Mythos is a myth, nothing mythical about it . Every software engineer worth his salt knows what's up, AND the markets are starting to catch up, so this pathetic lying administration jumps onto helping big tech bros with this Hollywoodic narrative of "national security risk" to hype up the stocks. I'm sure China is enjoying watching America in self destructive mode
It's only mythical, ie mythological thought. AI is software causality which has nothing to do with real events, thought, intelligence, completion. Reality and thought are massive, parallel oscillations. There's nothing clockwork, linear, compute in reality.
This is false, unstable technology that will never offer a step beyond optimization.
There are no adults in the White House. Worse, the vanishingly few adults in Congress and the Supreme Court have nowhere near the numbers they need to be able to regain control of either branch, regardless of electoral outcome.
Unless we step outside our usual role of "vote every two years", this nightmare will never end.
Danielle Church: I have to say, while I'm glad you are "doing something," the use of that four-letter-word, as a comprehensive/symbolic title especially, is and always will be to me, extremely off-putting. (I slip and use it myself sometimes but always regret it.) Is there something I'm not seeing here? If so, clue me in so that I can understand?
And yet, you still knew which word she was talking about! And I knew that she wasn't talking about that particular conjugation of the word, but about all of its forms. So it seems to me like she used the ideal communication for the context... is there something I've missed?
Danielle Church: I won't die from it in any version. However, and thank you for explaining your usage of that word (regardless of how many other letters surround it). To Joy: It stands out as a "four-letter word" precisely for the same reasons Daniel wants to use it as introductory.
The other thought is that I certainly don't want to "doll up" anything that has to do with Trump and Company. Fxxk works in that regard and at least the lowlife that inhabits at least some of MAGA will love it.
The "on the other hand" idea, however, (and you most likely knew that was coming next) is that some just won't take it seriously or even think it's too crass to identify with (even though I think it beats "disambiguate").
Enough . . . and though the method of communication is a significant part of any idea worth moving along, let's not make it any more of a diversion. In this case, like no other, doing nothing is not an option.
Exactly, exactly, exactly right, Catherine. Doing nothing is not an option.
I made the choice that I want to give my efforts the chance for the most emotive power and force they possibly can, even though I know I'm giving up a lot on the "first impression" front, and that my subscriber count just isn't going to climb as fast as it would if I'd gone with a cleaner title.
This world works best when we all stick to our strengths, though, and mine are an irreverent but extremely caring humanist warmth and a complete refusal to compromise in my vision of what "success" means. If what our world needs is a different kind of voice to lead it out of this darkness, then I am so okay with not being that voice! (So, so very okay. I am terrified of being a public figure.) I wouldn't be nearly as good at anything except for being the person that I truly am, and my patriotism leans much closer to "fierce" than "proud".
In my ideal world, I am one of many, or at least a handful, of voices all offering us true, actionable paths to restore harmony to our world. I am doing everything I can to make my effort the most quintessentially "Danielle Church" path that it can be, so that once we're all standing on this flaming stage of an Internet, the public will be able to choose the one that offers the best chance of success for the most people.
I know they must be out there. But, well, I've been working on UTW since January 3rd, and I know just how much work has gone into making it into the best version of itself it can be, and there have to be at least dozens of people out there on our planet who are all doing the same thing, and I'm going to be so thrilled when they burst onto this stage with me.
...But, uh, just in case the others have run into some snags along the way, I'm gonna move forward with UTW.
Of course, Catherine, I'm happy to share! There are a couple reasons, and the first is decidedly unglamorous: for UTW to have the amount of success it will need in order to make a change in the country, it needs to stand out and stick in people's minds. Most newsletters/podcasts avoid profanity in the title, because it is off-putting to a lot of people! I expect it'll harm my subscriber count that I've written it this way.
...but I also think that the people who don't subscribe will still remember the name. If the first time you see UTW is in a Facebook ad and you get the wrinkle-nose reaction, and the second time you see it is in a news article about a growing movement that might actually save the US, then maybe you'll think "this could really be a thing, huh?"
The other reason is the one I find more important, and it was why I came up with the title and the dialectic style in the first place: I don't want my readers passively absorbing content. I am offering hope, yes, but it is an active kind of hope, for people who want to make a difference, not for people who want to be told everything will be all right and we don't need to worry.
We do need to worry. We need to understand how serious things are. And so I open each episode with that profanity, as a subtle signal to my readers' and listeners' brains to "gird your loins, take this seriously, this is important". I occasionally do the same thing in the text of the article, for the same reason, but I try to do that sparingly because honestly, it makes me uncomfortable too. *Nobody* writes calm, uplifting, Fred Rogers-style prose but then laces it with curse words. (And oh, do I wish he were around now for our sake, but also I'm so glad he never had to see this.)
