Sam: GPT-6-Astra just solved 49 out of 50 unpublished theoretical physics problems — problems that hadn't been published anywhere, designed to test genuine research-level reasoning — at a cost of two dollars and forty cents per problem. And in a few cases where the model appeared to get the wrong answer, it turned out the human experts' reference solutions had errors, not the AI's. We need to talk about what that means.

Priya: Welcome to AI Revolution for Friday, October 16th, 2026. I'm Priya Nair.

Sam: And I'm Sam Kim.

Priya: Big show today. We've got that physics result Sam just mentioned, Google's new Gemini 4 Argon and a benchmark discrepancy that should make everyone uncomfortable, Meta dropping an 816 billion parameter open-weight model optimized for Apple silicon, a debate about whether we can actually read neural networks, Anthropic opening its doors to independent auditors, Microsoft betting big on custom inference chips, Bank of America mobilizing thousands to fix AI-discovered vulnerabilities, and AWS releasing tooling for production agent systems. Let's get into it.

Sam: So the GPT-6-Astra physics result. This comes from a benchmark called TheoPhysEval — 50 unpublished theoretical physics calculations, the kind of problems working physicists at top institutions would spend real time on. Astra solved 49 of 50 with accuracy matching leading specialists. The key detail here is "unpublished." These aren't textbook problems. They're not in the training data. This is the model actually performing multi-step theoretical reasoning in physics, not pattern-matching against solutions it's seen before.

Priya: And the cost angle is striking. Two dollars and forty cents per problem. If you're a research group that needs to validate a calculation, or explore whether an approach works before committing weeks of a postdoc's time, the economics of that are transformative.

Sam: Completely. But the researchers made a really honest observation that I think is the more interesting finding. They said physics hasn't had the same kind of AI-driven breakthroughs that mathematics has seen recently, and the reason is that these models lack what they called "taste" — the ability to identify which problems are meaningful to solve. The model can execute the calculation once you point it at the right question, but it can't tell you which questions matter. That distinction between problem-solving and problem-finding is where the frontier actually is right now.

Priya: And that's a meaningful bottleneck. The researchers did say they think autonomous top-level AI research is arriving soon, but "soon" in academic terms can mean a lot of things. Still, 49 out of 50 with some of the "failures" actually being human errors — that's a data point you can't wave away.

Sam: Let's move to Google. Gemini 4 Argon scored 97.5 percent on the ARC-AGI-3 benchmark, making Google the second lab after OpenAI to cross 95 percent on that test. ARC-AGI-3 tests abstract reasoning — pattern recognition, novel problem-solving, the kinds of tasks that are supposed to be hard to brute-force.

Priya: But there's a huge asterisk here. That 97.5 percent was under what the report calls "optimal conditions." Under the standard evaluation harness — the one designed for apples-to-apples comparison — Argon scored 49.8 percent. That's a 48-point swing depending on how you run the test.

Sam: Which tells you something important about where we are with benchmarks. A 48-point gap between optimal and standard conditions means the model's performance is extremely sensitive to prompting strategy, compute allocation, or whatever other knobs they're tuning. For practitioners, the standard harness result is closer to what you'd actually experience in production. 49.8 is a very different story than 97.5.

Priya: Argon also trails competitors in math and software engineering benchmarks. So this is a model that's strong on one specific abstract reasoning test under favorable conditions, but not uniformly at the frontier. Google is clearly competitive again at the top end, but the headline number requires a lot of context.

Sam: Meta released Muse Spark Open this week — 816 billion parameters, multimodal across text, images, video, and audio, with a million-token context window. And the interesting angle is a collaboration with Apple to optimize it for local execution on M5 Ultra chips.

Priya: So let's unpack why that matters. Running an 816B model locally — not through an API, not in someone else's data center — means your data never leaves your hardware. For enterprises with data sovereignty requirements, in regulated industries, or in air-gapped environments, that's a fundamentally different deployment model. You get frontier-class capabilities without the API dependency.

Sam: The M5 Ultra optimization is the enabler. Apple's unified memory architecture lets you fit models this size without the multi-GPU setups you'd need on traditional hardware. Meta and Apple collaborating on this is notable — it suggests both companies see on-device large model inference as a strategic priority. On benchmarks, Muse Spark Open performs roughly on par with the leading Chinese open-weight models from Qwen, GLM, and Kimi, which have been setting the pace in that space.

Priya: Now to something more foundational. There's a new essay circulating in the AI safety research community claiming that approximately 90 percent of neurons in neural networks serve clearly identifiable functions. This is an interpretability claim, and it's generating real debate.

