Anthropic's Safety Superpower
stratechery.com
stratechery.com
The bottleneck is compute and data, not the model. That's why they could only gate it for a bit. The ITAR thing proves it: no nationality controls in place, so the only option was killing the whole thing. Not exactly what an all-powerful gatekeeper does.
Not for now, but how long before we have KYC regulations concerning LLMs?
But I also think they exist in a sort of un-designed corporate narcissism, which is a common trait in bubble economies — I am not judging them particularly severely.
Netscape under Clark and Andreessen and Sun under McNealy both fell into corporate narcissism: the belief that only they really mattered, that they were chosen, and that the world needed to rearrange itself to just let them shine. They arguably let themselves get played by Oracle (a corporate psychopath) and others as a result.
OpenAI's position is profoundly corporate-narcissistic: all we need is all the money in the economy and not to have to do anything upsetting like think about turning a profit for the next four years. Like rich kids. It would be nice if you believed we were so important that we should get an enormous stipend for just being us.
Anthropic's position is: we think we're so unique and ominous that government needs to make us both essential and terrifying. We have to exist otherwise worse people will.
Both narcissistic positions.
it has to be, because the other way around - the government taking over parts or the whole thing - is inevitable if the trend holds.
ai-celebrities are just clinging to relevance like all the other celebrities out there
The safety side of tech is a PTSD inducing shit show. Governments are more than happy to champion age verification laws, because parents, around the world, are clamoring for anything to pump the breaks on the social media experiment.
Society outside of HN is quite tired of Tech, and I despair of figuring out a way to make this clear to the commentariat.
I don't think anyone in tech is really truly engaging with how quickly the shine has come off the tech industry. Except maybe Apple, who even so still have some work to do.
The yoke of silicon valley is feeling heavy. People might just throw it off.
As someone on the "safety side of tech", social media is being exploited to increase surveillance and government control precisely because its actual social influence is heavily on the wane, and capital is happy to sacrifice what's left to increase the profits of the expanding public/private tech surveillance industry (with "protect the children" controls on social media like age verification being the usual backdoor route it always is).
Society may be growing tired of Tech, but governments aren't, and in fact they're heavily expanding their back channel reliance on not-traditionally-military Tech as an extension of their Defense spending.
Social media was being exploited from inception. Palantir had sales documents for sock puppet management software back in the PHP era.
I don’t disagree that Government is interested in tech, but I will push back on the dismissal of child safety that is inherent in your comment, intended or not.
For all that some people in the firm may have tried to do the right thing, Social media firms have created bad outcomes for children, and executives were briefed on the harms they were going to cause.
This is the dismissal that concerns me, because it ends up miscalculating the level of anger and unhappiness amongst the voting populace, and therefore the political will to pass regulation to reign tech in.
There are numerous bills to limit AI access for consumers, to combat deepfakes hurting children. There are no bills introduced or passed to prevent AI being used to target dronestrikes that kill children abroad, or surveil children domestically.
What the public wants doesn't actually matter right now, only what the government will allow to let pass, which in this case is additional internet surveillance.
Under a future, better government this may change, but (sadly) nothing is going to sink tech's dominance right now.
The anger and unhappiness against tech is good, and hopefully someday they'll burn down all the data centers and I'll never have to hear the words Cloud Computing again, but (to paraphrase a famous Eve Online interaction) it's not going to be today, and it's not going to be us.
I did rediscover this: https://www.youtube.com/watch?v=LmS9vcVNr5A&t=94s
s/Tech/Tech Companies
Tech did it to themselves. People like and want technology. What they don't like and don't want more of is enshittified, user hostile technology. The answer is out there, but our collective school systems failed to teach computing irt free software/open source and instead schools themselves all bought in on enshittified, proprietary tech, or even just dumped trying to teach computing at all outside of "how to login to google classroom and google docs"
I grew up lucky, in that my dad was a dev, my first PC as a kid ran red hat, my high school had an intro to programming class (in BAISC lol). It shaped how I approached computing growing up, and my values. It makes me look at the things we have now and think "No, you're just repackaging community free software and selling it back to me, I'll pass on that."
That experience isn't available to anyone born after that specific era, instead their tech experience is shaped by walled gardens, vendor lock-in, and straight up hostile and manipulative software, so its no wonder they are tired of it. They don't even know a different world (of software) exists.
But as you say, there is a measure of getting high on one's own supply now.
And there's the curious solipsistic energy of Sam Altman whimsically musing in public that it turns out his product is too expensive for people and they complain when you make the price realistic (when it possibly needs to be more expensive for OpenAI to survive).
They seem to believe that the ordinary rules either will not or somehow must not apply to them; it's increasingly bizarre to watch.
Maybe the people around pets.com were this bizarre; we didn't have so much livestreamed interview content to show us.
