taking it seriously.
- We have been sold the lie that capitalism is about companies fairly competing in a free market but the shadow of the state looms over everything. AI will be no different if it becomes too consequential to ignore.
1) yes, it is definitively remarkable that now the gov. steps in to control the release of the most powerful models - in 2 instances!
2) Yes, the government would want to get an increasingly powerful AI under its control (be it a Trump gov. or even a more serious one).
3) the people inside Anthropic and Open AI might actually believe, that they are heading to a technology that is too powerful to make it freely accessible (and it might be even somehow True!)
Where I have a doubt:
1) With "moat" I mean, the gov. has not means to prevent, that tomorrow DeepSeek builds a model similarly powerful than todays best Anthropic model under gov. control. (or even - lets dream a bit - that we Europeans build one). The threshold to do that is a big chunk of cash - and hire some good people (again: look DeepSeek). The threshold to build an atomic bomb is MUCH higher.
2) The Trump administration will be there till late 2028. I doubt, that Amodei looks beyond 2029, and hands over the control over his baby to a Trump administration, hoping that in 2029 more decent people will govern.
3) The idea, that the gov. takes control is not compatible with the story of the IPO. (OK, you might think, that these IPOs will eventually not take place at all).
I’m not sure it is necessarily incompatible with an IPO. Like I think Boeing is a public company and they make commercial airplanes but also act as a defense contractor making military planes.
From what I’ve seen, it sounds like Mythos being so good at exploiting vulnerabilities was not intentional. And probably there wasn’t time to do a more elaborate re-training to try to keep the general coding capabilities while suppressing that. So they stuck guardrails on top to get Fable. But ideally there’d be separate models that are good for commercial purposes (and even pointing out security flaws), but don’t go as far as Mythos does vs another model that is more restricted and geared toward defense. GPT5.6 vs their security model might be more like what Anthropic would have preferred.
So just because capabilities keep getting better and the access to some models is restricted doesn’t necessarily mean the lab can’t keep selling useful stuff to the public that keeps getting better in its own way. It might start to really force specialization vs the everything all in one models we’ve been seeing.
....or is it that the U.S. government has been anthropocized? If democracy represents choice and natural diversity, Alberto Romero has just described what Chamath et al have been saying all over the media for the last week or two. And I am persuaded that the VC and the commentator are both right.
Real risk isn't the model getting held back. It's every delay creating a bigger opening for non-US labs. I covered that side here; export controls meant to box out China, ending up making US models optional for everyone else.
So do you think as this iterates, the next frontrunner in line is eventually going to decide NOT to build their next release? I suspect this doesn't go the way this administration imagines it going.
the speed mismatch is the real problem honestly. regulators are trying to write rules for something that changes every 3 months, and builders are shipping stuff without knowing what the rules will be. feels like everyone's just guessing at this point
Here is what I am thinking. If the performance on benchmarks is only a fraction of a percentage or just a few percentages higher than open source models, that is a cause for optimism. The regulatory swoop-in itself, in addition to the market economics, may bring more attention on smaller models. I see this as a good thing. I am not sold that scale is the only path to capabiities. It reads as a lazy man's path forward. Perhaps a silver lining by confluences. Constraint as a forcing function. Constraint as an instigator for, hopefully, innovation beyond autoregressive frameworks.
So, DeepSeek wins the LLM race and then...literally nothing. It's not a race America needed to win. There's no ROI, no moat, and not much future for LLMs. This tech is too expensive and not useful or reliable enough. It's a toy that was commoditized before we figured out what it was for.
What?
Which part do you not understand?
all of your claims are easy to understand, but none of them make sense. they are vague, informal claims based on vague intuitions. "there is no RoI" can be claimed about every investment if we don't specify the period. google had no RoI for years. how reliable is Meta? economic claims need to be based on data, not impressions.
There is no ROI after 3.5 years and no ROI in sight. Datacenters are not being completed, GPUs are being hoarded but unused until they break down, circular deals hide the fact that every frontier AI model is an utter economic failure. Emterprise is starting to rehire humans they thought they no longer needed. The whole thing is a heist on common sense.
The same part you don't: how you're going to falsify that.
If it wasn’t useful at all, people wouldn’t be using. It js certainly useful for coding. The ROI of throwing Opus at every problem probably isn’t there, but planning with it and using something cheaper for implementation. Or using a cheap-but-still-good model like GLM5.2.
Agreed!
