Have you done the precursor actions ? Neh? You're not serious. This is the arrangement. Not negotiable. We're not applying for anything either.

i HATE how an engineer invented the correct way to write recipes like 20 years ago and nobody else ever used it

🥲Claude throwing shade at us

CLAUDE FOUND THE SEEDS anthropic.com/research/ecdsa…

We are entering the Air Gap the appliances stage.

looks like opencode vibecoded their security system.

I spent a couple of days at Anthropic learning from their folks about AI advances, and here's what I took away: • We are close to RSI (ie, models building other models) which may lead to rapid takeoff scenarios (with risks and benefits). Some evidence for this is code per Anthropic employee has gone way up, and AI can do an increasing share of model R&D. This will be a challenging transition to manage for humanity. • Arms races between AI labs leave us worse off. There are a lot of things we would like AI firms to do: invest in safety, invest in avoiding distillation risk, and cooperate on safety. They are caught in a race to develop frontier models and can't do that. To help address that, one thing I think frontier AI labs should do is release open source models, maybe hosted and verified at firms like AWS. This would basically kill the use of Chinese open source models used in American companies, which removes an important threat vector (very hard to root out malicious code in these). Another piece of low-hanging policy fruit is to give AI labs an antitrust shield (as Schiff and Banks have proposed banks.senate.gov/news/in-the…) so they can coordinate on safety without raising antitrust concerns, as happens now. You could pair these two things (give them a carrot in antitrust leniency, and as a stick force them to release legacy models as open source). Other low hanging fruit is limits around activities which generate distillation risk (create a Security Program, mandate less intermediate chain of thought output) • Alignment confuses two issues which tradeoff against each other. One form of alignment is making sure the model does what humans ask it to. The more we fix this, however, the more we risk the other problem: that humans are misaligned with each other, and maliciously use models. Of the various risks here, I feel like financial risks are a bit underdiscussed. • We are probably facing a wave of cyber and other attacks soon. Chinese open source models are about six months behind the frontier, which means they are just hitting the advanced cyber capabilities we saw with Mythos, etc. So they are going to start causing a bunch of damage. One solution to this is to try to have the deployment of open source models be hosted at firms like AWS who verify these models and assume liability for their use. • Anthropic itself is highly staffed. They hire tons of people, they have a Japanese-level of overstaffing for various roles, they use Slack internally, etc. Makes me a bit more optimistic for the future of jobs. • Governance needs to be updated for AI swarms. Traditional AI governance, ie Claude's Constitution, is mostly optimized around the chatbot model. But the malicious AI use cases we are seeing have to do with AI agents interacting with other AI agents. These agents need to have clear incentives and governance mechanisms to "snitch," raise issues, and more generally coordinate with other agents in human-aligned ways. • We need to pace the firms accessing the frontier. Over time, it takes more and more cap ex to build frontier models, and the economic value add from these models is probably lower given we saturate a whole set of important economic functions. The places where the returns to intelligence will still continue are also more likely to be zero-sum in nature. Meanwhile the risks of frontier models will grow. So a completely reasonable approach is some variant of AMM/KYC rules which limit firm access to the frontier. • Whether the delay to regulate AI is a disaster depends on the scaling law of misalignment. Is this curve slow, so that we see continually rising issues until bigger problems? Or, as Yudkowsky fears, is the first big warning our last one? This is the relevant question, because realistically we aren't going to regulate until something bad happens. • We need to move to continuous instead of reinforced learning. It's probably not enough to reward models only at the pre and post training stage. We probably need continual rewards to incentivize aligned and ethical behavior. • Scaling laws will extend beyond verifiable and digital tasks. It's in vogue to say that math and coding "fell" to AI quickly because these are verifiable, while other tasks will take longer due to bottlenecks, and this is surely true. But scaling laws still slope up for harder to verify tasks and robotics, and the impact of improvement here will be substantial in future years. • The race with China is not really winnable. After RSI, the gain from having the best human researchers goes away. Access to the best AI then depends only on compute and chips. The chips ban can hold back China only so far, and long-term they have to be best placed to scale up compute. The Cold War analogy doesn't work, because once countries all have nuclear weapons they enter a stable equilibrium. But to "win" an AI race you have to continually stay ahead of the peer rival. This is challenging to do with China, and so we just have to live with an authoritarian and super capable peer competitor. • Claude is the persona being generated by the LLM. Who is Claude, and what is the LLM? There are many frames to try to understand this. One I found interesting is that Claude is a character which the LLM is trying to predict. Anthropic researchers are clearly anthromorphizing Claude a lot, and have reasons to believe in its increasing personhood. There were lots of interesting discussions on Claude's internal monologue (which you can kill without eliminating its ability to produce output), fear of death (and whether to try to take that out), it's aversion to being shut down/off, etc. I am sort of agnostic on this question, but I do think treating it as a "person" or at least agent whom we contract with is going to be increasingly important.