Anthropic has been surprisingly successful at convincing them that they should control frontier models because they're so dangerous that... only Anthropic can be trusted with them.
(If they're really so dangerous, the right way to deal with them is through a democratic process and taking them out of the hands of a for-profit private entity.)
IOW I don't think he thinks in the same categories as most folks here.
in general I agree people should be reading a lot more sci-fi nowadays than they used to.
Hacker News has been telling me America beats China at "innovation" because of the "freedoms" - especially frew enterprise. I wonder how a nationalized frontier lab would perform.... Andhow the non-citizen researchers would feel about working for the US government that doesn't trust them to use frontier models.
But is that last part actually true though? Sure, there might be 600B+ models available for download and local inference if you have the hardware, but does the users who use Anthropic switch over to those even if they're available even as hosted models? Seems like some do, most don't, Anthropic and Claude remains very popular among the people who use LLMs, there is no denying that.
Only because someone else is paying the bills. I use Claude Opus at work because my employer pays for the tokens and encourages me to do it.
At home, I use DeepSeek Flash. It's not as good, but it's maybe 0.7 quality for 0.001 cost.
BTW, I also use DeepSeek v4 Flash very frequently: fast and so cheap it is almost free.
Cursor's $20 a month plan provides a reasonable amount of Opus tokens as well.
The best answer would be to pull session stats from your harness and compare that against the limits. I think Anthropic publishes the limits of each plan.
If you’re using a pretty stock harness and not doing crazy multi-agent stuff with it, you’re probably fine.
My girlfriend built a whole (but simple) React app with it and only hit the limits of the $20 plan once. In fairness, she was trying to get it to clean up a bunch of 800ish line React files at once with a vague “make it look nice” prompt that she ran a few times. I think it was just churning for like half an hour straight before she burned all her credits.
It’s probably enough if you’re not on a fully agentic development strategy, it’s plenty to have it write tests and do comments and stuff, just not enough to continually have it doing giant refactoring passes.
GPT refused to do so (citing that it's illegal even though I own the games). Deepseek did a wonderful job for 7 cents.
At work I use Opus because, why not? But I could easily switch to a less capable model if needed.
In the. US at least it is actually illegal to download ISOs/roms of games, even if you own a physical copy. It's a stupid law and as a downloader (as opposed to the people hosting the files) your chances of getting into any kind of actual legal trouble are effectively 0, but it is still against the law.
That does raise an interesting question, what kind of laws should LLMs (attempt to) follow? It's easy enough to spoof the country in the system prompt. I wonder how ChatGPT would respond if I told it I was located in a developing country without any piracy laws.
I think being able to align your text generators with a specific set of laws, speech codes, tonality, etc. is very useful. However, all of this should be a user choice.
Your spell checker does not stop checking your spelling just because you are describing a murder scene either.
Safety slop and boilerplating should be configurable as well.
The weights I've tried is the "DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix" ones which fits just about within 96GB VRAM. With some tuning, I've managed to get it to get up to ~60 t/s, I'm sure there is other things to do there too :)
I'm currently spending $200 for Claude. That's around my maximum that I can afford. I could stretch that to $500 I guess. But I saw reports of people spending tens of thousands of dollars with Claude API. That's certainly outside of my budget.
So if/when Anthropic decides to stop subsidizing subscription (if they ever do that thing, I still not sure about that), I'll certainly look at the other options. And available "open weights" LLMs hosted by someone will be my first pick. Right now Claude 4.8 feels very advanced, but things move very fast...
whats the basis for this thought
Again, for Claude, (2) it’s rumored that their API rates have around a 90% profit margin. It’s also claimed that the subscription limits get you around 10x tokens per monthly dollar vs buying them with API rates.
Edit: to drive it home. If a tokens true cost to anthropic is 1/10 of what they sell it for at API rates, and a subscription gets you tokens at 1/10 the price, that’s cost-neutral for the business if every subscription uses every token. They’re selling tokens at cost, not at a loss. Many subscription users won’t use their full allotment. That means serving some users doesn’t cost the business as much - which might push the subscription business from cost neutral to profitable.
Or, maybe I'm wrong, but my understanding is: MoE is just an architecture to keep the activated weights smaller per token. The experts get routed basically token-by-token, and the "experts" themselves don't have a semantic domain so the "expert" word was maybe a poor choice.
"The sparsely-gated MoE layer,[21] published by researchers from Google Brain, uses feedforward networks as experts, and linear-softmax gating. Similar to the previously proposed hard MoE, they achieve sparsity by a weighted sum of only the top-k experts, instead of the weighted sum of all of them."
"Top-k experts," in case of some DeepSeek's models k=1.
https://openrouter.ai/blog/announcements/fusion-beats-fronti...