And if I write like nobody else in the world does, then maybe people will believe I also have an idea that no one else in the world does. That's my daydream, at least, and either I'm right and the world gets better for everyone from 2027 onward, or I'm wrong and I'm the only one who has invested anything of value, so I'm the only one who loses.
The problem is, the inmates like Andreessen and other bullish VCs are still running the asylum. People like you and Khanna know basically what needs to be done but this administratio is still in the pocket of the tech oligarchs! No chance we will get the control over AI that we need anytime soon sadly ☹️
You're absolutely right about where the control of our world currently lies!
I'm hoping I'll be able to convince you (and, you know, everyone) that the timeline of "getting control of our world back" can be much shorter than everyone currently believes—I'd love to count you as a subscriber, even if it's just so you can tell me "I told you so"! I expect we'll know by the end of launch week whether UTW has any legs or not. 👍
Exactly why we need an independent authority over AI that is not corrupt but also allows our domestic AI companies to compete effectively with China (and other nations)
There’s no such thing as a domestic AI company. Staff and IP from around the world, chips from Taiwan, fab from Europe, rare earths from Australia and Canada (and Russia and China and Africa). These are the months that future historians and economist will say this is when the post Bretonwoods period ended. The great advantage of having control of the financial system that makes all that possible, that makes the post war boom for R&D in the US possible, has been killed. The assumption that the US has the ability and will to keep the lines of trade open, that the USD and the US economy relies on, have been laid bare. The fact that one person without any oversight or transparency can shut off a service, and are willing to do so in such a manner, has given a sharp wake up call to every competent government and organisation around the world that digital electronic services emanating out of the US aer not actually protected by the institutional safeguards everyone assumed were there. History tells us it’s a downhill spiral from here until there is some kind of civilisational threatening event (but we can’t really do world wars any more to reset the economic, political and moral paradigms). It’s been fun while it lasted, and I think the uptake of distributed wind and solar versus centralised generation, is going to look like a gentle breeze compared to the windstorm of decentralised compute that’s about to break out. Gotta do something with all those electrons in the middle of the day, and let’s face it. There’s a lot of compute tasks that can be batched the one time for future programmatic recall rather than the chat bot experiments we’ve been doing to date. Saturday we were looking at a world like the day after like the day after the Berlin Wall came down.
Amazon owns a 21% stake in Anthropic. It does also own a stake of unknown size in OpenAI, but it is assuredly less than the 21% it owns in Anthropic (a currently more valuable company than OpenAI).
Unfortunately, to "lose favor with the White House" if you ever had it, one only has to (a) stand by your oath to the Constitution and to "the people," (b) do your job accordingly, (c) tell the truth (speak truth to power) (d) tend to your conscience; (e) don't be intimidated or bribed by powerful buffoons and people who own big oil and gas companies and (f) don't pay to play.
Chris Blue: I don't totally disagree with some of what you say, but most is way over generalized. For the record, your analysis is so full of logical fallacies that it's hard to read through your narratives here without tripping on them. And that's the thing about conspiracy theories--the always ride in on a little bit of truth, then expand on that as it's the whole story and everyone else is an idiot if they don't agree with you. No thanks.
the problem isn’t that limits arrived, it’s that they arrived through a process with no transparency and with wrong-looking motives. We do need regulation, that’s abundantly clear. But “grounded in technical facts” only means something if we’ve agreed on what the facts should be and right now the field has no shared way to measure what needs regulation.
The whole safety conversation is fixed on catastrophic misuse, while the everyday failures, confident fabrication, sycophantic drift under pressure, models agreeing with whatever the user wants, go essentially unmeasured. An independent agency is the right idea, but it’ll need benchmarks for reliability that don’t exist yet, or “technical facts” becomes whatever the room decides it means.
Ad hoc enforcement of "national security" concerns provides ample opportunities for grift and blackmail. Seems on-brand for this administration. Perhaps Anthropic just needed to make a donation to the "White House Ballroom Fund."
These models do not work, they do not exist and they are built from ARBITRARY sources: metaphors, words, symbols. It matters not if you can reduce them to symbolic mechanisms to make them more deterministic.
Frontier is nothing more than a charade that uses the symbolic to reach a more specific practice. Yet the same binary bottleneck of symbols arrests these models from ever gaining specificity.
So an arbitrary political response is in full parallel with the arbitrary nature of all existing AI.
The approach was totally in error, this is unstable compute that has nothing to do with thought, behavior, completion, intelligence, consciousness etc.
Nothing is going to change until there's a different administration. The grifting will just accelerate as the end of this term nears and the time to cash in shrinks. You know exactly why the US government made this opaque ruling. It is market manipulation, plain and simple. And when it works here with AI companies, the protection racket is only going to spread. Think about what the FDA could do to promote and suppress drug trials in favour of some companies and against others - just as a way to influence stock price.
