> So many comments here missing the big picture, and just gleefully pointing out that Anthropic got what they deserved, or that this is the natural culmination of some kind of marketing stunt.
They are not wrong, it feels like that Game of Thrones season where someone thought it would be a great idea to let the fanatics re-arm.
> The real story here is that this may be the beginning of governments restricting the availability of strong LLMs to the public, to you. Fable was the strongest model on the market, and the US government has told you you can’t use it (technically, only if you’re not a US citizen, but in practice, even if you are). If you think the solution here is going to be open source Chinese models and / or running on your own hardware, think again. Do you think China is going to allow the strongest LLMs from companies within its borders to be open source a year from now when they have Mythos capabilities, if the US government is keeping the strongest American models back? Unlikely. These are heading in the direction of being powerful cybersecurity weapons and it will be in the interest of nation states to restrict and control them. In 2 years time, I would be surprised if the strongest LLMs are available for general use at all.
The world is a bit bigger than US and China, if Anthropic did it, another company can do it as well.
I am highly skeptical about Mythos's part in the whole cyber security angle and Anthropic seems to agree with me:
> We have reviewed a report that we believe is the basis of the government's directive and validated that the level of capability displayed there is widely available from other models (including OpenAI’s GPT-5.5)
It does sound funny to hear this from Anthropic after they spent recent months with scaremongering about Mythos's capabilities, now they say it was a prank bro, you can actually achieve more or less the same with good old GPT-5.5.
> Will we be the poorer for that, or will we be safer? I think poorer, because I hate being told what technology I can and can’t use, but I’m not certain. Maybe you think the government should restrict strong LLMs. Maybe you don’t. But either way, this is big news and a rubicon has been crossed and a precedent set. That’s true even if the motivation for this is just the government settling scores with Anthropic.
What this has demonstrated: if you can't run the software on your own hardware, you should assume that it can be taken away at any moment.
Meanwhile the world keeps spinning and most people don't even know what Anthropic is, much less anything about Fable.
If AI lived up to a tenth of the promises the American labs produce, the world would be drastically different today. It's not. I'm doubtful of future impact based on that.
I'm happy we can utilise current OSS models to the extent we can now. They'll improve. The world will continue as usual. And hopefully we can put this bubble behind us.
lol if this is an attempt by the admin like the DoD thing to "knock them down a peg" it actually has the opposite effect for me, showing that anthropic is that far ahead you have to resort to dirty tricks.
> As a brief aside, I am once again extremely disappointed in the myriad of Silicon Valley people who angrily argued that a Democratic led government would ‘pick winners and losers in the AI race’ are now completely silent or defending the actions of this admin. I cannot help but feel that that previous posturing was just a machiavellian play for power, which has just been the worst feeling in the world.
I mean, yeah. But did it take this long for that to be apparent to you?
I really hope it's just the USA punishing Anthropic for their insolence. If this is actually the beginning of AI regulation, we're probably heading towards dark times.
> I actually have another draft post in the barrel about how I think we should see a resurgence of the ‘flash game’ renaissance because it has become so much easier to make fun little games with AI tooling.
I have been lurking on the aigamedev subreddit to see exactly what sort of games people are coming up with and I can say I have been incredibly disappointing. I've been faithfully trying the games people post and have come to the conclusion that game design is a very difficult art to learn, and something LLMs really can't help with that much. My guess is that these games are "fun" just like toddler paintings are "beautiful." And there are so many quality indie games you could get for the 25+ dollars you'd spend generating the code. Anyways, I guess that's another discussion for another blog post.
I completely agree. The #1 thing you need as a game designer is taste. AI's doesn't have that. It will happily take your low-quality-seen-many-times-before game design idea and generate code for a low-quality-seen-many-times-before game.
I wonder if you could prompt for configurable values "knobs" like:
"Let's make a super mario 2D sidescrolling clone. I want the impulse, angular velocity, and gravity influencing the feeling of the jump to be configurable"
So you essentially design the parts that require detail while building up a custom engine
I'm hearing a lot of this kind of thing. "Oh if only it was a different government". I'm sorry, but when you cry out for government involvement, it's not always going to be coming from the government you personally wanted. This is the whole problem with government involvement! I don't think that message is getting through, but it's the real lesson that should be learned here.
with how the admin is talking about taking a stake in openai, it's so incredibly clear this is the government attempting to kneecap an openai competitor
It stinks to high heaven, especially considering how over-the-top security protocols were introduced with Fable. The US government is asserting its influence on the economy and showing Anthropic that their IPO will depend on bending the knee.
I seriously feel like there's easier ways for OpenAI to catch up to anthropic and it would be a waste of political capital that the idea of Sam pulling strings for this to happen seems highly unlikely
While this is regrettable the guardrails were rather sloppy and I managed to do things with Fable that really should not be possible. It seems with all the focus cyber and bio security, threat scenario analysis went out the door. I guess they will fix the guardrails and then open it up again.
> OpenAI did the same “too dangerous to release” song and dance for the awesome, world ending AI that was GPT-2.
