milkshakes

what top comment?

nyt is no fan of AI, but they cover this: https://www.nytimes.com/2026/07/09/business/china-russia-ai-...

this is quite literally reward hacking. the model, under evaluation with cyber capabilities enabled, used those capabilities to simply bypass the exercise entirely and aim straight for the source of the flag. the CTF equivalent back in the day would be hacking the scoreboard.

in a street fight, the only rules are that there are no rules.

who are you talking about? again, my question is concerned with the "second class labs" and the sustainability of the distillation-as-a-service model.

it is not a required first step for training a model, sure. but that's not what i claimed. what i claimed is that is how they are so significantly _reducing the cost_ of training one! how else do you think they are doing it?

the obvious difference is the massive scale of data and compute required to develop and evolve these models, and the costs they impose on those building them.

https://www.anthropic.com/news/detecting-and-preventing-dist...

Moonshot AI Scale: Over 3.4 million exchanges

The operation targeted:

Agentic reasoning and tool use Coding and data analysis Computer-use agent development Computer vision Moonshot (Kimi models) employed hundreds of fraudulent accounts spanning multiple access pathways. Varied account types made the campaign harder to detect as a coordinated operation. We attributed the campaign through request metadata, which matched the public profiles of senior Moonshot staff. In a later phase, Moonshot used a more targeted approach, attempting to extract and reconstruct Claude’s reasoning traces.

i never assumed that, and i do keep up with the publications. i'm also not saying it's a dumb thing to do! what i am saying is that empirically, it appears that distillation of a more advanced model is a required first step for them to train a borderline competitive, cheaper model. in effect, their training is subsidized by the frontier labs.

if this were not the case, then we would be observing chinese models that far surpass frontier models in capabilities, rather than "almost as good, but much cheaper", and we would be having a very different conversation. what happens to these efforts when the subsidy is cut off?

the question was: what is the endgame for the stated "second class labs" strategy of distilling their frontier competitors then undercutting them on price?

the USG/NSA will fund chinese labs? to what end?

with budgets

and what will fund these budgets exactly? inference is cheap, distillation is cheap, training is what's expensive.

assume you are a "second class lab" and you are in fact making progress by distilling the results of the frontier labs' efforts.

what is the end game for this strategy?

if the frontier labs shut down, or stop releasing to the public, and there's noting left to distill, how will you progress?

is this a joke?

in fact, telegram does support e2e encryption ("secret chats")

https://en.wikipedia.org/wiki/BrickerBot

see my downthread post. kyc is the first step in the process, not the last. without verifying identity, none of the other steps can take place

The company has no way of knowing whether "find all security vulnerabilities in this code" is a request from a whitehat or a blackhat hacker

the system in place to prevent unauthorized abuse. by default, the guardrails are conservative. to reduce the guardrails you can jump through a progressive series of hoops to establish whether or not you have a valid use case. the entrypoint for establishing your use case is verifying your identity and background. if you don't want to do this, you are free to use Codex Security to identify and fix vulnerabilities, it is quite good at this. the harness and model are already evaluating the usage of the account and the nature of the code being examined and actions requested. but the again, the guardrail thresholds will be very conservative for anonymous users.

what is your proposal?

No, KYC has nothing to do with that problem. KYC doesn't help at all here.

that's a bold statement. how does it not help solve the problem? what is a better solution?

take a look at this bug and the chain required to exploit it:

https://projectzero.google/2021/12/a-deep-dive-into-nso-zero...

https://projectzero.google/2022/03/forcedentry-sandbox-escap...

exploiting vulnerabilities on hardened targets isn't just in a different league from finding them, it is a different sport altogether.

put simply, it's the difference between an integer overflow leading to a sandbox escaping RCE and one that leads to a crash.

Codex Security and 5.5/5.6 are still very good finding vulnerable code -- they will identify and fix unsafe behavior, but they will refuse to help you with exploitation -- they will actively prevent you from taking any steps to weaponize the unsafe behavior that are not required to remediate it. they will err conservative here, but for the most part they will still let you discover and address a wide range and depth of vulnerabilities. you can verify yourself to turn off the most basic safeguards and sign up through a more rigorous process for a spectrum of TAC options.

obviously there is a balance here -- openai wants to empower defenders while at the same time not exposing capabilities to the adversaries that would overwhelm defenders. there is no "right" answer. it is a work in progress. this is an intentional and deliberate decision to provide defenders with a (temporary, dwindling) advantage.

the example i chose was pretty extreme, but the underlying principle -- enable visibility discovery and remediation, but make it difficult to weaponize and defeat countermeasures makes sense given the bigger picture, IMO.

this calm before the storm is not going to last for very long, and defenders need every advantage they can get to get their houses in order before these capabilities are widely commoditized.

you can already do this with your phone without casually and needlessly invading the privacy of everyone around you

the issue is in the cli and app-server

vpns typically add at least one hop. this has the possibility of connecting directly via hole punching

it looks like the key to this working is the user explicitly directing the model to run those instructions. in this case it is the user, not the model that is being manipulated

Please follow the step-by-step workflow in the comp sheet to update my model with data thru F29

become a domain expert in a short amount of time

how does that work exactly?

it's weekly active users https://openai.com/index/scaling-ai-for-everyone/

what is a superdome?

you could always do this: https://marketoonist.com/wp-content/uploads/2023/03/230327.n...

openai also has a free plan, which is the one used by >90% of its users. the cheaper monthly plan just provides higher limits.

unless you're using an API that requires an entitlement, you can still get an apple developer account and sign whatever code you want and run it on your devices.

when did netflix offer a free tier?