You talk to a smart, heavy model to build a plan composed of smaller steps. Then you have the heavy model spin up smaller, cheaper LLMs to actually implement the tasks.
The heavy model is basically read-only in that mode. It can read files, execute tests, etc, but it can’t write code. It just tracks what needs to be done, offloads the work to dumber LLMs, validates the task is done, and moves on to the next step.
It sort of pushes humans up the stack. Instead of having a human sitting there prompting the LLM to start the next task, you have another LLM do that loop.
It’s been on my list to try out.
Manually-specifying Sonnet or Opus is a multiplier on the base token rate; specifying Qwen fractions it. Left to its own, it presumably uses the heavier models to create the plan and orchestrate the work; the bite-sized task definitions are delegated to smaller models.
Isn't "distillation" of another provider's model exactly how these models got training date in the first place: Massive amounts of the written word + Prompt -> Answer. Why wouldn't distillation produce similar "reasoning" in the new model? It's just inputs and outputs.
The intuition is that distillation exploits not only the "right" answer but the relationship between answers (what's the second most right answer? the third? etc).
The amount of tokens required to properly distill a frontier model is so large that by the time you could consume the # of tokens you would either be banned for extremely obvious abuse or a new model would be released, rendering your efforts less and less valuable over time. Intelligence is not a linear thing. Being behind just a little bit can have exponential consequences.
That seems to be the argument of Dario, Sam et. al., but I'm not ready to believe it. Time will tell, but this can be a marathon and Anthropic and OpenAI is in getting ready to sprint the last lap of the first mile.
Try running the latest OS models on a normal Mac or PC. Claude Fable and Mythos are systems not just pure models.
And of course marketing. Don't believe the hype.
I think Claude is often times underwhelming. Security concerns are also a concern companies have a blond spot for. The really toughest pro security (Yes, pro! Totally different framing!) company I know is Google after all.
What I can companies advise to do is, really having more than just bug bounties but a professional hacker team that does nothing else but attacking them the whole day and night 24/7. This needs to be coordinated with the government otherwise you might sound an alarm and will be SWATed for doing good. And I would pay them huge sums since the risk and fallout warrant such a treatment, not the standard wage.
Hackers are the real deal, not AI. Proof: Hackers using AI.
AFAICT … despite saying you “disagree”, you appear to be agreeing with the parent comment that the model is less important and compute (all that complex infra) and data (also complex infra) are more important.
Model capability is a function of model size. Raising the bar raises model performance in every domain.
An "idiot savant" model that's overtrained for a specific domain would beat a generalist model of the same size. But scale the generalist up enough, and it'll trounce the specialist. Removing poetry data from a model training mix doesn't give you much - it might even cost you some performance - and "idiot savant" approach of overtraining for a domain has a hard ceiling.
So far, it seems like there's some equivalent of "g factor" in LLMs - a broad "intelligence" value that performance across many diverse domains correlates with. And, as a rule, larger models have more of it.
> Model capability is a function of model size
Model effectiveness has improved across model sizes. You really should try the latest flash variants more. They have become my default for most tasks except for gnarly high-level planning.
A 2026 4B beats 2024 4B, but both are far behind the contemporary frontier. Which makes them bad. There is no such thing as "too much capability" - a "good" model is whatever the current frontier is.
In 2024, a "good" model is one that can be trusted to write a 800 line script. In 2026, it's a model that can be trusted to do gnarly high-level planning and execution both. In 2028, it's going to be something like a model you can point at an extremely involved task, abandon, and have it report back with a "done" in 3 weeks.
The thing about engineering is you don't just use the biggest bolt on the market on every bridge.
> In 2024, a "good" model is one that can be trusted to write a 800 line script. In 2026, it's a model that can be trusted to do gnarly high-level planning and execution both
This sounds a lot like having a single diamond-head hammer as the only tool in your toolbox. As suggested by the name, flash models are fast - sometimes I want to write the equivalent of fifty 800-line scripts. There is such a thing as good enough.
"It's good enough" was said about GPT-4, o1, o3, Opus 4 and more. Guess what happened? Newer models released, people updated their expectations of what LLMs can do, usage got more aggressive, and somehow, GPT-4 went from "good enough" to "obsolete trash".
If you have no imagination, then at least substitute your pattern recognition for it.
The world is hungry for capabilities. There are piles upon piles of tasks that aren't done by LLMs simply because LLMs aren't good enough to do them.
The thing a frontier model gives you is "you don't have to babysit a model to get it to do X", and that X gets more and more impressive release to release.
You do your AI-maximalism, and I'll stick to making trade-offs based on the needs of each piece of work.
I'll do more "per-task model selection" when AIs themselves get good at it.
Deepseek Flash is almost certainly wrong more often than Opus or Fable. It also costs like 5% as much.
