surinder

23, professional tool user.

[surinder @0xsurinder]

@Creative_Math_ it realized you were on some llm tomfoolery and secretly rerouted you to a braindead model

haha jk

unless….?

[surinder @0xsurinder]

one of the worst “technical” reports i’ve had the displeasure of reading. didn’t even catch the LLM-isms til i finished 5.7.1. rough way to kill 20 minutes

[Alexander Whedon @alex_whedon]

Here is the technical report on SubQ 1.1 Small.
https://t.co/bu8AEc4lsk

This is the second iteration on our Subquadratic Sparse Attention (SSA) model, and the first to be deployed with design partners in the coming weeks.

The results are compelling and verified by @AppenResearch.

- Near-perfect long-context retrieval up to 12M tokens on the needle-in-a-haystack test, with up to nearly 1,000x attention compute reduction.

- A balance of long-context optimization and general reasoning ability, with strong performance retained across knowledge, coding, and non-coding enterprise agent benchmarks.

- At 1M tokens, SubQ 1.1 Small requires 64.5x less compute than dense attention and runs 56x faster than FlashAttention-2.

These results highlight a significant scaling advantage thanks to the efficiency gains from the SSA architecture.

We included some details and learnings from the development process which may be helpful to the community.

Comment with questions, I’ll try to respond!

[surinder @0xsurinder]

Guy has been on a tear with the f-bombs since that first one in June of last year. Doesn't hit the same anymore though unfortunately.

[Insider Paper @TheInsiderPaper]

JUST IN - President Trump told Axios: "Why did Bibi have to do a fucking attack? I was so pissed off. I let him know. He has no fucking judgement. I let him know that"

[surinder @0xsurinder]

Champs baby!!!!!!

[surinder @0xsurinder]

@fkasummer which is why if they nationalized it tomorrow under the guise of keeping jobs safe or fighting against deepfakes or some BS, almost everyone my age (23) would support it. Very sad. A two-tier world.

[surinder @0xsurinder]

?

Not what I said.

The two options available to Anthropic today are (1) compliance or (2) publicize the jailbreak if they don't think it's actually that harmful. I don't see any other avenues in the short term.

The long term is (3) establish an independent agency/board who has the credibility and technical ability to verify if a model is truly dangerous, but most of that talent is already within Anthropic itself. And that can't be done now.

I am not arguing for either (1) or (2), just listing them.

USG demands have basically put a stasis on the industry.

Telling foreign nationals they have no right in the upside of creating great models encourages them to look elsewhere if USG wants to pull the plug at any time. Who wants to help build a model they can't even use?

Finite resources like compute and talent become more spread out across more places outside of the US, because adversaries just learned that relying on US models is a no-go but really they always knew that. The key is allies in Europe learned that too.

Frontier labs are less motivated to pursue improvements if the business use-case dies like it died yesterday where now you cannot serve the best models to foreign nationals even domestically. Why would Google or OpenAI pursue research and spend 100s of billions if at the end of the road there is no one to sell to anymore because of regulation and all users have to provide proof of American citizenship.

US hegemony weakens. I think you'd reasonably agree with that much.

All this because USG believes the threat to be credible (despite having their hands in the model for the last 4 months and not being able to produce it themselves), Ant wants an actual process in place and doesn't believe in the threat (and is likely the best informed to make that call as the producers of the model themselves), and you and me get to argue on the internet without any of the pertinent technical information like the actual severity of the "jailbreak" itself because again, we don't have a process!

[surinder @0xsurinder]

sacks said they refused to fix the jailbreak. clearly ant thinks it isnt a valid concern. private squabbles leading to models getting pulled is not good for the industry.

nor is overzealous regulation in the name of broad-sweeping 'safety.'

comply with USG or publicize the jailbreak if it's not that concerning.

im not sure what is so confusing about the paths that anthropic has available to it right now in the short-term?

either do it this way or get a regulatory agency that can independently verify if models are dangerous and do so efficiently in such a way to not create a backlog. which is what has been proposed by ant. but that can't be done instantly right now in this very moment. it is a very hard technical and bureaucratic problem but the only way forward in the long-term.

else this just devolves into scenarios like we have now with the he-said she-said.

[surinder @0xsurinder]

@ShakeelHashim Anthropic asks for non-arbitrary regulations

Gets arbitrarily regulated

"Haha, they got what they asked for."

Nuance is dead on the internet.

[surinder @0xsurinder]

be tech twitter

hate dario because he never shuts up about safety

“bro thinks he’s the AI crossing guard”

be USG

have been in the loop for 4 months

tell dario his model is too dangerous even with the insane guardrails

say you need to clamp down harder

find one minor jailbreak

dario explains this is a deeply complex, basically unsolvable problem

also explains the supposed issue is not actually an issue

USG says no actually you do not understand your own model

“ackshually, just perfectly separate good cyber from bad cyber, infer user intent from vibes, prevent every multi-step workaround, fix jailbreaks forever and please solve interpretability while you're at it, this is all really easy dude"

“we cannot believe you are not complying”

nuke the models

lab vs USG/Amazon squabble causes widespread instability

tech twitter immediately flips

“wow I cannot believe Dario was so reckless with these dangerous models”

same people were calling him the safety hall monitor 3 nights ago

many such cases

[David Sacks @DavidSacks]

I’ve had a number of conversations with folks inside and outside government about the current situation with Anthropic, and here is what I believe to be true:

— As we know, Anthropic publicly released its Mythos class models earlier this week under the commercial name Fable.

