[mekpro]

It’s clear that Anthropic has run out of the compute capacity needed to serve Mythos publicly.

They’re using security concerns to mask their inability to deliver the model at scale, while still trying to maintain their lead over OpenAI. As a result, they’ve chosen to release it privately under the banner of an “ethical” rollout.

[mofeien]

Jack Clark, co-founder of Anthropic said the following at an Oxford lecture last week ([0], at around 10 and 12 mins):

    "It's a technology that we do not fully understand because it's more grown than made. And it is a technology that you can concoct plausible scenarios where it could kill every single person on the planet. So to think building this technology is without risk would be an act of hubris or insanity.

[...]

    The technology is in fact so powerful that I should clearly state that if it was possible to elegantly slow the development of this technology to give ourselves more time as a species to deal with it, that would likely be a good thing. ... But in the absence of a coordinated global slowdown, we are left with the current situation, which is a powerful technology being developed at breakneck speed by a variety of actors and a variety of countries locked in a competition with one another where commercial and geopolitical rivalries are often drowning out the larger existential-to-the-species aspects of the technology being built. This isn't an ideal situation, but it's the one we find ourselves in."

They know they are in a race that no one will win.

[0] https://www.youtube.com/watch?v=8zIcP5WlShw

[tomjakubowski]

It's worth noting that Clark's career started in PR and journalism.

[protocolture]

>you can concoct plausible scenarios where it could kill every single person on the planet.

Idiots can scary black box their way to that concern. Plausible? Not so much.

[graerg]

It is already very plausible (and has been since the 1950s) without the advent of LLMs. This is just another layer on top of the preexisting and very plausible existential threats we already face.

[protocolture]

Detail it. Justify it.

Your comment about before LLMs is a non sequitur. Demonstrate that an LLM can kill everyone on the planet.

[patcon]

Task a squirrel with justifying the risk of a fox, but from the biomolecular level. That is the level of the task you are setting out.

There can be arms-races in domains that are unfathomable to the participants. A small mammal will die a billion times over before it understands the evolutionary mechanisms and the genetic playing field on which it loses. Actors are not necessarily privy to understand the means by which they will lose, and humans have only existed in a small window of time in which we fashioned a manicured garden, in which that full understanding was briefly possible. It is not favoured in the universe for us to fully understand our environment imho

If the risk must be exhaustively detailed before it is given credence, we are already doomed, and deservedly so

[protocolture]

>Task a squirrel with justifying the risk of a fox, but from the biomolecular level. That is the level of the task you are setting out.

Thats a really deep thought for a 12 year old.

>There can be arms-races in domains that are unfathomable to the participants.

You cant even justify LLMs as being unfathomable. Oh watch out I am fathoming them. You cant stop me fathoming all over the place.

>A small mammal will die a billion times over before it understands the evolutionary mechanisms and the genetic playing field on which it loses.Actors are not necessarily privy to understand the means by which they will lose, and humans have only existed in a small window of time in which we fashioned a manicured garden, in which that full understanding was briefly possible. It is not favoured in the universe for us to fully understand our environment imho

Non Sequitur. One that sounds like it was made up for that "What the Bleep" garbage.

>If the risk must be exhaustively detailed before it is given credence, we are already doomed, and deservedly so

The risk needs to be justified as something more substantial than weird people writing wannabe edgy messages on the internet. If someone on the internet told you that we need to drastically reverse living standards because there's a risk that modern technology will summon King Kong any reasonable person would ask for the working out instead of running for a cave.

[delusional]

"Oh what peril we are in where I must get rich by killing all of you" Is a statement that should make you disregard anyone saying it at any time. Either they are liars, or they are so morally bankrupt that they are willing to sacrifice the species for short term satisfaction. Either option makes them more fit for a mental hospital than a stage.

[airstrike]

It's worth remembering that people do not always say what they believe. Instead, they often say that which benefits them the most.

[clbrmbr]

Was a good watch, tho would have liked to be there in person. Props to Brenden & his Cosmos team for really setting the bar.

[WarmWash]

Thank you Mr. Altman for firing the starting gun when no one else wanted to race.

(The ambiguity of sarcasm is intentional here.)

[sumedh]

Didnt Google start the race with their paper?

[WarmWash]

Google (and other labs) wanted to keep the tech internal because of the obvious safety concerns. Once they were confident that the tech was understood and under control, the public could start being drip fed. Everyone on the ground back then was hyper cautious.