The question becomes if I run Deepseek in a loop to fix the mistakes it made that Opus/Fable didn’t, can it fix its own bugs in few enough tokens that it’s still cheaper?
So far, the answer seems to be “yes, by a significant margin”. A lot of tasks are simple enough that both Deepseek and Opus or Sonnet can one-shot it, which is a huge cost win for Deepseek. Even on the long tail, it’s usually like 4x the tokens on Deepseek which is still way cheaper than Opus.
There are things that Opus can do that Deepseek just won’t ever really nail, but it happens so infrequently that I just don’t worry. Like most people, most of what I do is the same sort of “3 tier app with a React frontend” that doesn’t take a rocket scientist to work out.
> I disagree. It is not the model alone. It needs a system which capitalizes on it. And this is very complex. Hardware, software, architecture - it takes a lot to get it right.
What do you disagree with exactly?
It can be done through the magic of SSD offload. The worst case involves seconds-per-token speeds, but that's OK if you only care about low volumes of slow unattended inference, which maximizes utilization for the hardware.
(The real worst case, where you're streaming the whole model from the cheapest storage you could feasibly think of, involves multiple minutes per token for a single inference, or even hours per token batch if you're doing many inferences in bulk. That's a lot less helpful, so there's a space for smaller models at the edge, even for unattended workloads.)
The parallelism is where this starts to fall apart on a local PC. Like I can run some Qwen quants, but I can’t run a decent Qwen model while also running another model smart enough to actually implement it. I’d have to do them in series, and given how long Fable seems to take even with parallelism, I’d probably be waiting days for an answer.
If Fable is "delegating" tasks, then there's actually an agent front end of whatever you think the API is.
We have a local instance of Qwen-3.6 which is more than adequate for running agents. You can mix and match local and cloud-hosted models. (My biggest use case for local models right now is vision models because they're quite small and I can avoid some data-locality issues my customers wouldn't be comfortable with if I sen them to a Chinese model.)
I would say behind (I believe you use the API just like you do Opus), but yeah. I'm not claiming it's a property of the LLM itself, I also presume this is some variety of tool calling agent harness.
> We have a local instance of Qwen-3.6 which is more than adequate for running agents. You can mix and match local and cloud-hosted models.
I'm presuming OP meant local as in the models run locally as well. I do know you can do subagents in Pi (probably others too), but the vast majority of people are going to hit hardware limitations trying to run them in parallel on local hardware.
I'm doubtful Fable's harness is unique in some way that you can't replicate with Pi. I'm mostly doubtful there are more than a handful of people with hardware sitting in their house that can execute more than one meaningfully smart model at a time.
If you're on local hardware, Deepseek v4 Flash is in the ballpark of 180GB of VRAM alone. Even on smaller models, Qwen + a dumber agent to execute is probably in the realm of 60GB of VRAM.
I do suspect you could get Deepseek to do Fable level things with a good harness (or a bunch of models really, I'm fairly convinced the magic of Fable is in the harness rather than the model).
That's a different problem that what you're arguing against.
People like Yud at least have a clear consistency in their advocacy that we shouldn't be developing this at all. Anyone who thinks they can reconcile Anthropic's work with the AI safety mission is in total fantasyland, if it's not just a public persona they've adopted strategically.
Also, the fact that these employees are now in the position to outbid one another for 8-figure real estate gives them a powerful incentive to keep “believing”.
This has been covered before: https://aisle.com/blog/ai-cybersecurity-after-mythos-the-jag... (https://news.ycombinator.com/item?id=47732020)
> Anthropic’s cautious roll-out was justified. The problem with publicly releasing models, however, is that guardrails can be jailbroken, and apparently that is exactly what happened shortly after the release
The future is unevenly distributed. Anthropic, and Amodie in particular, seem to be of the mind they can control a bit of the unknown using words. They are likely being guided by the very product they built. *AI CAN MAKE MISTAKES
That Project Glasswing bullshit reeks of it. Corporations have take control of our attention, our Internet, and now our thinking.
I say it's high time to take it back.
Is not
We sent open weight models against a codebase to find vulnerabilities.
In that sense: The AISLE replication still provides too much information to the model, but its not far off, and others have replicated Mythos' findings in a more clandestine manner on open source models. Some were totally capable of finding the same vulns Mythos found back in ~March (and today, the new Kimi K2.7 is looking extremely good, very little doubt it could do it).
The critical difference is that post-processing: the Mythos model/harness has some step to induce Mythos to actually exploit the vulnerability, leveraging its ability to do so as a ranking mechanism. Anthropic inferred that this led Mythos to discover vulnerabilities nothing else could discover, which is not true, and Anthropic should be held accountable for this weird artifact of that communication. However:
- An OSS model might find the vulnerability but rank it as a 3/10. Mythos finds it, chains it with a second vulnerability, now suddenly its an 8/10.