— Fable is Mythos with guardrails. But if those guardrails fail, then you’ve exposed Mythos and its advanced cyber capabilities to people who shouldn’t have them. (Keep in mind that Anthropic itself widely promoted the idea that Mythos was a cyberweapon and needed to be regulated as such. They asked for government regulation of Mythos and championed the guardrails on Fable. If there is a vulnerability — big or small — it is Anthropic’s responsibility to patch.)

— A highly credible trusted partner of both Anthropic and the USG who was testing Fable came forward with a jailbreak of those guardrails. The Admin asked Dario to fix the jailbreak or de-deploy the model. Dario refused.

— In their blog post, Anthropic defended its decision by saying the jailbreak isn’t serious. That is not what the trusted partner and the USG believe; nor is that kind of minimizing language consistent with Anthropic’s brand as the AI safety company. It’s difficult to fathom how they could claim a jailbreak allowing operability of a cyber weapon could be defined as not “serious.”

— In the past, Anthropic has always said that safety must be top priority and taken super seriously. In this case, Anthropic prioritized the continued offering of the consumer model over safety.

— In reaction, the Admin issued the export control. The Admin did this reluctantly. It’s been very surprised that Anthropic hasn’t wanted to cooperate with a reasonable safety request (ie fixing the jailbreak issue). Anthropic’s reaction is very much at odds with their branding and ethos as a safe AI research community.

— The Admin’s hope now is that Anthropic remediates the safety issue, the export control is lifted, and Fable goes back into general release. The Admin wants all of this to happen as soon as possible. It is frankly bewildered that Anthropic hasn’t wanted to comply with safety requests that it previously said were its highest priority.

— Those trying to misdirect and tie this action to the prior DoW/Anthropic issues are wrong. The Admin values Anthropic’s technical capabilities and feels that this issue, while serious, should be easily resolved. The ball is in Anthropic’s court.

[surinder @0xsurinder]

@corsaren personally did a complete 180 on disliking anthropic because of all the dogshit i saw on the TL last night. people genuinely arguing for anthropic to get nationalized because they don’t like dario’s incessant safetyisms. complete inability to think about 2nd order consequences.

[Emerson T. Brooking @etbrooking]

life comes at you fast I guess

[surinder @0xsurinder]

@far__el you can hate anthropic all you want for the holier-than-thou attitude and the moral grandstanding but this falls squarely on the govt

[surinder @0xsurinder]

@mattyglesias i vote (c), use this as leverage to make all labs deferential to the government at all costs and reign in Anthropic, or else we prevent your employees from using your own models and then make your entire business model go poof

pretty dystopic stuff.

[surinder @0xsurinder]

my immediate thoughts are ai tech just got neutered hard today. it's a miracle that anthropic even got this far with the gov breathing down its neck and it self-regulating so much. if this doesn't get fixed by monday morning/no clarification im expecting a market blood bath.

[surinder @0xsurinder]

@DoWCIODavies @POTUS @SecWar Good to know it's not the brightest at the top.

[surinder @0xsurinder]

I'm absolutely appalled at the amount of shitposting and dumbassery on the timeline today.

But really the immediate gut instinct to pile the blame right onto Anthropic is what pisses me off the most, instead of against the government overreach that took place today.

Anthropic is the only company on the planet willing to have intelligent discourse and sacrifice their own revenue to safeguard AI because whether you like or not, they are idealogues and true believers in their models. Whether or not you agree (and I certainly don't with Anthropic or Dario on most things, especially what he posted a few days ago), the ability to have a public conversation is the most important.

Instead the gut reaction has been to take the piss on Anthropic and that they 'deserved it' and that this was a result of their marketing.

This was going to happen regardless because of the inherent capabilities of the model, regardless of how hard they 'marketed' it.

Does it sound great to you to have the best models in the world reserved only for government use? Or how about government elites having backdoor conversations as to whether or not a model is acceptable to release to the public? I don't like the idea of either of those things.

To the government, this is a card you only get to play once. Now our adversaries know that their foundation is built on quicksand and will naturally accelerate the building of their own models.

Domestic companies will never pursue better models if they have reason to believe the government will pull them at any time. That genie is never going back into the bottle after today.

Whether you like it or not, the frontier of the industry has been built on the research contributions of foreign nationals for the past decade, at the very least.

If there is such a credible threat of adversaries building Mythos-tier models why only wait until the model is public to pull the plug? There are tons of foreign nationals in all the leading AI labs today, and they have all been there for years.

The US certainly took a step back today. They created an unenforceable policy with zero understanding of the industry and ramifications, and now the entire industry and the country will suffer for it, but please, continue to shitpost about how this is all Anthropic's fault because the vibes today are about shitting on Dario.

[surinder @0xsurinder]

@superwhisper joking about government overreach to market ur product, very classy.

[surinder @0xsurinder]

@AnthropicAI The best models will be for the government and the government alone. They will force you to comply under the guise of national security. Dystopic.