The damage will outlast the administration. The Senate just confirmed a federal judge that the non-partisan Bar review board gave a rare Unqualified rating. She had never even been lead attorney in a court case. It is a lifetime appointment. And no doubt more to come.
Other news: In early June, Qatar's Investment Authority increased its investment in Anthropic in a $65 billion funding round, while yesterday, WaPo uncovered the fact that Qatar has been secretly back channeling with Iran for months to avoid getting bombed.
The US bombed Iranian water reservoirs recently. 20,000 Iranians out of water.
If Iran bombs Qatar desalination plants in retaliation, Qatar becomes unlivable. The same applies to the other Gulf States and Israel.
Qatar and the other states now know the US can't protect them, so the US protection racket is over. Best make an accommodation with Iran because Iran is the new fourth world power (lower than the top three but still a regional power.)
I cannot say I blame Qatar. Who wants to get bombed? Seems they are stuck in a hard place between the US and Iran and they have the money to comfortably afford payments to both.
Chris Blue: I did not say your notes were 100% false. I said your notes are full of logical fallacies. (If you don't know what that means, look it up.)
I usually agree with you, Gary, but this time, I read the situation differently. (I have worked in software security and AI, so I know a lot about both.) I previously linked to this Anthropic blog post: https://red.anthropic.com/2026/n-days/ It's long, but here's the money quote: "Mythos Preview is a step change improvement in turning a crash into a full exploit." This makes the model significantly more dangerous, in the wrong hands, to software security, than any previous model.
Of course Anthropic have worked, probably very hard, to train the released model (Fable) not to respond to requests to create exploits; and of course, LLMs being LLMs, these guardrails are not perfect. In my view, Commerce has done the right thing here, at least provisionally, and yes, it was an emergency.
I take David Sacks's tweet mostly at face value. What he and the people he has spoken to don't understand, though, is that Anthropic CAN'T just fix the problem. LLMs don't work that way. Anthropic can continue to make it harder to jailbreak the model, but they probably can't ever completely prevent it — if the information is in the weights, there are going to be ways to get it out.
In my view, this is another black mark on LLMs' report card (I wanted to say "nail in their coffin" but alas, they aren't dead yet). And it's one that jibes very well with your larger thesis. I strongly recommend you rethink your position on this.
The only reason Mythos is dangerous to software security is because software development practices are so poor industry-wide.
What LLMs have exposed is the crap quality of the software industry, where craft practices are referred to as "engineering" while violating every principle of engineering.
As I said in a previous post in this thread, there are no "too dangerous to release" models. Every model can do what Mythos does, if used by someone familiar with hacking. That has been proven by several organizations studying exactly that. The only advantage Mythos has is that it's faster because it's running on huge compute engines.
No, this is either another stupid Trump administration move or another Trump administration corruption move. Or both.
That this damages Anthropic - the one company that makes ALL its PR about how their models are "conscious, have emotions, have a soul, are dangerous" BS - is icing on the cake for me. I despise Anthropic, I despise Dario Amodei for being a huckster and a fraud, and I would like to see Anthropic sued for fraud.
So pass the popcorn. But personally I'm more interested in the Iran war and its effects on the AI bubble than this nonsense.
I agree about the state of the software industry, but:
> Every model can do what Mythos does, if used by someone familiar with hacking.
Perhaps, with enough effort, one can get an earlier model to generate the same exploits. But it will take longer, and many of the attempts will fail and have to be retried, perhaps several times. Sounds to me as if Mythos will bring the token cost of an exploit down by a factor of 3 to 10, maybe. I don't think that democratizing cybercrime is on Anthropic's mission statement.
There appears to be an argument that a frontier model can 1) scan a large codebase better than a small model, and 2) chain a series of vulnerabilities together to make an exploit better than a small model.
The opposing argument is that this is not how hackers do it. Hackers don't scan an entire codebase. They work the system and try what has worked before using general principles of known software mistakes in the past. In edge cases, they hypothesize an exploit and look for evidence that it exists in the code which necessarily means restricting examination to likely areas of the code.
A hacker using a local model can do exactly the same thing. The local model makes it just as fast a process as using a frontier model to run over the entire codebase - and at a cheaper cost, most likely, given the cheaper hardware.
It also appears that the harness is as important as the size of the model, although again, when scanning an entire large codebase, compute capability does have an edge.
The point is that Mythos may be better at generating long-chain exploits out of a large codebase, but many other, even much smaller, models are fully capable of producing exploits when used by competent hackers.
As the report above concludes:
QUOTE:
Even if Mythos is never released, the public-model results are already significant on their own. The models are past the “just crash it” stage: several can build target-specific primitives, and GPT-5.5 crossed the line into full code execution on a small number of cases. That puts public models in an important middle ground: not replacements for elite exploit developers, but increasingly useful accelerators for skilled researchers working through the exploitation ladder on hard targets.