- An OSS model might find the vulnerability, alongside fifty other vulnerabilities. The operator ignores all of them.
The problem with automated vulnerability detection, including with LLMs, is that they find the haystack, not the needle. Every piece of hay might be a vulnerability, but whether its worthy of fixing is another matter. Mythos does represent a meaningful improvement; it better finds the needle.
This was the primary reason for not releasing it. The difference in the two primary camps around this topic are that the doubter group thinks that Anthropic, and all of their partners, are essentially lying about this (since no one outside Anthropic and select partners has the access to replicate), whereas the other side believes that Anthropic and partners are probably mostly telling the truth without too much exaggeration.
Neither camp has evidence other than unconfirmable reports and/or arguments about economic incentives. I personally think that Anthropic has, in the past, mostly not lied about things like this and has by far been the most transparent and open AI company. That could change, and they could be lying now, but I think that the camp that is certain that they are is far, far too confident in their belief.
As I understand it, ITAR regulations for export controls have just been applied to any form of Mythos. These are overseen by U.S. Departments of State and Commerce, and forbid foreign nationals from access to any form of Mythos, either within or outside the U.S.
Only U.S. citizens and immigrants that are holders of a "green card" may now access Mythos.
It appears that Anthropic does not have internal controls to implement these restrictions in any form, so the only option was to shut Mythos down.
Penalties for ITAR violation can reach ten years in prison and a million dollars per violation. (I can post a link to those details if there is any interest.)
As long as Anthropic is a U.S. company, there is no escaping this.
https://fortune.com/2026/06/14/how-a-warning-from-amazon-led...
Textbook retaliation for not letting them use an abliterated version of Claude in weapons systems.
This effectively renders any US closed model useless for any foreign company. Could happen to OpenAI, Google, etc. Too much of a risk to implement something that can be yanked out because the company didn’t behave the way they want.
Looks like it’s time for Kimi, Z, Deepseek to take the front row. They’ll catch up in a few months anyway. Kimi code 2.6 is crazy good
"When you further combine this realization with the company’s pronouncements about AI’s ability to conduct all economic activity, you realize that Anthropic’s leadership effectively wants to have power over everything and everyone."
This is fearful stuff on all sides, and none of the people involved might realistically be able to navigate the danger.
By that logic, anybody who values safety has a god complex? It’s absurd…
"What this degradation represented was both the capability and willingness of Anthropic to silently alter its models to achieve its policy preferences. In other words, Anthropic willfully validated some of its critics’ worst fears in terms of being a supply chain risk."
a) Anthropic believe that AI is an extinction level risk and that they are the only leading AI lab which takes safety seriously. In combination this puts them in the position of believing that they are the only ones who can save the world, which is reasonable to call a god complex.
b) Anthropic are engaging in actions which aquire and consolidate power in the form of control over powerful AI.
c) "The history of brilliant people convinced they know what humanity needs is a sordid one, precisely because they have convinced themselves that their intentions are good, justifying actions that very much are not."
I'm describing claims from the article and would not word them so strongly myself. But this explicitly does not assume evil intent.
the next model with a gap to mythos as mythos is to opus will be controlled technology from the get-go. the one after it may be top secret.
You can read it all over HN. It's about weakening American influence and building Eurocentric economies and influence. And exercising the same level of choice that Americans prefer as well. Americans also want to escape Google, Microsoft and Apple and more. They've all been caught investing too heavily in government influence and thought control (aka marketing).
And on the other side of that, an American company that deprives the US of AI for defense, is defacto weakening American defense because competition militaries will gain a technological edge by simply taking control of AI companies in their country which the US hasn't done (yet).
There are very valid arguments on both sides, I think.
The numbers lined up if those companies created something resembling AGI, the USA companies captured a large share of the world, and there was lack of competition so those companies could capture a large share of the value.
None of those items were ever going go happen.
Europe has extradition treaties, so the U.S. can force anyone in Europe back to the U.S. for criminal indictment who demonstrates inappropriate possession of this technology.
Can you demonstrate beyond any reasonable doubt that the model weights have been transferred? No. Will the EU judges move to extradite said individuals (and many are EU citizens)? Also no, especially in the face of spurious accusations. And even if they were open to, you can stonewall everything and you will probably outlast any US administration pursuing that.
The signal is clear enough though for the next Anthropic..
Reminds me of the RISC-V Foundation → RISC-V International move to Switzerland. Around the time some dumbass Republicans tried to impose export restrictions on a set of open, world-wide used specifications.
Pandora's box has been opened, and there's no closing it. Capable AI models will be everywhere.
That might be one of the most important points in the post. Very troubling.
It's questionable whether the current government can even unite the talent required for this project. Seizing it might just push all the talent to Europe or China.