Wasn't that when Dario, et al were at the company. One way to view this is that OpenAI expelled the cultists and they went on to form their own organization that continued using the same tactics.
Certainly some of the Anthropic press around Fable seems to me to be just marketing but I also think there's a core of people there who really believe it. I also think like all good advertising/lies there's some truth to the claims even if they're exaggerating.
Nice summary. Reading this reminds me about the strong encryption discussion.
> We optimize what we can measure, not what we actually want to achieve. We hope and pray that these are the same thing, but they often aren’t.
He points out the core problem with LLMs. I believe it is impossible (or extremely expensive) to ensure that the models are aligned safely for everyone and any intention. And 'safe' can mean different things for a different audience.
Counterintuitively, this is a huge win for misAnthropic and other closed labs in the US. They can nerf the models, ask for IDs from users and do what it takes to comply with whatever regulation they've been fighting for.
Foreign labs releasing open source models won't be able to comply, and as a result open source models will remain stunted at pre-mythos levels or their use will be criminalized.
We should look past the petty fights these closed labs have, and see their common interest in banning open source and/or local models.
67 comments
[ 2.8 ms ] story [ 42.4 ms ] threadThey are not wrong, it feels like that Game of Thrones season where someone thought it would be a great idea to let the fanatics re-arm.
> The real story here is that this may be the beginning of governments restricting the availability of strong LLMs to the public, to you. Fable was the strongest model on the market, and the US government has told you you can’t use it (technically, only if you’re not a US citizen, but in practice, even if you are). If you think the solution here is going to be open source Chinese models and / or running on your own hardware, think again. Do you think China is going to allow the strongest LLMs from companies within its borders to be open source a year from now when they have Mythos capabilities, if the US government is keeping the strongest American models back? Unlikely. These are heading in the direction of being powerful cybersecurity weapons and it will be in the interest of nation states to restrict and control them. In 2 years time, I would be surprised if the strongest LLMs are available for general use at all.
The world is a bit bigger than US and China, if Anthropic did it, another company can do it as well.
I am highly skeptical about Mythos's part in the whole cyber security angle and Anthropic seems to agree with me:
> We have reviewed a report that we believe is the basis of the government's directive and validated that the level of capability displayed there is widely available from other models (including OpenAI’s GPT-5.5)
It does sound funny to hear this from Anthropic after they spent recent months with scaremongering about Mythos's capabilities, now they say it was a prank bro, you can actually achieve more or less the same with good old GPT-5.5.
> Will we be the poorer for that, or will we be safer? I think poorer, because I hate being told what technology I can and can’t use, but I’m not certain. Maybe you think the government should restrict strong LLMs. Maybe you don’t. But either way, this is big news and a rubicon has been crossed and a precedent set. That’s true even if the motivation for this is just the government settling scores with Anthropic.
What this has demonstrated: if you can't run the software on your own hardware, you should assume that it can be taken away at any moment.
If AI lived up to a tenth of the promises the American labs produce, the world would be drastically different today. It's not. I'm doubtful of future impact based on that.
I'm happy we can utilise current OSS models to the extent we can now. They'll improve. The world will continue as usual. And hopefully we can put this bubble behind us.
I mean, yeah. But did it take this long for that to be apparent to you?
I have been lurking on the aigamedev subreddit to see exactly what sort of games people are coming up with and I can say I have been incredibly disappointing. I've been faithfully trying the games people post and have come to the conclusion that game design is a very difficult art to learn, and something LLMs really can't help with that much. My guess is that these games are "fun" just like toddler paintings are "beautiful." And there are so many quality indie games you could get for the 25+ dollars you'd spend generating the code. Anyways, I guess that's another discussion for another blog post.
"Let's make a super mario 2D sidescrolling clone. I want the impulse, angular velocity, and gravity influencing the feeling of the jump to be configurable"
So you essentially design the parts that require detail while building up a custom engine
I'm hearing a lot of this kind of thing. "Oh if only it was a different government". I'm sorry, but when you cry out for government involvement, it's not always going to be coming from the government you personally wanted. This is the whole problem with government involvement! I don't think that message is getting through, but it's the real lesson that should be learned here.
But at this time Dario was at OpenAI and was a co-author on the GPT-2 research paper announcing the model.
The "too dangerous to release" approach has been him the whole time, at both companies.
Wasn't that when Dario, et al were at the company. One way to view this is that OpenAI expelled the cultists and they went on to form their own organization that continued using the same tactics.
Certainly some of the Anthropic press around Fable seems to me to be just marketing but I also think there's a core of people there who really believe it. I also think like all good advertising/lies there's some truth to the claims even if they're exaggerating.
> We optimize what we can measure, not what we actually want to achieve. We hope and pray that these are the same thing, but they often aren’t.
He points out the core problem with LLMs. I believe it is impossible (or extremely expensive) to ensure that the models are aligned safely for everyone and any intention. And 'safe' can mean different things for a different audience.
Foreign labs releasing open source models won't be able to comply, and as a result open source models will remain stunted at pre-mythos levels or their use will be criminalized.
We should look past the petty fights these closed labs have, and see their common interest in banning open source and/or local models.