I think the problem with this argument is that it assumes AGI-seeking LLM-based models are incapable of referencing, and coordinating, all of the more reliable data and computing methods of 2022 (and since) as effectively as a massive group of (also plenty fallible) humans can. Which isn’t the case: those models can already do so significantly more effectively. And in areas like cybersecurity, one of Anthropic and the US Government’s primary concerns, that dynamic is making for something that really is starting to make the “Manhattan Project” claims look a bit less hyperbolic. For an example, see the day zero findings by Mythos that caused the scare about its release.
While I agree that products like ChatGPT and Claude were initially commoditized as toys that didn’t have much real use yet, and that the doomsday/utopia claims of 2023 were largely marketing BS at that point, that isn’t the case anymore… so you’re behind the times in still claiming that. I don’t think it’s reasonable to continue calling models like the ones in question toys when they already could easily destroy every computer system and most electrical infrastructure on the planet if given open access to the internet with the instructions and hardware support to do so. We’re not going back to 2022.
You've swallowed a bunch of lies. This is all theater. None of the claims about the capabilities of models like Mythos have been corroborrated with evidence. It's pre-IPO marketing hype aided and abetted by a corrupt government that also seeks to profit from the fraud.
We’ll see. I’m a retired military officer whose background includes some cyber warfare and high-level strategy, now looking at this as an academic since 2022. To my knowledge, the cyber threat is legitimate—and based on my experience, the DoD and Three-letter support for these policies means their actual concerns are probably worse than those that have been publicly released. While I agree that there’s plenty of hype and corruption involved in what we see in the news, that doesn’t mean there isn’t also something real going on behind it all that experts in this field can’t keep denying based on outdated assumptions.
I would have 100% agreed with you in 2023-24, and appreciated your playing hardball devil’s advocate with the “AI” true believers that actually did fall for the pure theater back then. But, at this point, I wouldn’t have bothered mentioning the emergent problem with the “just a toy” argument if i’d seen your page first: I’d incorrectly assumed you just hadn’t been following how things have changed since that argument was a good one. I have no skin in the game itself, so don’t have a good reason to argue with someone so deeply committed to any specific mantra on the subject, and was just trying to help someone I thought was here for reasons similar to mine.
Are you sure you're not the one too deeply committed to believing the hype around mythos? The AISI report clearly found that the model is about as capable as a mediocre average hacker who already has easy access to a compromised system. It only completed the task 3 out of 10 times. Quote: "Mythos Preview’s success on one cyber range indicates that it is at least capable of autonomously attacking small, weakly defended and vulnerable enterprise systems where access to a network has been gained." So no different from the average security script kiddie. How is it a national security threat? It's like you're not reading what's actually written. My only crime is keeping a level head and knowing how often these oligarchs and politicians lie through their teeth to exact their desired level of control over us.
Yes, that is the concern. A mediocre hacker that can autonomously find minor vulnerabilities, then theorize and chain them into breaches
Most real world 'hackers' just use social engineering, and/or scripts that check a system against a list of known vulnerabilities/exploits
People capable of actually finding/chaining together *new* exploits are pretty rare (and a big chunk of these folks work for nation states, and do not make their findings public). Mythos acts like a mediocre version of one of *these* hackers, not like a mediocre script kiddy
If the model was made public without safeguards, then your run of the mill script kiddy would be capable of doing much, much more damage
The concern isn't about the quality of the attacks Mythos produces (the best humans do beat it). It's about the sheer volume of specific individualized attacks it could perform, in the hands of the wrong people
And my understanding is at this point it wouldn’t necessarily do that good at autonomous hacking on its own. In terms of being able to get around in a system for an extended period without detection. But that doesn’t mean that the ability to find exploits and figure out which ones are useful together to make a more serious exploit aren’t super powerful on their own. As you said, finding new usable exploits is a big deal.
So, we will go back to the times before AI developments?
Back to 2022, not the end of the world.
Best writer on AI there is
🙏🏻🙏🏻🙏🏻
Thank you for this informative update. I can see that both of the competing AI frontier (Anthropic & OpenAi), have serialized or cloned in parallel each of their model development side by side with advancement of nuclear development,therefore it is indeed a "highly dangerous" instrument. While the technology is further developing at a rapid pace & phase, and by leaving it into a level of uncobtrollable momentum, in an alarming pressure point, which can probably be unleashed or use as an alternative very lethal pseudo kind of authority to replace the government as they think it fit a legitimate entity & authority next in power to government, in which case it can become a probable condition which theory and assumption refers it back from previous paragraph.