Although ExploitBench does not measure easier targets, the results strongly suggest they are likely already capable of full exploitation on less hardened software with fewer layers of defense. That does not mean they can exploit everything. It means that public models have moved into real exploit-development on their own, or as a powerful assistance tool for practitioners.
END QUOTE
As for Anthropic's mission statement - their goal was to get in with government. They explicitly developed Mythos to max out the benchmarks for cybersecurity capability. It wasn't just an "emergent property" of their models. They'll deny it, of course, and absent internal Anthropic documents I can't prove it. But after the Pentagon fiasco, it seems highly likely that this was deliberate.
And it worked - Anthropic is now embedded in the NSA, even going to far as to deploy its own engineers to the NSA to assist in the implementation.
I should have added: I do agree that going forward, an independent agency and a transparent process will be very important. And if you want to criticize the administration for not having already set all that up, I'm with you. But since they haven't, I think this action was necessary, for the moment anyway.
I'm not defending Anthropic. And I agree that the risks need to be studied more carefully, and might turn out to have been overstated. But reaching that conclusion will take time. The question is whether, given what the Commerce Department knows today, the apparent risks justify this emergency action.
I think they do. Mythos/Fable seems to be more than "marginally" better than previous models at developing full exploits. I know you don't think so, but think about the mess we could be in if you're wrong. A little caution seems to me to be in order here.
The brain drain point cuts the other way too. Europe spent three years building the most comprehensive AI regulatory framework on earth and has exactly one frontier lab to show for it, propped up by Macron personally. Now the US is handing every European "sovereign AI" advocate the best argument they've ever had.
The problem is Brussels will respond to Friday with a position paper. GDPR took four years from proposal to enforcement. The AI Act took three. AI moves in quarters. The window where "sovereign AI" could have meant something closed while Europe was still writing rules for American products.
Khanna's independent agency is the right call for the US. Europe doesn't need another agency. It needs to build something before there's nothing left to regulate.
The other option that Europe has is to stay away from the current genAI/LLM craziness and fund some more useful research. Current genAI is unlikely to have economic impact proportional to the amount of investments. There are some good use cases for the current genAI, but it is unlikely to 10x GDP or even replace many workers.
Naked, gangster capitalism, a return to an earlier age and with no one to stop it or even have a clue about the implications of AI's effect on economics, politics, culture, on reality itself. It's a bit like an unstoppable virus has been let loose on society and its now eating its way through the fabric that holds it all together. And the government? What government? It's busy waging war on the planet and making as much money for the 1% as it can, while it can. Sociopathology writ large.
William Bowles: I don't think anything good can or will happen (barring a miracle) if the present administration keeps any of its present power. And I haven't stopped head-slapping over the state of the John Roberts Court and some of his incredibly ridiculous statements and positions (talk about "out of touch"). I do have hope (if not "optimism") about things that ARE doable--the institution still exists--and is not "dead in the water," so to speak; but those involve still must actual do it.
I keep waiting for Trump to insert himself into the World Cup, if he hasn't already set it up. But if you want to get a sense of optimism, even if it's temporary, watch the camera scans over the crowds at the games, and the players, coaches, and referees. Their enthusiasm and happiness are "catching." It's a wonderful thing to see. Who cares who wins. And how about those knicks?
Ah, the Knicks. As a former New Yorker but not a b-ball follower, I'm aware of the love/hate relationship between the team and Nuyorkers, in fact I read a piece yesterday about the team, that it finally, actually won!! I too have a love/hate relationship with Nuevayork, I lived there for 17 years and had an amazing, creative life in the city.
William Bowles: I think anyone who still has 'the disease' of CLAIMing to believe Trump really won the election should be asked to explain themselves and all the evidence to the contrary--every day.
An addendum: I don't know how Trump can keep expecting people to like him and seeing his face plastered on everything when he has made himself into the monstrous male equivalent of a mean girl.
Mythos is a myth, nothing mythical about it . Every software engineer worth his salt knows what's up, AND the markets are starting to catch up, so this pathetic lying administration jumps onto helping big tech bros with this Hollywoodic narrative of "national security risk" to hype up the stocks. I'm sure China is enjoying watching America in self destructive mode
It's only mythical, ie mythological thought. AI is software causality which has nothing to do with real events, thought, intelligence, completion. Reality and thought are massive, parallel oscillations. There's nothing clockwork, linear, compute in reality.
This is false, unstable technology that will never offer a step beyond optimization.
There are no adults in the White House. Worse, the vanishingly few adults in Congress and the Supreme Court have nowhere near the numbers they need to be able to regain control of either branch, regardless of electoral outcome.
Unless we step outside our usual role of "vote every two years", this nightmare will never end.
Danielle Church: I have to say, while I'm glad you are "doing something," the use of that four-letter-word, as a comprehensive/symbolic title especially, is and always will be to me, extremely off-putting. (I slip and use it myself sometimes but always regret it.) Is there something I'm not seeing here? If so, clue me in so that I can understand?