Then Altman made ChatGPT public, and the race began.

[sumedh]

No that is not true, Google didn't release it because they feared it would kill their search business, nothing to do with safety.

[NiloCK]

I find this line of reasoning highly dubious.

Yes, Anthropic is compute constrained, even after the SpaceX Colossus deal.

But supply constraints are the normal operating mode of any market. Anthropic could choose to serve whatever models it pleases at whatever price points it chooses and let the market decide where the value is.

If Mythos at $X overwhelms their capacity, they could just charge $X+1. If still overwhelmed, there are larger prices as well.

[OtherShrezzing]

During periods of market exuberance, it’s in the vendors interest not to reveal where exactly x+1 is. At the moment, everyone just guesstimates the company’s TAM. Bringing certainty to that guesstimate cuts Anthropic off from the most exuberant market participants, bringing their post-IPO price down unnecessarily.

[deaton]

The question is, will anyone pay enough for Mythos to offset the opportunity cost of offering that much Opus? You don't want to end up in a spot where you don't have enough compute and your service's reliability degrades to an unusable state like xAI.

[orrito]

I feel like there's always a demand for the very best models, even at insane prices. If the opportunity cost is x times opus, maybe few but there will always be companies willing to pay x+1 times opus.

[suttontom]

Isn't that kind of what they're doing with this rollout? Except they're just hand picking the companies.

[ikiris]

Only if the price is under the competition, which does exist now.

[JumpCrisscross]

> Mythos at $X overwhelms their capacity, they could just charge $X+1

This may not be as valuable in the long term as getting committed customers hooked at a sustainable price.

[tiahura]

Sort of, but valuation models depend on X being in a certain range. If it > this range, revenue and therefore valuation are impacted.

[malfist]

And then the bubble would collapse. Corps are already putting limits on token usage across the board because of costs. Increasing costs would significantly contract the hype bubble.

[mattnewton]

No insider info, but just wanted to mention that pricing signals things too. If Mythos is only servable at $X*Y dollars and isn’t Y times better than $X of compute at another provider, it’s quite possible that affects the IPO price negatively versus the halo of having the worlds most expensive model that is “too powerful to release” unpriced and unbenchmarked.

I think that most people at Anthropic are true believers from my interactions with them so I don’t believe this theory anecdotally. The simplest explanation is that it really is taking a while to gain confidence they won’t be used for a spree of bad cyber attacks. Knowing how long it takes institutions to fix security issues when filed by humans I would be more suprised if this wasn’t the case.

But I would forgive anyone who did think it was deliberately sandbagged; given the staggering sums at play, true believers might believe the ends justify the means to a little “marketing” like this.

[jb_briant]

It is not "clear", as your comment suggests, it's hidden. Which is semantically the opposite of clear. Regarding your theory, might be true, might be false. But it's highly speculative.

[Forgeties79]

All of us, including you, know that he is not saying "they are being transparent." When someone says "it's clear that..." in this way they're saying "It's clear to us what is really happening here.

[jb_briant]

It's not clear, there is no tangible proof that Mythos is not released because they don't have compute power to serve it. Saying that would imply that the "too dangerous" is a lie. Nobody has proof. It can feel "clear" for you, but it's not. Hence, I correct it.

[jb_briant]

Yes I got how they used the phrase. And it was wrong, so I wanted to react. Thanks for your addition, it dissipates any doubt on the intention of OP: he thinks Anthropic is hiding the lack of power by pretending it's too dangerous. But he is wrong to assume that without proof, hence my reaction.

[Forgeties79]

Agreed, but I'm talking about how they are, very clearly, using the phrase.

[WhitneyLand]

The not clear comment is valid by either interpretation.

To a lot of us it’s not clear that’s what’s happening. It’s speculation and one possibility.

It may also be a secondary consideration and not the primary gating factor.

Anthropic has had their missteps but it’s still plausible to take what they say at face value.

[jb_briant]

I agree, saying "it's clear" when at best, "it's plausible" doesn't let the conversation happen. And pretending to know what is going on behind the scene, anon on HN is not credible

[baq]

I had to patch my Linux boxes daily at some point in the past couple months. I don’t want Mythos to be publicly released for as long as it is economically feasible for Anthropic. I hope they have a gentleman’s agreement with OpenAI and DeepMind about this, too.