I fear that Alberto is right about this. Choice has been withdrawn from the market. The best model(s) are to be withheld from average citizens. The ultimate elitist scheme is coming to pass. Possession of Fable or Mythos will become punishable and offenders will be deprived of their right ever to own the best models because they are deemed unworthy. And it was accomplished without the nasty, unending fight that is being waged over gun rights.
I ask you, which is more dangerous to human happiness, guns or Fable?
I’m afraid to inform you - and you should know, that AI is genuinely existential as a risk and therefore, regulating it is something that should have happened, not now, but months or years ago.
Regulation is not bad. This post is not about regulation.
You trust Mr. Trump with the future of humanity?
I trust that if there is no regulation at all, it will work out as well as the accelerating car wheels have worked out for small animals on the road.
And I am particular about my children having a future.
Yup, what's shocking is not that it's happening, it's that it took so long to be this blatant. I wrote this on Medium almost 2 years ago:
AI makes me a better programmer. AI makes engineers better engineers, doctors better doctors, automobile technicians better automobile technicians. Hell, it would even make politicians better politicians if any of them gave a crap about the quality of their policy.
But AI also makes scammers better scammers, astroturfers better astroturfers, propagandists better propagandists, misinformation peddlers better misinformation peddlers, sociopathic power brokers better sociopathic power brokers, etc.
The point is that, yeah maybe someday the agency of an AI will actually be a problem. Yeah we should think about that. But look at the major players who, right now, are promoting the AI agency fear. If you exclude anyone who is personally benefitting from the fear mongering (because conflict of interest), you're left with tech corporations, their representatives, and politicians. These are the very people who want to control or facilitate the control of AI into the hands of the few.
Your incident report not withstanding, Gen-AI is not even on the horizon. What we have today, right now, is the very real threat of people, not AI, subverting our public institutions and threatening our future. If we don't solve the problem of people with too much power and no accountability, then we won't ever need to face the problem of a gen-AI run amuck.
The Manhattan Project parallel keeps showing up across every major AI thinker for a reason. It's not a metaphor anymore. It's a description. Staggered release, government-approved access lists, trusted partner tiers, capability classified as strategic. Every one of those steps has a direct precedent in how nuclear technology moved from commercial research to national security asset in the 1940s, and the transition happened the same way: gradually, then all at once, and by the time anyone outside the programme realised the rules had changed, the rules had already been in place for months.
The part that should keep every developer outside the US awake tonight is the asymmetry. China responds to this by releasing open-weight models under MIT licence with no regional restrictions, which means the restriction doesn't contain the capability. It relocates who distributes it. The US built the best models and gated them. China built good-enough models and gave them away. A year from now the global AI infrastructure outside America will run on Chinese foundations not because those foundations are better but because they were the ones that showed up without conditions. The commercial era of American AI didn't end because the technology got too dangerous. It ended because the government decided the technology was too important to share, and the rest of the world went shopping elsewhere.
Do you really believe Anthropic's grand plan is to get nationalized by the US government?
Yes, that's instrumental for their goals
I find that hard to believe.
Why?
For starters, it would upset a lot of investors in Anthropic. They just filed for IPO.
You would need to believe a) there is no path to profitability, and b) they view nationalization as an off ramp. Is that what you believe is happening here?
Or do you genuinely believe Anthropic believes their technology is so scary they desperately want to hand over all responsibility?
The last one is the correct one. Yes they do believe that (you stated it in a way that's perhaps too exaggerated but essentially yeah)
That doesn’t really line up with any of their public policy proposals or advocacy, does it? I just find it hard to believe that what we are seeing is the result of some covert agenda pushed on the part of Anthropic. The order of events seem to be largely driven by incompetence.
Agree with your comment. For me the idea does not sum up:
1) a nationalized AI would eventually not be competitive with other AIs that evolve in the free market, including chinese ones. Governments are not good entrepreneurs. The Atomic bomb had no private competitors, but AI has extremely strong ones.
Lets assume, Anthropic reaches a sort of "AGI" and this is then under gov. control. But what we see, as that the competition is behind a couple of months: what is the performance of Anthropics frontier model today is reached by cheaper competitors a little later. There is no way for any government in the world to project this moat.
Last but not least I have a hard time to believe, that its the goal of Dario Amodei to have Trump and Hegseth as his Bosses tomorrow.