This reminds me of the issue of the translation problem from yesterday. "Four letter word, = "unfucking," really? Looks like 9 letters to me.
And yet, you still knew which word she was talking about! And I knew that she wasn't talking about that particular conjugation of the word, but about all of its forms. So it seems to me like she used the ideal communication for the context... is there something I've missed?
Danielle Church: I won't die from it in any version. However, and thank you for explaining your usage of that word (regardless of how many other letters surround it). To Joy: It stands out as a "four-letter word" precisely for the same reasons Daniel wants to use it as introductory.
The other thought is that I certainly don't want to "doll up" anything that has to do with Trump and Company. Fxxk works in that regard and at least the lowlife that inhabits at least some of MAGA will love it.
The "on the other hand" idea, however, (and you most likely knew that was coming next) is that some just won't take it seriously or even think it's too crass to identify with (even though I think it beats "disambiguate").
Enough . . . and though the method of communication is a significant part of any idea worth moving along, let's not make it any more of a diversion. In this case, like no other, doing nothing is not an option.
Exactly, exactly, exactly right, Catherine. Doing nothing is not an option.
I made the choice that I want to give my efforts the chance for the most emotive power and force they possibly can, even though I know I'm giving up a lot on the "first impression" front, and that my subscriber count just isn't going to climb as fast as it would if I'd gone with a cleaner title.
This world works best when we all stick to our strengths, though, and mine are an irreverent but extremely caring humanist warmth and a complete refusal to compromise in my vision of what "success" means. If what our world needs is a different kind of voice to lead it out of this darkness, then I am so okay with not being that voice! (So, so very okay. I am terrified of being a public figure.) I wouldn't be nearly as good at anything except for being the person that I truly am, and my patriotism leans much closer to "fierce" than "proud".
In my ideal world, I am one of many, or at least a handful, of voices all offering us true, actionable paths to restore harmony to our world. I am doing everything I can to make my effort the most quintessentially "Danielle Church" path that it can be, so that once we're all standing on this flaming stage of an Internet, the public will be able to choose the one that offers the best chance of success for the most people.
I know they must be out there. But, well, I've been working on UTW since January 3rd, and I know just how much work has gone into making it into the best version of itself it can be, and there have to be at least dozens of people out there on our planet who are all doing the same thing, and I'm going to be so thrilled when they burst onto this stage with me.
...But, uh, just in case the others have run into some snags along the way, I'm gonna move forward with UTW.
Of course, Catherine, I'm happy to share! There are a couple reasons, and the first is decidedly unglamorous: for UTW to have the amount of success it will need in order to make a change in the country, it needs to stand out and stick in people's minds. Most newsletters/podcasts avoid profanity in the title, because it is off-putting to a lot of people! I expect it'll harm my subscriber count that I've written it this way.
...but I also think that the people who don't subscribe will still remember the name. If the first time you see UTW is in a Facebook ad and you get the wrinkle-nose reaction, and the second time you see it is in a news article about a growing movement that might actually save the US, then maybe you'll think "this could really be a thing, huh?"
The other reason is the one I find more important, and it was why I came up with the title and the dialectic style in the first place: I don't want my readers passively absorbing content. I am offering hope, yes, but it is an active kind of hope, for people who want to make a difference, not for people who want to be told everything will be all right and we don't need to worry.
We do need to worry. We need to understand how serious things are. And so I open each episode with that profanity, as a subtle signal to my readers' and listeners' brains to "gird your loins, take this seriously, this is important". I occasionally do the same thing in the text of the article, for the same reason, but I try to do that sparingly because honestly, it makes me uncomfortable too. *Nobody* writes calm, uplifting, Fred Rogers-style prose but then laces it with curse words. (And oh, do I wish he were around now for our sake, but also I'm so glad he never had to see this.)
And if I write like nobody else in the world does, then maybe people will believe I also have an idea that no one else in the world does. That's my daydream, at least, and either I'm right and the world gets better for everyone from 2027 onward, or I'm wrong and I'm the only one who has invested anything of value, so I'm the only one who loses.
The problem is, the inmates like Andreessen and other bullish VCs are still running the asylum. People like you and Khanna know basically what needs to be done but this administratio is still in the pocket of the tech oligarchs! No chance we will get the control over AI that we need anytime soon sadly ☹️
You're absolutely right about where the control of our world currently lies!
