Total nonsense. Taking literally the notion that ‘knowledge is power’ without regard for the requirements of physical implementation. Or should we believe that knowing how to build a nuclear weapon is the same as building one? This is essentially a God complex.
Thank you so much for posting this essay! I used to work in a biology lab and, at a gut level, I have never agreed with the Yud take along the lines that, "AGI will make a virus to kill everyone." But also, I wasnt smart enough to explain my sense in words. Like the author said, the claim is so baseless and wrong that it's hard to even know where to begin. YES: "Dario’s writing is all premised on some tacit assumption that performance on these tasks can in some sense be dramatically improved with sufficient supply of cognitive labor."
Also, I don't agree with what Trump did but it is 100% hilarious that the big brains (human and otherwise) at Anthropic didn't see Trump's action coming. Maybe they should have asked Mythos? If Mythos is so smart, shouldn't it have been able to game out the consequences of fearmongering about its capabilities? Why didn't it come up with an effective plan to prevent Trump to issuing the export ban? Surely, once the all-wise Mythos identified the risk of an export ban, it could have suggested helpful changes to Dario's ridiculous behavior, suggestions that would have both been effective and been likely accepted by Dario. Isn't that what recursive self-improvement would require? It's almost as if the events of the past couple days show the limits of cognitive labor.
I think the author makes some fair points about Anthropic’s biology claims, benchmarks, and biosecurity rhetoric. I am not sure I agree with all of them, but they are debatable and worth taking seriously.
Where I disagree is with the framing of Dario Amodei as somehow deserving this. Dario has been clear and public about his concerns. Maybe he is right, maybe he is too cautious, and maybe Anthropic deserves criticism on execution. But he has behaved in a principled way.
The Fable and Mythos 5 order looks heavy-handed, and it has the feel of singling out one of the few AI billionaires who has not bent the knee to Trump. Criticize Anthropic’s claims if warranted, but the bigger concern, in my opinion, is the Trump administration using state power in a blunt and politically suspicious way.
The US government did not ban the model. It issued unprecedented export controls that target nationality instead of market. Decent regulation also would not make a one-off decision for one specific model but ones that specify criteria for all models that pose a risk.
That is not quite right. Anthropic did not simply say “our model is dangerous” and then get banned as the logical consequence.
Anthropic had been cooperating with the government for over two months before release. The government was given access to Mythos-class capabilities through Project Glasswing starting in early April, well before Fable was released as the guarded version for broader use.
The issue cited later, which is the excuse given for the export control, was a narrow potential bypass, not a universal jailbreak. The odd part is that Anthropic says the same kind of capability is available in other frontier models, including GPT-5.5. So why was only Anthropic hit with export controls?
Maybe it was a misunderstanding. Maybe the government overreacted. Or maybe Anthropic is being treated differently because Dario has been one of the few AI leaders willing to publicly disagree with Trump. I don’t know which is true, but “Anthropic said it was dangerous, so the government banned it” leaves out a lot of important context.
I feel we can have both positions be somewhat true: yes, having a large amount of increased intelligence will accellerate e.g: drug development (I am a professional modeler in big pharma, and I know we're seeing large speedups everywhere, in large part thanks to the newfound ability of specialists to translate their complex ideas into testable code - e.g: a PBPK scientist can have a bayesian inference ready model for a new idea he has or a new paper in a few hours or days, instead of waiting a year for one of the company's statisticians to do it by hand). On the other hand, in silico is very imperfect, and nothing can replace in-vitro/ in-vivo reality checks. these are SLOW - even if you imagine a perfectly automated retrosynthesis + synthesis engine to make the compounds you want to test that will take a lot of time (thermodynamics and kinetics are sadly hard problems to mitigate :), some assays even in vitro mechanistically take days (even without random QC failures, shipment delays etc) and in vivo means litterally raising an animal and monitor before an analysis step, which is very much not quick (and the genetic variability between animals make perfect repeatability impossible). growing a mouse and making a molecule are always going to be limiting factors time-wise, and that means a rate limitation on intelligence's ability to self-calibrate.
Maybe if models develop: perfect genetics understanding of model organisms, fast cloning without defects, the most incredible assays (which include theoretical material science for inert vials, etc), an understanding of chemical reactions and retrosythesis that allow for optimal synthesis of arbitrary compounds, perfect proteomics, better imaging, etc... then the bottlenecks would be minimized.
Also, no one in big pharma could make real use of Fable (even before the takedown) due to the classifiers. any chem/bio discussion was shut down immediatly.
Great stuff, Cremieux. Work on my longevity blog brings an instant demotion to Opus. Anthropic has clearly lost the plot on how to make the world safer from bio weapons.
Total nonsense. Taking literally the notion that ‘knowledge is power’ without regard for the requirements of physical implementation. Or should we believe that knowing how to build a nuclear weapon is the same as building one? This is essentially a God complex.
yes the guy with a PhD in Biophysics has no idea how biology works lol
Did you actually think that was a substantive comment?