--> That said, I still value a lot Albertos inspiring writing about this stuff - even if I dont agree with any of his conclusions !
Thank you for engaging Jochen. Let me clarify a few things that I may have not conveyed well.
- AFAIK, the largest/most important projects humanity has achieved are all government-based or government-adjacent and I don't think AI will be different.
- This doesn't mean the best people in the world stop doing what they're doing. Commercial pressures were always a race to the bottom and essentially no one in the safety space wanted that. They've always advocated for tighter control.
- The moats are there but only for those with eyes to see. The government has more moats than anyone else has (eg intelligence agencies, military, fed, etc., which are extremely useful for ambitious projects)
- This is not about the Trump administration (who can trust then?). It's about a national project for AI being at the frontier. This is not about any political party but about taking AI seriously. Those who claim we're going to have a "country of geniuses in a datacenter" are taking it seriously.
- We have been sold the lie that capitalism is about companies fairly competing in a free market but the shadow of the state looms over everything. AI will be no different if it becomes too consequential to ignore.
thank you, Alberto for responding!
Where I think you are right:
1) yes, it is definitively remarkable that now the gov. steps in to control the release of the most powerful models - in 2 instances!
2) Yes, the government would want to get an increasingly powerful AI under its control (be it a Trump gov. or even a more serious one).
3) the people inside Anthropic and Open AI might actually believe, that they are heading to a technology that is too powerful to make it freely accessible (and it might be even somehow True!)
Where I have a doubt:
1) With "moat" I mean, the gov. has not means to prevent, that tomorrow DeepSeek builds a model similarly powerful than todays best Anthropic model under gov. control. (or even - lets dream a bit - that we Europeans build one). The threshold to do that is a big chunk of cash - and hire some good people (again: look DeepSeek). The threshold to build an atomic bomb is MUCH higher.
2) The Trump administration will be there till late 2028. I doubt, that Amodei looks beyond 2029, and hands over the control over his baby to a Trump administration, hoping that in 2029 more decent people will govern.
3) The idea, that the gov. takes control is not compatible with the story of the IPO. (OK, you might think, that these IPOs will eventually not take place at all).
I’m not sure it is necessarily incompatible with an IPO. Like I think Boeing is a public company and they make commercial airplanes but also act as a defense contractor making military planes.
From what I’ve seen, it sounds like Mythos being so good at exploiting vulnerabilities was not intentional. And probably there wasn’t time to do a more elaborate re-training to try to keep the general coding capabilities while suppressing that. So they stuck guardrails on top to get Fable. But ideally there’d be separate models that are good for commercial purposes (and even pointing out security flaws), but don’t go as far as Mythos does vs another model that is more restricted and geared toward defense. GPT5.6 vs their security model might be more like what Anthropic would have preferred.
So just because capabilities keep getting better and the access to some models is restricted doesn’t necessarily mean the lab can’t keep selling useful stuff to the public that keeps getting better in its own way. It might start to really force specialization vs the everything all in one models we’ve been seeing.
....or is it that the U.S. government has been anthropocized? If democracy represents choice and natural diversity, Alberto Romero has just described what Chamath et al have been saying all over the media for the last week or two. And I am persuaded that the VC and the commentator are both right.
Good point on Z.ai and DeepSeek.
Oh dear, that is good to know. Thank you 🙏🏽
Real risk isn't the model getting held back. It's every delay creating a bigger opening for non-US labs. I covered that side here; export controls meant to box out China, ending up making US models optional for everyone else.
https://thisweekinaiclub.substack.com/p/the-state-of-ai-the-us-is-losing
Worth reading.
Why people are hell bent on killing things in this country?
So do you think as this iterates, the next frontrunner in line is eventually going to decide NOT to build their next release? I suspect this doesn't go the way this administration imagines it going.
Sounds like another Growthfroth reset.
the speed mismatch is the real problem honestly. regulators are trying to write rules for something that changes every 3 months, and builders are shipping stuff without knowing what the rules will be. feels like everyone's just guessing at this point
Here is what I am thinking. If the performance on benchmarks is only a fraction of a percentage or just a few percentages higher than open source models, that is a cause for optimism. The regulatory swoop-in itself, in addition to the market economics, may bring more attention on smaller models. I see this as a good thing. I am not sold that scale is the only path to capabiities. It reads as a lazy man's path forward. Perhaps a silver lining by confluences. Constraint as a forcing function. Constraint as an instigator for, hopefully, innovation beyond autoregressive frameworks.