Sam: GPT-5.5 Instant just became the default model for every ChatGPT user. OpenAI is claiming 52.5% fewer hallucinated claims on high-risk topics — medicine, law, finance. That number is from internal testing, so we should hold it loosely, but the direction is meaningful. And there's a new feature called "memory sources" that shows you exactly which stored context shaped a given response. That transparency piece is actually underappreciated. We'll get into why.

Priya: Welcome to AI Revolution, Wednesday May 6th, 2026. I'm Priya Nair, joined as always by Sam Kim. Today is a dense one — we've got a major model default swap from OpenAI, a $200 billion cloud commitment from Anthropic that's worth pausing on, Google's new TPU architecture designed specifically around agentic workloads, a significant expansion of the Pentagon's AI supplier list, and a sobering essay from an Anthropic co-founder about recursive self-improvement. Let's get into it.

Sam: So let's start with GPT-5.5 Instant. The headline number is that hallucination reduction figure, and I want to spend a minute on what actually makes hallucinations hard to reduce, because it contextualizes why a 52% improvement — if it holds up — is genuinely notable.

Priya: Right, because hallucinations aren't a bug you can just patch. They emerge from how these models generate text probabilistically.

Sam: Exactly. The model is always doing next-token prediction based on learned distributions. When it doesn't have reliable information, it still produces fluent-sounding output — it doesn't know to stop and say "I'm uncertain here." So reducing hallucinations requires either better calibration of the model's own uncertainty, or better retrieval mechanisms that ground responses in verifiable sources, or both. The 5.5 Instant architecture almost certainly involves improvements on both fronts, though OpenAI hasn't published the technical details.

Priya: The high-risk topic framing matters here too. Medical and legal queries are particularly dangerous because users may act on wrong answers. The model producing less confident-sounding wrong information in those domains is a different kind of improvement than general fluency.

Sam: And then there's the memory sources feature, which I think is actually a bigger deal than it's getting credit for. Previously, if ChatGPT was drawing on something you'd mentioned six months ago, you'd have no idea. Now it surfaces which stored context influenced a response. That's a meaningful step toward auditability.

Priya: From a systems perspective, personalization introduces a real traceability problem. You're getting an output that's partly a function of your history, and without visibility into that, you can't understand why you got a different answer than someone else asking the same question.

Sam: The Gmail integration and file personalization being limited to Plus and Pro users for now makes sense as a staged rollout — that's a more sensitive data surface. But the direction is clear: the model is moving toward something that has persistent context about you across many sources.

Priya: Let's move to the Anthropic-Google Cloud number, because $200 billion over five years is a figure that requires some processing.

Sam: It's more than 40% of Google Cloud's entire committed backlog. From a single customer. And when you put OpenAI and Anthropic together, they account for roughly half of two trillion dollars in committed cloud revenue across Amazon, Microsoft, Google, and Oracle combined.

Priya: Both of those companies are currently operating at a loss. So the question that number immediately raises is: does the revenue trajectory actually support this level of compute spend?

Sam: The implicit bet is that the capability improvements from scaling justify the cost — that you train at this scale, the models get meaningfully better, and you monetize that through enterprise contracts and API usage. The uncertainty is whether the monetization curve catches up fast enough.

Priya: And what it tells you about the competitive dynamics. If you can't access this level of compute, you can't train at the frontier. That's a structural constraint that shapes who stays in the game.

Sam: The concentration is striking. And it means the cloud providers have extraordinary leverage over the labs, even as the labs are individually enormous customers.

Priya: Now Google's new TPU generation, which was also announced this week, is relevant here because it shows what they're building to serve exactly this kind of demand.

Sam: What's technically interesting about the 8th-gen TPUs is that Google designed two distinct chip variants rather than a single general-purpose accelerator. One optimized for large-scale training runs, one optimized for agent inference workloads.

Priya: That split is meaningful. Training and inference have always had different compute profiles, but the agent case makes inference genuinely harder in a specific way.

Sam: Right. A standard inference request is roughly: take input, run forward pass, return output. An agentic workflow is: take input, generate a reasoning step, call a tool, process the result, generate the next step, potentially spin up sub-agents, loop. The memory access patterns are different, the latency requirements are different, and you often have multiple models in the loop simultaneously.

