Today in AI — Aug 7

[THEO] Alright, Mara, you know I love a good puzzle, and this week in AI, we've got a couple of big ones, particularly around security and, well, how smart these models are getting. The big news making waves has to be Anthropic's Claude, and its role in knocking a quantum-resistant cryptography algorithm out of the running.

[DR. MARA] Indeed, Theo. It’s a significant development, especially in the realm of post-quantum cryptography, or PQC. For those unfamiliar, PQC is about developing cryptographic systems that can withstand attacks from future quantum computers, which theoretically could break many of our current encryption standards. The National Institute of Standards and Technology, or NIST, has been running a multi-year competition to standardize these new algorithms. This particular algorithm, known as a third-round candidate, was under consideration for official U.S. standardization.

[THEO] So, picture this: NIST is running this high-stakes competition, like a scientific Olympics, to find the strongest, most resilient encryption methods, right? And then Claude comes along, takes a look at one of the top contenders, and finds a weak spot. It’s like a digital X-ray vision, exposing a hidden flaw. What exactly did Claude do here?

[DR. MARA] Well, researchers at Anthropic used a preview version of their Claude Mythos model to discover improved ways to attack cryptographic algorithms. In this specific instance, it helped uncover a critical flaw in one of these PQC candidates. This flaw was significant enough that it rendered the algorithm broken, meaning it could no longer be considered secure against potential attacks. It highlights the increasingly sophisticated role AI models can play in security auditing and cryptanalysis. It's not that Claude *invented* a new attack; rather, it rapidly and effectively explored the attack surface of the algorithm and identified vulnerabilities that humans or traditional computational methods might have missed or taken far longer to find.

[THEO] So it's less about the AI being a mastermind hacker, and more about it being an incredibly diligent and fast-acting debugger, sifting through possibilities to find the chink in the armor? That makes sense. And speaking of clever AI, OpenAI has been quietly teasing its next major model, Astra, which apparently just solved ten long-standing math problems.

[DR. MARA] That's correct. OpenAI revealed that an internal version of Astra, their upcoming model family, produced new results for ten previously unsolved problems in mathematics, quantum complexity, and theoretical computer science. What's particularly notable is the claim that these solutions were generated using a relatively modest amount of computational resources – around $2,000 worth of tokens at their API rates.

[THEO] Two thousand dollars to crack ten major math problems? That's like hiring a super-genius tutor for pocket change. And it wasn't a big, splashy announcement either; some people are noting it was almost "smuggled" into a blog post about math. It feels like a very understated flex, doesn't it?

[DR. MARA] It does. The discourse online suggests a mix of awe and a certain amount of… well, suspicion, given the understated nature of the announcement. Some are arguing this is a significant step forward for scientific reasoning capabilities in AI, while others are questioning the precise nature of "solving" these problems and whether the proofs are fully novel or merely re-derivations. The emphasis from OpenAI is on making better intelligence more accessible, as they’ve also updated their public-facing ChatGPT models.

[THEO] So the takeaway is, AI is getting sharper, not just in finding flaws but also in creating solutions. And while Demis Hassabis is stepping into a new role to focus on long-term strategy at Google DeepMind, and Sam Altman is maintaining an optimistic outlook about pushing boundaries, it’s clear the frontier labs are continuing to push these capabilities in surprising ways. It’ll be fascinating to see how these newly discovered math solutions hold up under scrutiny.

[DR. MARA] Agreed. The peer review process for these mathematical proofs will be the real test.