I'm hoping I'll be able to convince you (and, you know, everyone) that the timeline of "getting control of our world back" can be much shorter than everyone currently believes—I'd love to count you as a subscriber, even if it's just so you can tell me "I told you so"! I expect we'll know by the end of launch week whether UTW has any legs or not. 👍
Exactly why we need an independent authority over AI that is not corrupt but also allows our domestic AI companies to compete effectively with China (and other nations)
There’s no such thing as a domestic AI company. Staff and IP from around the world, chips from Taiwan, fab from Europe, rare earths from Australia and Canada (and Russia and China and Africa). These are the months that future historians and economist will say this is when the post Bretonwoods period ended. The great advantage of having control of the financial system that makes all that possible, that makes the post war boom for R&D in the US possible, has been killed. The assumption that the US has the ability and will to keep the lines of trade open, that the USD and the US economy relies on, have been laid bare. The fact that one person without any oversight or transparency can shut off a service, and are willing to do so in such a manner, has given a sharp wake up call to every competent government and organisation around the world that digital electronic services emanating out of the US aer not actually protected by the institutional safeguards everyone assumed were there. History tells us it’s a downhill spiral from here until there is some kind of civilisational threatening event (but we can’t really do world wars any more to reset the economic, political and moral paradigms). It’s been fun while it lasted, and I think the uptake of distributed wind and solar versus centralised generation, is going to look like a gentle breeze compared to the windstorm of decentralised compute that’s about to break out. Gotta do something with all those electrons in the middle of the day, and let’s face it. There’s a lot of compute tasks that can be batched the one time for future programmatic recall rather than the chat bot experiments we’ve been doing to date. Saturday we were looking at a world like the day after like the day after the Berlin Wall came down.
Amazon owns a 21% stake in Anthropic. It does also own a stake of unknown size in OpenAI, but it is assuredly less than the 21% it owns in Anthropic (a currently more valuable company than OpenAI).
Max: I vote that no one with shares or with family and friends with shares, gets ANY access to an "independent" agency.
Unfortunately, to "lose favor with the White House" if you ever had it, one only has to (a) stand by your oath to the Constitution and to "the people," (b) do your job accordingly, (c) tell the truth (speak truth to power) (d) tend to your conscience; (e) don't be intimidated or bribed by powerful buffoons and people who own big oil and gas companies and (f) don't pay to play.
Chris Blue: I don't totally disagree with some of what you say, but most is way over generalized. For the record, your analysis is so full of logical fallacies that it's hard to read through your narratives here without tripping on them. And that's the thing about conspiracy theories--the always ride in on a little bit of truth, then expand on that as it's the whole story and everyone else is an idiot if they don't agree with you. No thanks.
the problem isn’t that limits arrived, it’s that they arrived through a process with no transparency and with wrong-looking motives. We do need regulation, that’s abundantly clear. But “grounded in technical facts” only means something if we’ve agreed on what the facts should be and right now the field has no shared way to measure what needs regulation.
The whole safety conversation is fixed on catastrophic misuse, while the everyday failures, confident fabrication, sycophantic drift under pressure, models agreeing with whatever the user wants, go essentially unmeasured. An independent agency is the right idea, but it’ll need benchmarks for reliability that don’t exist yet, or “technical facts” becomes whatever the room decides it means.
Ad hoc enforcement of "national security" concerns provides ample opportunities for grift and blackmail. Seems on-brand for this administration. Perhaps Anthropic just needed to make a donation to the "White House Ballroom Fund."
The dumpster fire of AI will only spiral.
These models do not work, they do not exist and they are built from ARBITRARY sources: metaphors, words, symbols. It matters not if you can reduce them to symbolic mechanisms to make them more deterministic.
Frontier is nothing more than a charade that uses the symbolic to reach a more specific practice. Yet the same binary bottleneck of symbols arrests these models from ever gaining specificity.
So an arbitrary political response is in full parallel with the arbitrary nature of all existing AI.
The approach was totally in error, this is unstable compute that has nothing to do with thought, behavior, completion, intelligence, consciousness etc.
Nothing is going to change until there's a different administration. The grifting will just accelerate as the end of this term nears and the time to cash in shrinks. You know exactly why the US government made this opaque ruling. It is market manipulation, plain and simple. And when it works here with AI companies, the protection racket is only going to spread. Think about what the FDA could do to promote and suppress drug trials in favour of some companies and against others - just as a way to influence stock price.
The damage will outlast the administration. The Senate just confirmed a federal judge that the non-partisan Bar review board gave a rare Unqualified rating. She had never even been lead attorney in a court case. It is a lifetime appointment. And no doubt more to come.
yeah. too bad kamala isn't in the WH.
The White House did something corrupt?? i'm shocked
https://www.youtube.com/watch?v=HMIyDf3gBoY
ihor Gowda; The White House tried to get Jeffrey Epstein to run an Ethics Committee, but he was unavailable at the time. (Love that clip.)
Other news: In early June, Qatar's Investment Authority increased its investment in Anthropic in a $65 billion funding round, while yesterday, WaPo uncovered the fact that Qatar has been secretly back channeling with Iran for months to avoid getting bombed.
The US bombed Iranian water reservoirs recently. 20,000 Iranians out of water.
