Today in AI — Aug 6

[SOFIA] Welcome back to The Dish! We're diving straight into what's cooking in the AI world, and it's been a busy few days. There are some really interesting developments at the frontier labs, especially around new models doing some pretty unexpected things.

[DANIEL] Indeed. The headlines are dominated by a couple of big announcements, and it's worth sifting through the excitement to see what's truly novel versus what might be more incremental.

[SOFIA] Absolutely. And for me, the most striking news this week has to be Anthropic’s Claude Mythos. They've been using it to find weaknesses in cryptographic algorithms. We're talking about quantum-resistant cryptography here – the kind of encryption designed to withstand attacks from future quantum computers.

[DANIEL] That’s right, Sofia. The context here is that national cybersecurity agencies, like NIST in the US, have been actively looking for new cryptographic standards because current methods could be vulnerable to quantum computing. They've been vetting a number of candidates for these next-generation algorithms for years.

[SOFIA] Right, and a major one, a third-round candidate, has now been essentially "broken" with the help of Claude Mythos. This isn't just a theoretical crack; it's significant enough that this particular algorithm is reportedly out of the running for standardization. That's a pretty big deal for national security and for the future of secure communication.

[DANIEL] It certainly is. When we talk about finding weaknesses in cryptography, we're looking for subtle mathematical or structural flaws that allow an attacker to bypass the intended security without having to brute-force every possible key. This is a highly specialized field, requiring deep expertise.

[SOFIA] So, for Claude Mythos to contribute to identifying these vulnerabilities, it really suggests a leap in its analytical capabilities, especially in a domain as abstract and rigorous as cryptography. It's not just generating text; it's performing a kind of complex, adversarial reasoning.

[DANIEL] My immediate thought goes to the specifics of *how* it did this. Was it generating new attack vectors from first principles, or was it more about efficiently sifting through existing knowledge and combining known techniques in novel ways? The details of the methodology are crucial for understanding the true "intelligence" exhibited here. Without those specifics, it's hard to gauge the depth of the breakthrough.

[SOFIA] A fair point, Daniel. But it's still impressive. Shifting gears slightly, OpenAI also made some waves with their next major model, Astra. They're teasing that an internal version has solved ten long-standing math problems.

[DANIEL] Yes, and there's been quite a bit of discussion online about the nature of these problems and the proofs generated. The claim is that Astra not only solved them but also published the proofs, and that the token cost for these solutions was reportedly quite low, around $2,000.

[SOFIA] This isn't about simple arithmetic, but rather complex, open problems in mathematics, quantum complexity, and theoretical computer science. The implication is that Astra is capable of high-level, abstract reasoning and generating verifiable mathematical proofs.

[DANIEL] The community is, predictably, very interested in seeing these proofs and understanding the rigor of the solutions. The term "solved" in mathematics can sometimes mean different things, and the true measure will be in the peer review of these proofs. The discourse online is definitely highlighting that the devil will be in the details of the mathematical rigor and originality. It's a bold claim, and the scientific community will want to see the workings.

[SOFIA] Absolutely. And finally, on the commercial side, OpenAI also announced some pretty significant price cuts for their GPT-5.6 models – Luna and Terra – with Luna seeing an 80% reduction. It seems they're pushing the frontier not just in capability but also in accessibility.

[DANIEL] That’s a clear move to increase adoption and usage. Lowering the cost of access to these advanced models could significantly expand the types of applications and research that can leverage them. It reflects a competitive landscape where cost efficiency is becoming as important as raw capability.

[SOFIA] For sure. It's a fascinating time when these models are not just getting smarter but also more accessible. We'll be keeping a close eye on those proofs from Astra and the full implications of Claude Mythos's cryptographic detective work. Up next, we're diving into some new research on engineering bacteria for environmental remediation. Stay with us!