Chinese labs will force their hands, until then let’s hope maximum number of projects get patched at a reasonable pace.

[cmxch]

I hope that such agreement gets broken hard and given the MSRC cold shoulder. If that means abliterated Qwen et al embarrasses Mythos to deliver a wider rollout, I’ll take that.

Trusting Anthropic to deliver is like asking Microsoft to pay out for bugs.

[simonw]

They started Glasswing before they struck that $1.25B/month deal with xAI/SpaceX for their (notoriously dirty) Memphis data centers.

So they have a whole lot more compute now than they did last month.

[mekpro]

Yes, 300 MW from SpaceX helps a lot, but I think that’s mainly to support Opus demand, which has grown faster than expected. If Mythos is roughly 5× more expensive to serve than Opus, as the pricing suggests, then 300 MW is nowhere near enough to enable large-scale deployment of Mythos.

As an ordinary developer who relies on a $20–$200/month subscription, I feel disappointed by the release of a paper describing a model that I can’t actually use.

[aspenmartin]

Ok but they can easily upsell this to enterprise customers at a market price reflective of their capacity constraints. Big corps would pay it, this is clearly a major update.

[nickthegreek]

But that compute might not be available to then long term. Hard to make big moves with a contract like that.

[simonw]

I don't know if any of the big AI labs have confidence in planning for the long term.

For all they know they'll find a new optimization that lets them serve Opus class models for half the computing cost next month. Or someone will invent the next OpenClaw and demand will 10x over night.

[mrbluecoat]

While I tend to be cynical with big tech, if this statement is indeed true we owe them some thanks for staving off a zero day tsunami.

> 50 initial partners ... found more than 10,000 high- or critical-severity security flaws.

[cobolcomesback]

So why is OpenAI also releasing 5.5-Cyber in a private manner? Are they also out of compute?

[LiamPowell]

OpenAI has been pulling this marketing trick for years. Remember how GPT-3 was too dangerous to release? It's also probably bad PR if script kiddies have access to GPT model with no guardrails even if it doesn't enable any significant attacks.

[signatoremo]

I suppose you meant GPT-2, but for years? Did they say the same about subsequent models?

[LiamPowell]

They did it for 2 and 3, however it looks like they didn't for 4 and 5.

GPT-2: https://slate.com/technology/2019/02/openai-gpt2-text-genera...

GPT-3: https://www.itpro.com/technology/artificial-intelligence-ai/...

[neuronexmachina]

For GPT-2 and GPT-3 it seems like the concern was that they hadn't yet figured out how to properly write safeguards for it yet:

> The company believes making its API generally available was made possible due to its progress with safeguards, and that opening up the API to all developers will help see applications developed faster. ...

> A large emphasis has been placed on safe use of the tool, which in the past has been criticised for a range of shortcomings, including racism and prejudices against specific genders and religions.

[LiamPowell]

Maybe, but they certainly used it for marketing too. At the time they contacted a bunch of publications and gave them access but told them they could only share snippets of the output [1]. The only reason to set restrictions like that is marketing.

[1]: https://youtu.be/TfVYxnhuEdU?t=102

Transcript of the timestamped part:

> Now, OpenAI's terms of service don't let me give you the full list. I have to curate them, and show you a sample. Those are the terms and conditions I agreed to.

[AgentME]

GPT-4 was announced in March 2023 and wasn't made available to all developers until July 2023.

[atleastoptimal]

Why do you think that? All these rumors about compute constraint just seem like speculation and not based on any data or information. All they would need to do is increase their prices to free up compute capacity.

[lossolo]

Probably. This is an 8-12 trillion-parameter model, which is why it costs so much, that is also a major reason, besides RL and synthetic data, why it suddenly gained these new capabilities. They claim it was not fine-tuned or trained specifically for cybersecurity, but is instead a general purpose model.

[notahacker]

The security concerns argument would have worked better if a forum full of people hadn't promptly obtained access by the extremely sophisticated tactic of guessing its URL...

[pshirshov]

I bet Huawei and co would be happy to sell them some cheapo chips for inference!

[Almondioco]

Or they actually take the 'technology can kill' serious.

[cute_boi]

Also, they just want to jack up the price by creating sensation.

[y0eswddl]

it's also a marketing ploy.