If Iran bombs Qatar desalination plants in retaliation, Qatar becomes unlivable. The same applies to the other Gulf States and Israel.
Qatar and the other states now know the US can't protect them, so the US protection racket is over. Best make an accommodation with Iran because Iran is the new fourth world power (lower than the top three but still a regional power.)
Smart of Qatar.
I cannot say I blame Qatar. Who wants to get bombed? Seems they are stuck in a hard place between the US and Iran and they have the money to comfortably afford payments to both.
Chris Blue: I did not say your notes were 100% false. I said your notes are full of logical fallacies. (If you don't know what that means, look it up.)
I usually agree with you, Gary, but this time, I read the situation differently. (I have worked in software security and AI, so I know a lot about both.) I previously linked to this Anthropic blog post: https://red.anthropic.com/2026/n-days/ It's long, but here's the money quote: "Mythos Preview is a step change improvement in turning a crash into a full exploit." This makes the model significantly more dangerous, in the wrong hands, to software security, than any previous model.
Of course Anthropic have worked, probably very hard, to train the released model (Fable) not to respond to requests to create exploits; and of course, LLMs being LLMs, these guardrails are not perfect. In my view, Commerce has done the right thing here, at least provisionally, and yes, it was an emergency.
I take David Sacks's tweet mostly at face value. What he and the people he has spoken to don't understand, though, is that Anthropic CAN'T just fix the problem. LLMs don't work that way. Anthropic can continue to make it harder to jailbreak the model, but they probably can't ever completely prevent it — if the information is in the weights, there are going to be ways to get it out.
In my view, this is another black mark on LLMs' report card (I wanted to say "nail in their coffin" but alas, they aren't dead yet). And it's one that jibes very well with your larger thesis. I strongly recommend you rethink your position on this.
The only reason Mythos is dangerous to software security is because software development practices are so poor industry-wide.
What LLMs have exposed is the crap quality of the software industry, where craft practices are referred to as "engineering" while violating every principle of engineering.
As I said in a previous post in this thread, there are no "too dangerous to release" models. Every model can do what Mythos does, if used by someone familiar with hacking. That has been proven by several organizations studying exactly that. The only advantage Mythos has is that it's faster because it's running on huge compute engines.
No, this is either another stupid Trump administration move or another Trump administration corruption move. Or both.
That this damages Anthropic - the one company that makes ALL its PR about how their models are "conscious, have emotions, have a soul, are dangerous" BS - is icing on the cake for me. I despise Anthropic, I despise Dario Amodei for being a huckster and a fraud, and I would like to see Anthropic sued for fraud.
So pass the popcorn. But personally I'm more interested in the Iran war and its effects on the AI bubble than this nonsense.
I agree about the state of the software industry, but:
> Every model can do what Mythos does, if used by someone familiar with hacking.
Perhaps, with enough effort, one can get an earlier model to generate the same exploits. But it will take longer, and many of the attempts will fail and have to be retried, perhaps several times. Sounds to me as if Mythos will bring the token cost of an exploit down by a factor of 3 to 10, maybe. I don't think that democratizing cybercrime is on Anthropic's mission statement.
See, as an example:
System Over Model, Tested: Reproducing Mythos's FreeBSD Find on Local Open-Weight Models
https://clearbluejar.github.io/posts/system-over-model-tested-mythos-freebsd-local-openweight/
Also, see here:
AI benchmarking report: Measuring the exploitation ladder for AI models
https://www.bugcrowd.com/blog/ai-benchmarking-report-measuring-the-exploitation-ladder-for-ai-models/
There appears to be an argument that a frontier model can 1) scan a large codebase better than a small model, and 2) chain a series of vulnerabilities together to make an exploit better than a small model.
The opposing argument is that this is not how hackers do it. Hackers don't scan an entire codebase. They work the system and try what has worked before using general principles of known software mistakes in the past. In edge cases, they hypothesize an exploit and look for evidence that it exists in the code which necessarily means restricting examination to likely areas of the code.
A hacker using a local model can do exactly the same thing. The local model makes it just as fast a process as using a frontier model to run over the entire codebase - and at a cheaper cost, most likely, given the cheaper hardware.
It also appears that the harness is as important as the size of the model, although again, when scanning an entire large codebase, compute capability does have an edge.
The point is that Mythos may be better at generating long-chain exploits out of a large codebase, but many other, even much smaller, models are fully capable of producing exploits when used by competent hackers.
As the report above concludes:
QUOTE:
Even if Mythos is never released, the public-model results are already significant on their own. The models are past the “just crash it” stage: several can build target-specific primitives, and GPT-5.5 crossed the line into full code execution on a small number of cases. That puts public models in an important middle ground: not replacements for elite exploit developers, but increasingly useful accelerators for skilled researchers working through the exploitation ladder on hard targets.