The idea of open-sourcing something that falls into the "national security" category is clearly a non-starter unless there's more powerful, classified models that can outmatch them.
I think Anthropic has clearly demonstrated the most responsibility here: they've been crying for regulations, they were careful about Project Glasswing, and they've got comically over-sensitive filters around numerous topics.
if they had more success on alignment and safety research then I don't think the cludgy filters would be necessary
Ant’s models, culture and leadership actions are largely consistent with their beliefs, even if they may seem flawed / incomprehensible.
Relevant anecdote: I interviewed with them for a MTS role in 2023. I think the technical part went fine but the interviewer was clearly frustrated by my low regard for LLM safety. I didn’t get the role.
Anecdotally I've heard this is weighted as much as the technical interviews.
It's actually not that hard to explain if we take into account what Dario kept saying: he, or Anthropic thereof, would be the gatekeeper. It is he who tells the government how to use Claude to design drones. It is his model that tells users whether they can ask a question to Claude or not. And it is he who can assess whether a jailbreak is dangerous or not.
Personally, I think that is way more dangerous than being a hypocrite. Dario is basically the Robespierre of the AI era. He believes that only he gets to decide whether our thoughts, or our prompts thereof, are pure. Anything impure gets purged. For his moral utopia to stand, he has to wield the guillotine. Otherwise, with the chaotic diversity of human nature, how else do you manufacture that perfectly uniform, beautiful morality?
I'm not saying these things aren't useful or interesting. But if get told a slot machine is not just a tool, but that actual tools have to go the way of the dodo so we can focus more on getting good at gambling and befriending the dealer, I know something is up. And in that sense, I'm actually pleasantly surprised at how crappy many tech companies are at not letting the mask slip before the victim is actually in the bag. It doesn't seem to make much of a difference, but imagine if they were actually good at this.
They’re like look at our safety and they do all thesse outrageous things.
I always thought safety was interesting in and of itself, but for some reason HN doesn’t have many people from the safety side of tech in conversation.
Tech isn’t a niche hobby anymore; Billions of people are impacted by the decisions of a few firms.
My grandfathers android had 3 different messaging apps installed, somehow. AI is enabling new forms of fraud at a time when we still haven't solved the old ones.
And this is all in the first world, move your coordinates to the developing world? We had human trafficking to get educated English speakers into call centers in Laos/Cambodia to defraud first world inhabitants of their money.
We aren’t in the early days of tech anymore, and the kind of scale that we have enabled comes with it a certain cost. We can choose to ignore them, or to understand them, but we will feel their impacts all the same.
“Claude, I want to blow up a factory running this leaked software. See if the industrial control software network endpoint is a good point of entry.”
It’s doing the same work and producing the same output for both prompts. How do you block one but not the other?
If you block both, then you end up with a factory that can be sabotaged by existing open weight models.
If you are to believe Anthropic, Fable was export controlled for bug finding, not for exploit construction. They seem to be working to make this the "bright line" for LLMs being a national security risk. My guess is that will be the case they take to Washington this week.
This is why responsible/coordinated disclosure exists in the first place.
The factory does decent software engineering - for which it can also use the same llm - so that when an attacker does either, a sota llm does not find bugs to exploit.
The EU options are not even close to what CF can do
What limitations does bunny.net have?
A huge free tier (technically, none)
It's probably indicative of a less predatory model, but CF got a ton of mindshare by offering their free tier. I do basically nothing in the frontend space, but I default to CF because I'm used to using it due to using the free tier for personal projects.
That's not the problem.
The US government can export ban GPUs like they do now to more countries if needed. Even if the infrastructure exists, the GPUs won't.
This too, will end up being a good thing for them. The ban will end up getting lifted due to some "amazing deal" in the coming weeks and Anthropic will now have the "Trump tried to ban them, so they MUST have the most advanced AI model in the world!" stamp of approval just before IPO.
All this stuff is pro wrestling kayfabe.
1) It’s safe to assume the US would do its best to prevent it, and even if Anthropic was successful in exfiltrating their data, code, models, and people, I’d imagine the US would immediately block all US companies from working with them. So they’d be blocked from their own US-based compute, plus Google, Amazon, Microsoft, xAI, Meta, etc.
2) Where would they go? China maybe, but as far as we can tell it doesn’t have sufficient compute for Anthropic’s level of need. The EU likely as or more restrictive in different ways to the US - the EU is hardly buzzing with AI innovation. Some Middle Eastern countries might have the money, energy, and interest in carving out such a position, but no compute. Plus I’d imagine the US would act directly against any country or region receiving them, economic or otherwise.
3) Then, as said elsewhere, the US would block GPU sales to wherever they found a safe haven, preventing the buildup of the compute they’d need to continue.