Sam: To give some context — the conventional wisdom has been that neural networks are deeply opaque. Individual neurons might respond to many different things, features are distributed across many neurons, and understanding what a network is "thinking" is extremely hard. If it turns out that 90 percent of neurons have clear, singular roles, that changes the tractability of the whole interpretability problem dramatically.

Priya: Think of it this way. If you're trying to audit a model for safety — looking for deceptive behaviors, or checking whether it might produce dangerous outputs in specific scenarios — there are two approaches. You can test it behaviorally, throwing inputs at it and watching what comes out. Or you can inspect its internals directly, the way you'd review source code. If most neurons are interpretable, that second approach becomes viable. You could potentially inspect a model's circuits and identify problematic patterns before deployment, not after an incident.

Sam: The debate is about whether the thesis actually holds up at scale, in production models, across different architectures. The 90 percent figure is provocative, and some researchers are pushing back hard. But even if the real number is significantly lower, progress on interpretability at this level would be a big deal for safety.

Priya: Speaking of safety and accountability — Anthropic announced it's giving METR, an independent nonprofit focused on model evaluation, employee-level access to its models, training processes, and safety practices. And critically, Anthropic has waived the right to block or modify METR's published findings.

Sam: That "no veto" clause is what makes this meaningful. Plenty of companies invite external reviewers in, but retain editorial control over what gets published. Anthropic is explicitly saying METR can publish whatever it finds — risks, incidents, criticisms of their practices. That's a genuinely different level of transparency.

Priya: And it's happening against the backdrop of an FTC investigation into both Anthropic and OpenAI over whether security incidents have caused harm to Americans. The FTC is reportedly considering fines calibrated to data breach disclosure penalties. So you have voluntary transparency from Anthropic on one hand, and regulatory pressure on the other. These dynamics are going to shape how the industry handles accountability going forward.

Sam: On the infrastructure side, Microsoft announced it's scaling its Maia accelerator to multi-gigawatt deployments starting 2027 as the primary inference engine for its own products. That's a massive commitment to custom silicon independence from Nvidia.

Priya: The technical detail worth noting is the architecture. Maia 300 collapses scale-up and scale-out onto a single Ethernet fabric and enables KV state transfer between prefill and decode workers without offloading to host memory. In current inference architectures, moving that key-value cache around is a real bottleneck. If they've solved that at the fabric level, it could meaningfully improve inference latency and throughput. But practitioners have publicly questioned the tokens-per-watt efficiency claims, so we'll see how those numbers hold up under independent testing.

Sam: Quick hit on a fascinating application story. Bank of America has assigned 4,000 people — 3,000 internal staff plus 1,000 outside contractors — to remediate software vulnerabilities discovered by AI models, including Anthropic's Mythos. This is one of the largest known deployments of AI for offensive security discovery in a regulated enterprise.

Priya: What's telling is the ratio. The AI found the vulnerabilities, but you still need thousands of humans to actually fix them. That's the current shape of AI-augmented security — the discovery scales dramatically, but remediation is still human-intensive work.

Sam: And briefly, AWS released two open-source projects for agent development: Strands Decider 2B, a small model purpose-built for decision-making within multi-step agent workflows, and Strands Box, a sandboxed execution environment for agent code. These target two real production pain points — reliable tool-use orchestration and safe code isolation. Having both from a major cloud provider, open-source, lowers the barrier for teams building production agent systems.

Priya: Looking ahead — the thread connecting several of today's stories is the gap between capability and trustworthiness. We have models solving research-level physics, but we can't verify their reasoning the way we'd check a human's work. We have a benchmark result that swings 48 points depending on test conditions. We have interpretability claims that could transform safety auditing, if they hold up. And we have a major lab voluntarily opening itself to independent scrutiny. The capability curve keeps steepening, and the infrastructure for verifying and trusting that capability is racing to keep up.

Sam: The question I'm watching is whether the interpretability result — even a weaker version of it — combines with the kind of access Anthropic is providing METR to create a real accountability framework. If you can inspect model internals and you have independent evaluators doing the inspecting, that's a fundamentally more robust safety regime than behavioral testing alone.

Priya: And on the physics result — watch for whether other domains start seeing similar performance. The "taste" gap the researchers identified is real, but if problem-solving at this level generalizes across scientific disciplines, the implications for research productivity are enormous, even without autonomous problem-finding.

Sam: That's the show for Friday, October 16th, 2026.

Priya: Show notes and links to everything we discussed are at cleartext.fm. Have a great weekend, everyone. We'll see you Monday.

Sam: See you Monday.