Although ExploitBench does not measure easier targets, the results strongly suggest they are likely already capable of full exploitation on less hardened software with fewer layers of defense. That does not mean they can exploit everything. It means that public models have moved into real exploit-development on their own, or as a powerful assistance tool for practitioners.
END QUOTE
As for Anthropic's mission statement - their goal was to get in with government. They explicitly developed Mythos to max out the benchmarks for cybersecurity capability. It wasn't just an "emergent property" of their models. They'll deny it, of course, and absent internal Anthropic documents I can't prove it. But after the Pentagon fiasco, it seems highly likely that this was deliberate.
And it worked - Anthropic is now embedded in the NSA, even going to far as to deploy its own engineers to the NSA to assist in the implementation.
That is not a "safety first" company.
I should have added: I do agree that going forward, an independent agency and a transparent process will be very important. And if you want to criticize the administration for not having already set all that up, I'm with you. But since they haven't, I think this action was necessary, for the moment anyway.
Did you read the blog post?
I'm not defending Anthropic. And I agree that the risks need to be studied more carefully, and might turn out to have been overstated. But reaching that conclusion will take time. The question is whether, given what the Commerce Department knows today, the apparent risks justify this emergency action.
I think they do. Mythos/Fable seems to be more than "marginally" better than previous models at developing full exploits. I know you don't think so, but think about the mess we could be in if you're wrong. A little caution seems to me to be in order here.
The brain drain point cuts the other way too. Europe spent three years building the most comprehensive AI regulatory framework on earth and has exactly one frontier lab to show for it, propped up by Macron personally. Now the US is handing every European "sovereign AI" advocate the best argument they've ever had.
The problem is Brussels will respond to Friday with a position paper. GDPR took four years from proposal to enforcement. The AI Act took three. AI moves in quarters. The window where "sovereign AI" could have meant something closed while Europe was still writing rules for American products.
Khanna's independent agency is the right call for the US. Europe doesn't need another agency. It needs to build something before there's nothing left to regulate.
The other option that Europe has is to stay away from the current genAI/LLM craziness and fund some more useful research. Current genAI is unlikely to have economic impact proportional to the amount of investments. There are some good use cases for the current genAI, but it is unlikely to 10x GDP or even replace many workers.
Shoot first, ask questions later. The USA has literally returned to the Wild West. And there is no honest sheriff in sight. The bandits are in charge.
Where are Sheriff Wyatt AI-rp and Doc Hollid-AI?
Returning Law to the Wild West of AI
Naked, gangster capitalism, a return to an earlier age and with no one to stop it or even have a clue about the implications of AI's effect on economics, politics, culture, on reality itself. It's a bit like an unstoppable virus has been let loose on society and its now eating its way through the fabric that holds it all together. And the government? What government? It's busy waging war on the planet and making as much money for the 1% as it can, while it can. Sociopathology writ large.
William Bowles: I hear that . . . except that the voters CAN stop it, and perhaps almost at the finish line? WE will.
I wish I had your optimism.
William Bowles: I don't think anything good can or will happen (barring a miracle) if the present administration keeps any of its present power. And I haven't stopped head-slapping over the state of the John Roberts Court and some of his incredibly ridiculous statements and positions (talk about "out of touch"). I do have hope (if not "optimism") about things that ARE doable--the institution still exists--and is not "dead in the water," so to speak; but those involve still must actual do it.
I keep waiting for Trump to insert himself into the World Cup, if he hasn't already set it up. But if you want to get a sense of optimism, even if it's temporary, watch the camera scans over the crowds at the games, and the players, coaches, and referees. Their enthusiasm and happiness are "catching." It's a wonderful thing to see. Who cares who wins. And how about those knicks?
1. DHS Docs: Govt bracing for Nationwide anti-AI riots, preparing to crack down on dissent
Alan MacLeod
https://www.mintpressnews.com/dhs-docs-govt-bracing-for-nationwide-anti-ai-riots-preparing-to-crack-down-on-dissent/290985/
2. Trump's Rolling Coup
And what must be done to stop it.
https://robertreich.substack.com/p/warning-trumps-rolling-coup-is-underway
Ah, the Knicks. As a former New Yorker but not a b-ball follower, I'm aware of the love/hate relationship between the team and Nuyorkers, in fact I read a piece yesterday about the team, that it finally, actually won!! I too have a love/hate relationship with Nuevayork, I lived there for 17 years and had an amazing, creative life in the city.
William Bowles: I think anyone who still has 'the disease' of CLAIMing to believe Trump really won the election should be asked to explain themselves and all the evidence to the contrary--every day.
An addendum: I don't know how Trump can keep expecting people to like him and seeing his face plastered on everything when he has made himself into the monstrous male equivalent of a mean girl.
The current administration is a clear and present danger to the whole field of AI.
If only the danger was limited to AI.
If only…