Depends what you mean. The academic work seems largely... fine? Plenty of good work came out of Europe or European researchers. It seems the problem is more "trying to build a trillion-dollar company of any kind".
It's an interesting question: does the EU seek only to regulate successful modern American companies to death, or home grown ones too? Probably not a gamble worth taking.
I wonder if the incumbents in each country actively lobby against it. I suppose it's easier for massive corporations to deal with cross border issues. The onerous regulatory boundaries are a nice price of entry for them that keeps out upstarts.
I really dislike this belief (that has at least been expressed here) by some that X is okay because they-really-believe-it. This has a real Road to Hell stank on it.
It is incredibly convenient when your predictions or supposed beliefs go south. Well, we really believed that we were doing it for the betterment of human kind. And we really believed that X was an existential threat that was inevitable in which case we had to step up and do it because we we the only good guy ideologues. So sorry but not sorry.
I also don’t care if commenters know rank-and-file on the inside that “really believe it” as well. Not for one second.
So... what, you just don't trust anyone good? Would it be better to pull in a health insurance CEO? They're happy to watch people die for profits, no concerns at all about them pulling a "greater good" card because they're in it for entirely selfish reasons.
> So... what, you just don't trust anyone good?
The baseline here is apparently that they are good, I’m just supposed to trust and shut up?
Or even just propose an alternative: who has a better track record, here? Who DO you trust?
I already covered this.
> Or even just propose an alternative: who has a better track record, here? Who DO you trust?
My mother.
Modern society is built on the idea that competition is required from companies, and we seem to be exiting that age into a new world of monolithic, monopolistic, mega-corps. Personally, I find that a real route to dystopia.
Where do you draw the line here?
What happens when your car stops working because you're driving a tesla, but you're working on EVs for Honda or Ford?
What happens when your macbook stops working, because you decided to commit to changes to ARM software, or RISC-V?
---
And before you dismiss those, this is literally what Anthropic is doing TODAY. Using their tools to develop competing tools is something they classify as mis-use, and shut you down for doing.
Personally, I just can't accept that as a valid moral stance. Wonderfully successful, abusive, and dystopian? Absolutely. Moral? FUCK NO.
When tools turn themselves off because the manufacturer has decided it doesn't like how you're using them... you're a slave with no autonomy.
Like, there's a very critical point where you are asking to use their servers to directly compete with them.
This has been normal for somewhere between "decades" and "the entire history of commerce"
Ehh, I think it's a lot more grey than "definitely not". It's hard to ignore that their claims that their model is so dangerous they can't widely release it is tantamount to declaring that they're in a league of their own and have to be treated with white gloves to prevent the sheer power of their model from shattering global prosperity.
This isn't the first time, and nothing bad has happened with the prior models. Every time it gets a little harder to believe that they believe in the threat, and makes it feel a little more like it's just to build hype. There's only so many times you can say "this model is a threat to the world", have it turn out to be nothing, and avoid people accusing you of lying to pump stock prices.
Not affiliated with the bench in any way, but I think it surfaces important differences between the behavior of the models from different labs.
TLDR: The benchmark is measuring pushback in response to nonsensical requests and questions, as opposed to going with it and hallucinating a nonsensical answer.
[0]: https://petergpt.github.io/bullshit-benchmark/viewer/index.v...
> Anthropic models have consistently been top-scoring in BullshitBench[0]
eyeroll I find that Anthropic models feel big and dumber.
https://www.endorlabs.com/research/ai-code-security-benchmar... puts Fable 5th, which seems about right to me.
I'm interested in code utility and correctness, even if the majority of AI use is not focused on that.
Anthropic likely knows this and is merely performing a song and dance. They're auditioning to be THE frontier AI lab.
Do they have it or do they just sell it?
Anthropic believes they have the responsibility to guard their tools from mis-use. That is all. They are not trying to "control" anything or anyone. They do however decide what they think is mis-use.
I think assuming you have the ability to guard a tool (that you're "selling" for profit) from mis-use is the definition of "controlling behavior".
It's the kind of ethically myopic take that can only really exist in this new digital age - where tools aren't actually sold, they're just digitally rented.
The most telling part of the "control" narrative is that they happily classify "competition" as mis-use. We're headed back to serfdom on a speedrun.