Excellent read sir, thank you
Thank you so much for posting this essay! I used to work in a biology lab and, at a gut level, I have never agreed with the Yud take along the lines that, "AGI will make a virus to kill everyone." But also, I wasnt smart enough to explain my sense in words. Like the author said, the claim is so baseless and wrong that it's hard to even know where to begin. YES: "Dario’s writing is all premised on some tacit assumption that performance on these tasks can in some sense be dramatically improved with sufficient supply of cognitive labor."
Also, I don't agree with what Trump did but it is 100% hilarious that the big brains (human and otherwise) at Anthropic didn't see Trump's action coming. Maybe they should have asked Mythos? If Mythos is so smart, shouldn't it have been able to game out the consequences of fearmongering about its capabilities? Why didn't it come up with an effective plan to prevent Trump to issuing the export ban? Surely, once the all-wise Mythos identified the risk of an export ban, it could have suggested helpful changes to Dario's ridiculous behavior, suggestions that would have both been effective and been likely accepted by Dario. Isn't that what recursive self-improvement would require? It's almost as if the events of the past couple days show the limits of cognitive labor.
I think the author makes some fair points about Anthropic’s biology claims, benchmarks, and biosecurity rhetoric. I am not sure I agree with all of them, but they are debatable and worth taking seriously.
Where I disagree is with the framing of Dario Amodei as somehow deserving this. Dario has been clear and public about his concerns. Maybe he is right, maybe he is too cautious, and maybe Anthropic deserves criticism on execution. But he has behaved in a principled way.
The Fable and Mythos 5 order looks heavy-handed, and it has the feel of singling out one of the few AI billionaires who has not bent the knee to Trump. Criticize Anthropic’s claims if warranted, but the bigger concern, in my opinion, is the Trump administration using state power in a blunt and politically suspicious way.
Anthropic said their model was dangerous. The US government did the logic thing and banned it.
The US government did not ban the model. It issued unprecedented export controls that target nationality instead of market. Decent regulation also would not make a one-off decision for one specific model but ones that specify criteria for all models that pose a risk.
I think the US government used whatever leverage and urgency it could, given Anthropic's claims of how dangerous their model was.
That is not quite right. Anthropic did not simply say “our model is dangerous” and then get banned as the logical consequence.
Anthropic had been cooperating with the government for over two months before release. The government was given access to Mythos-class capabilities through Project Glasswing starting in early April, well before Fable was released as the guarded version for broader use.
The issue cited later, which is the excuse given for the export control, was a narrow potential bypass, not a universal jailbreak. The odd part is that Anthropic says the same kind of capability is available in other frontier models, including GPT-5.5. So why was only Anthropic hit with export controls?
Maybe it was a misunderstanding. Maybe the government overreacted. Or maybe Anthropic is being treated differently because Dario has been one of the few AI leaders willing to publicly disagree with Trump. I don’t know which is true, but “Anthropic said it was dangerous, so the government banned it” leaves out a lot of important context.
I feel we can have both positions be somewhat true: yes, having a large amount of increased intelligence will accellerate e.g: drug development (I am a professional modeler in big pharma, and I know we're seeing large speedups everywhere, in large part thanks to the newfound ability of specialists to translate their complex ideas into testable code - e.g: a PBPK scientist can have a bayesian inference ready model for a new idea he has or a new paper in a few hours or days, instead of waiting a year for one of the company's statisticians to do it by hand). On the other hand, in silico is very imperfect, and nothing can replace in-vitro/ in-vivo reality checks. these are SLOW - even if you imagine a perfectly automated retrosynthesis + synthesis engine to make the compounds you want to test that will take a lot of time (thermodynamics and kinetics are sadly hard problems to mitigate :), some assays even in vitro mechanistically take days (even without random QC failures, shipment delays etc) and in vivo means litterally raising an animal and monitor before an analysis step, which is very much not quick (and the genetic variability between animals make perfect repeatability impossible). growing a mouse and making a molecule are always going to be limiting factors time-wise, and that means a rate limitation on intelligence's ability to self-calibrate.
Maybe if models develop: perfect genetics understanding of model organisms, fast cloning without defects, the most incredible assays (which include theoretical material science for inert vials, etc), an understanding of chemical reactions and retrosythesis that allow for optimal synthesis of arbitrary compounds, perfect proteomics, better imaging, etc... then the bottlenecks would be minimized.
Also, no one in big pharma could make real use of Fable (even before the takedown) due to the classifiers. any chem/bio discussion was shut down immediatly.
Thoughts on this new grant for APOE2(Christchurch mutation added) gene therapy in APOE4 homozygotes? https://reporter.nih.gov/search/TGw9-7-6gEWnwqK4EpwoKg/project-details/11291103 - https://news.weill.cornell.edu/news/2026/07/weill-cornell-medicine-awarded-grant-for-alzheimer%E2%80%99s-gene-therapy-research
Great stuff, Cremieux. Work on my longevity blog brings an instant demotion to Opus. Anthropic has clearly lost the plot on how to make the world safer from bio weapons.