Priya: So optimizing a chip for that kind of continuous multi-step reasoning is a different design problem than optimizing for throughput on batch inference.

Sam: And it signals that Google is treating agentic AI as a first-class workload, not just a scaled-up version of chatbot inference. That's an architectural commitment.

Priya: Let's cover the defense and government stories together because they're interrelated. The Pentagon expanded its approved AI vendor list — Microsoft, Amazon, Nvidia, and a company called Reflection AI now join OpenAI, xAI, and Google for use in classified operations.

Sam: Reflection AI is the notable name there. They haven't released a publicly available model. Being approved for classified Pentagon use before you've shipped anything public is an unusual sequence.

Priya: Separately, five labs now have formal agreements with the Commerce Department's Center for AI Standards and Innovation — CASI — to provide pre-release models with reduced safety guardrails for classified security testing. Google DeepMind, Microsoft, and xAI joined Anthropic and OpenAI in those agreements this week.

Sam: The reduced guardrails piece is important to understand correctly. The purpose is red-teaming — you want to probe what a model will do without the standard refusals in place so you can characterize its actual capabilities and risks. That's a legitimate evaluation methodology. The concern is always about how those findings are used and who has access.

Priya: And it establishes a formal government-industry pipeline for evaluating frontier models before they're publicly deployed, which is a structural change worth tracking.

Sam: One counterpoint to all of this expansion: Google DeepMind's UK staff voted to unionize this week, with the primary stated motivation being to block the use of DeepMind's models in military applications. So at the same time the US government is broadening its AI defense supply chain, there's organized labor pushback from within one of the key labs.

Priya: That tension is real and it's not going away. Anthropic also put out ten preconfigured AI agents for the financial sector this week — investment banking, asset management, insurance. Research, risk assessment, compliance, accounting. It's part of an IPO preparation story, but technically it represents something worth noting.

Sam: The move from model APIs to preconfigured domain agents is significant. Instead of a customer building their own workflow on top of an API, you're getting a template that encodes specific task logic. That lowers the barrier to deployment but also means the lab is making architectural choices about how the agent reasons through, say, a compliance check.

Priya: Which puts a lot of responsibility on getting those templates right for regulated domains.

Sam: Before we close, we should spend a few minutes on Jack Clark's essay. He's a co-founder of Anthropic, so this isn't an outside observer speculating.

Priya: His argument is that the building blocks for AI systems training their own successors are largely already in place — automated data generation, automated evaluation, the ability to run training pipelines at scale. And he puts 60% probability on recursive self-improvement being meaningfully underway by end of 2028.

Sam: What makes the argument technically grounded rather than speculative is that he's pointing to existing components rather than hypothetical breakthroughs. Automated data synthesis is real. Automated evaluation is real. Models contributing to their own training signals is real in limited forms already.

Priya: The gap he's focused on is supervisory capacity. Humans can evaluate model outputs when models operate at roughly human level. When models are generating code, reasoning, or research at significantly superhuman speed and volume, the oversight loop breaks down.

Sam: And that's a genuine systems problem. It's not about whether the models are aligned or misaligned in some abstract sense — it's about whether the feedback mechanisms humans use to detect and correct errors can keep pace with output volume.

Priya: He's not saying this is inevitable or catastrophic. He's saying the probability is high enough that it warrants serious preparation now. Coming from someone who helped build the safety infrastructure at Anthropic, that framing carries weight.

Sam: What's the thread across today's stories, to bring it together?

Priya: Scaling is accelerating on every dimension simultaneously — model capability, compute investment, hardware specialization, government adoption. And the oversight and governance infrastructure is running behind all of it.

Sam: The memory sources feature in GPT-5.5 is a small example of what auditability looks like in practice. The CASI pre-release testing agreements are an attempt at government-level oversight. Jack Clark's essay is pointing at the harder version of that problem at capability scales we haven't reached yet but may be approaching faster than expected.

Priya: That's the question to keep watching: not just what the models can do, but whether the systems we have for understanding and verifying what they're doing are keeping up.

Sam: That's the show for today. Show notes and links to everything we covered are at cleartext.fm. We'll be back tomorrow.

Priya: See you then.