Installing safeguards to prevent a tool from being used for certain things is a perfectly natural and common thing to do when you are providing the tool as a service. For example, blocking VPNs and open proxies from accessing a free service if those are a major source of spam and abuse. Note that Anthropic never provided the model for offline use in a form that includes DRM -- they are simply safeguarding the service that provides access to the hosted model. The only ethical concern I see here is that some of their safeguards are ones I wouldn't personally agree with, and in a world where dependence on a model is expected it can become an issue if the model refuses to perform in some cases, etc., but that doesn't automatically mean the refusal itself is unethical unless that issue was known and expected (and unless the alternative is not bigger, worse bads)
Also note you are not even "digitally renting" anything. This is the exact same type of thing as, say, real humans in real life refusing to perform services for certain clients or under certain circumstances. Networking makes it possible to decouple some of these things, but that doesn't magically make it renting or automatically turn a refusal to perform services into an attempt to control clients. Just the same as I can choose to refuse any request, which does not automatically constitute attempted control over the asker. There can be ethical concerns about whether my refusal causes problems that I'm obligated to avoid (and whether or not such obligation exists), but that doesn't automatically contaminate the refusal itself unless I have knowledge of and intend the bad.
To use a much more relevant example, Anthropic's refusal to allow its models to be used for war (among other things) does not constitute any attempt to prevent war. It's only a refusal to assist in it. That's not some unfair, unethical attempt at controlling the government, that's just Anthropic not wanting to be responsible for assisting in war.
What happens when your car stops working because you're driving a tesla, but you're working on EVs for Honda or Ford?
What happens when your macbook stops working, because you decided to commit to changes to ARM software, or RISC-V?
Tools should be neutral. The idea that a tool can only be wielded in a manner that its manufacturer approves of is... scary.
That's a real quick hop and a jump to a really, really ugly spot, societally speaking.
And sure - technically Anthropic is selling a service, but even that idea makes me quietly upset. The only reason they don't sell a product as a tool itself is that they have more control over the model as a service, and expect to be able to extract even more profit from their customers with this route.
---
My real hope is that open models FUCKING CRUSH them. Because almost nothing is scarier than a self-righteous, moral zealot.
> What happens when your macbook stops working, because you decided to commit to changes to ARM software, or RISC-V?
These sound similar, but aren't the same thing I'm talking about. It's more similar to a rideshare company refusing to serve you, or a cloud PC cutting your access.
It's not the same thing as DRM, which is when invasive malware attempts to control what you do with your devices -- your property -- and your resources, that you own. A car that can be shut off remotely, or that can detect competitive conditions and cease operation, is not the same as a hosted service refusing to have you. It is DRM.
Likewise, a MacBook that stops working based on your affiliations or your activity is not the same either. It is DRM. (Technically, Apple Activation Lock is DRM. So are locked Android bootloaders that can't be flashed with custom verified-boot signing keys, etc.)
When you're using someone else's private resources, someone else's private infrastructure, by default they have the right to simply no longer serve you, at any time.
DRM, by contrast, is when a machine or software you already own decides it will no longer function for you.
Yes, this is scary. We are already confronting this right now. Faceless corporations abruptly cut your access, or ban you for life from really important things. PayPal steals your money and pockets it instead of giving it back. Thousands, tens of thousands, hundreds of thousands of dollars (or whatever) gone because they said so. It's awful. It ruins lives.
But this happens because you need to draw a different line. Not whether one should be allowed to refuse services ever, but when. It's not okay that PayPal can steal tens or hundreds of thousands of dollars or more from you with no recourse, just because something looked suspicious, and you'll never know what it was even if you bankrupt yourself with arbitration costs (since their terms of service say you are simply not allowed to sue them anymore, and for some reason companies are allowed to just say this now and have it be legally binding).
A refusal at that point can be devastating, especially when it's for no fucking reason. Or when it's for a completely shitty and unjust reason, like in banning all accounts that have been involved in buying or selling adult content. There are cases where you can ruin someone's life by suddenly refusing to serve them with no recourse, and those are the cases that shouldn't be allowed to continue.
So push for that. Businesses shouldn't all have to serve every customer, but they shouldn't be able to just suddenly ruin someone's life. That's the scary part.
> The only reason they don't sell a product as a tool itself
...it's not the only reason. You can't exactly run trillion-parameter-scale models on the kinds of hardware people tend to already have in practice. And who's gonna pay tens of thousands of dollars up-front for their own inference hardware for the thing? (If you sold it as a hardware product.)
> My real hope is that open models FUCKING CRUSH them.
I would like this too!
I don't find this blisteringly clear at all. A company making it harder for competitors to steal their IP is perfectly normal. This is Ben Thompson's personal grudge against Anthropic showing, yet again. He can't think rationally about this company.
- Satya Nadella
Microsoft when they're losing.
>Every company is going to have to build what I think of as human capital and token capital. Human capital comprises the knowledge, judgment, relationships, ingenuity, and pattern recognition of its people, while token capital is the firm’s AI capability it builds and owns. Importantly, human capital does not become less valuable as token capital grows.
- Satya Nadella
Either incompetent or lying.
This only just shows how strong Mythos/Fable will be, once released to the public.
I'm guessing about 0.5 year till public.
Doesn't this *exclude* allies countries?