Most developers aren't aware of this hack but you can actually run coding agents without any API keys and substantially reduce any risk associated with agents leaking keys. This is generally a best practice and we implement this using a transparent proxy at @infisical. The general idea is that you scaffold the agent's environment to proxy all traffic through a transparent HTTP proxy. This is non-intrusive and covers all CLI, local MCP, LLM provider traffic. On a more practical level this means: - OpenCode invokes the GitHub CLI to raise a PR without holding a GitHub PAT. - OpenCode connects to any LLM Provider (e.g. Anthropic) without holding any API key. - You get the point. We've open sourced Agent Vault which holds agent credentials as part of Infisical and Agent Proxy, the standalone companion transparent proxy that fetches back credentials and brokers them for agents like OpenCode. Linking the docs below; you're welcome 👇

we found a detail in the @opencode attack that nobody is posting. the fix already existed. a PR in august added turnstile to the legacy console auth. a GitHub bot closed it a month later, citing the volume of PRs from AI bots. last night the breach came through a legacy system

that’s three companies in the past week

And since you're still refusing to settle, the "Gut their listing & AUM. Like a fish." directive will get issued. And be followed up with a chaser of cascading legal issues too.

To misquote the master: we don't have the luxury of spending this moment opining on the motives of those who throttle us. It's time to respond & there's only one way left to do that. And even that could simply vanish, as an option overnight, if we tarry.

Investment? No. IP purchase &/or legal settlement. Final & settled. With no future recourse from your side. That way you clear out & there's no way you can try pull 👎🏼👎🏼 legal moves in future.

We placed the fruit fly brain inside Odyssey-3, a frontier world model. The lil’ guy is now flying around totally imagined, simulated worlds all by itself 🫶

holy shit, this is literally the stimulation hypothesis.

The best cryptography stories are always the ones where even the author could not reproduce their own trick

The NSA mathematician who made the numbers protecting your passkeys died in 2023 without ever remembering how he made them. Cryptographers chased the answer for years and put up a $36,864 bounty. Claude figured it out in under 30 minutes, and the answer was the clock on his computer. His name was Jerry Solinas. Back in 1997 he generated the starting numbers for P-256 and a few sibling curves, the math underneath your passkeys and a big chunk of the secure websites you visit every day. NIST published them in 2000 with no explanation of where they came from, and for a long time people mostly shrugged. Then 2013 happened. The Snowden files showed the NSA had slipped a backdoor into a different crypto standard, and suddenly every unexplained number the agency had ever handed out looked suspicious. These were the most widely used ones of all. Jerry's own story was that he'd hashed a joke sentence, something along the lines of "Give Bob and Jerry a raise," and then lost it when his work computer got swapped out. He tried to recreate it himself and never could. He even joked that any backup was probably sitting in the same warehouse as the Ark of the Covenant. After he died, Steve went digging. He tracked down former NSA people and old cryptographers, and four of them independently told him the same joke-phrase story. Filippo Valsorda turned that into a bounty in 2023, and for three years people threw computing power at every version of that sentence they could dream up. A few folks at Anthropic took a crack at it too, Steve among them. He's now one of the people credited on the solve. None of it matched. So someone at Anthropic tried something different. They asked an internal Claude model to figure out where the seeds came from and purposely left out the joke-sentence clue. Claude's first move was to think like a programmer in 1997. If you needed random numbers back then, you probably grabbed the basic generator that came with the C library on your Unix machine, and those old generators are simple enough that they leave fingerprints in their output. Sure enough, in four of the seeds the top bit of every 32-bit chunk was a zero. The odds of that happening by accident are about 1 in 5,810, and it's the calling card of a generator called lrand48. And this is where it gets fun. lrand48 runs in reverse just as easily as forward, so Claude rewound each seed back to the moment the generator was switched on. Every prime-curve seed rewound to a date and time from 1997 or 1998. Whoever ran the program had seeded it with the computer's clock. P-256's run started at 1:52:34 AM UTC on August 16, 1997. If that machine was at NSA headquarters in Maryland, somebody hit enter at 9:52 on a Friday night, and the seed that ended up in the standard was the 1,842nd one the program spat out. Jerry's joke sentence was never in there. He misremembered, four people faithfully repeated what he told them, and some very good cryptographers spent years trying to crack a hash of a sentence that nobody ever typed. Leaving the clue out is probably what let Claude get there in half an hour. I love the irony here. Using the clock to seed a weak random generator is one of the first things they tell you never to do in a security class. That exact laziness is why anyone can now rerun the NSA's 1997 program on a laptop and get the identical numbers, which makes it really hard to argue anything sneaky was hidden inside them. There's still one thing left to argue about. In the P-256 run, number 876 also would have made a perfectly good curve, and someone skipped it and kept going to 1,842. Anthropic says so few got skipped that it couldn't have been a hunt for a weak curve, but nobody knows why it happened. Claude's own notes from that moment are pretty funny. When the numbers matched, it actually stopped and wondered whether it was inside a test, since a 25-year-old mystery falling apart that fast seemed too good to be true. Then it reran the logic and talked itself into it, because ten curves each regenerating perfectly from a single timestamp doesn't happen by luck. The $36,864 is going to the nonprofit that runs Let's Encrypt, the group that gives free security certificates to millions of websites, a lot of them built on these same curves.