Sam: A new benchmark just put a number on something a lot of us have felt intuitively. When you take the best AI models available today and give them actual knowledge work — not synthetic puzzles, not isolated coding challenges, but the kind of multi-step, context-heavy tasks that professionals do every day — they fully complete about three percent of them. Three percent. That's a number worth sitting with, and we're going to dig into what it actually means.

Priya: Good morning, and welcome to AI Revolution for Friday, June 19th, 2026. I'm Priya Nair.

Sam: And I'm Sam Kim.

Priya: We've got a packed show today. Beyond that sobering benchmark result, OpenAI has a genuinely interesting new alignment approach that works differently from what we've seen before. Google DeepMind published a security framework that treats AI agents like insider threats. There's a messy situation with Anthropic, export controls, and SK Telecom. Amazon is making a big move to sell its custom chips externally. A startup claims to have cracked the quadratic attention bottleneck. And we've got FERC fast-tracking data center grid connections, plus some notable funding and hiring news. Let's get into it.

Sam: So this benchmark. The key insight is methodological. Most AI benchmarks test isolated capabilities — can the model write a function, summarize a document, answer a factual question. This one, and we'll link to the full paper, constructs tasks that mirror what a real professional actually does. That means you're dealing with ambiguous requirements, multiple tools, information scattered across different sources, and judgment calls about what "done" even means. The model has to plan, execute across steps, handle errors, and produce output that actually satisfies the intent of the request.

Priya: And three percent completion tells us something specific. It's not that the models can't do any part of the work. They can often handle individual subtasks competently. The failure mode is in orchestration — maintaining coherent intent across a multi-step workflow, recovering when something doesn't work as expected, knowing when to ask for clarification versus when to make a reasonable assumption.

Sam: Right. If you've worked with AI agents in production, this probably matches your experience. The model is great at the atomic operations but falls apart at the workflow level. And this has direct implications for how teams should be scoping AI automation. If you're expecting end-to-end task completion on complex knowledge work, you're going to be disappointed. If you're using AI to accelerate specific subtasks within a human-managed workflow, you'll get much more value. That gap between benchmark scores on isolated tasks and this three percent figure on realistic work — that's the gap teams need to plan for.

Priya: Let's shift to something more encouraging. OpenAI published research on what they're calling beneficial trait training, and the mechanism here is genuinely interesting.

Sam: So the core idea: instead of training a model on a long constitution of rules — which is roughly how Anthropic approaches alignment — OpenAI used reinforcement learning to optimize for specific behavioral traits. Things like truthfulness and corrigibility, which is the model's willingness to be corrected and to defer to human judgment. What's surprising is the generalization. They trained these traits using health domain data, and the improvements transferred to completely unrelated domains. The model got better at detecting deception even in contexts that had nothing to do with health.

Priya: Walk me through why that might work. Why would training on truthfulness in one domain transfer?

Sam: The hypothesis is that these traits correspond to something like internal representations of honesty and deference that aren't domain-specific. When you reinforce truthfulness in a health context, you're not just teaching the model medical facts — you're strengthening whatever internal circuitry the model uses to distinguish between "I know this" and "I'm confabulating." And that circuitry gets applied everywhere. They saw improvements on 44 out of 53 benchmarks, which is a broad signal.

Priya: What makes this distinct from constitutional AI is that it's optimizing for behavioral properties directly rather than teaching the model to evaluate its own outputs against a set of principles. Both approaches seem to work. Having multiple viable paths to alignment is a good thing for the field.

Sam: Absolutely. And the small-dose aspect matters. They didn't need massive amounts of training data for this to work. That suggests it could be relatively cheap to apply, which matters for adoption.

Priya: Next up, Google DeepMind published what they're calling an AI Control Roadmap, and the framing is notable. They're treating their own AI agents as potential insider threats.

Sam: This is a security framework, and it's structured around a specific principle: tie your security controls to measured agent capabilities. As an agent gets more capable, it gets more access but also more monitoring and containment. Think of it like how you'd handle a new employee with escalating privileges. They analyzed a million coding tasks done by their agents and found that most problems aren't the agent trying to do something malicious — they're the agent being overzealous. It takes an instruction, interprets it too broadly, and takes actions that go beyond what was intended.

Priya: That matches the benchmark story. The failure mode isn't malice, it's poor judgment at the workflow level. DeepMind's framework essentially says: assume the agent will sometimes do the wrong thing, build containment around that assumption, and scale controls with capability. They're also warning that the window for establishing global standards around this is closing, which is a pointed message.

Sam: For anyone building agentic systems, this paper is worth reading closely. It's one of the more operationalizable frameworks I've seen.

Priya: Now let's talk about Anthropic and export controls, because this situation is a mess. Anthropic still cannot distribute Claude Mythos or Fable 5. The trigger was alleged ties between SK Telecom and China. SK Telecom had access to Mythos through Anthropic's partner program, Project Glasswing, and when US officials flagged the China connection, the White House stepped in.

Sam: What's striking is the opacity. There's no clear public explanation of what specific rule Anthropic violated. The export control framework is being applied in real time, and the rules appear to be written as they go. Anthropic didn't knowingly distribute to a sanctioned entity — they had a partner relationship with a major South Korean telecom company, and the geopolitical risk surfaced after access was granted.

Priya: This has real implications for any company distributing frontier models. Partner vetting now has to account for third and fourth-order geopolitical relationships. If your distribution partner has business ties that might alarm US national security officials, that's your problem. And the lack of clear rules makes it nearly impossible to know in advance where the lines are.

Sam: It also creates competitive asymmetry. If the rules are ad hoc, companies with better political connections navigate them more successfully, regardless of actual security risk.

Priya: Let's move to infrastructure. Amazon is making a significant move — AWS is in active talks to sell its custom Trainium and Inferentia chips to third-party data centers. Not just offering them through AWS cloud, but actually selling the silicon externally.

Sam: Andy Jassy has framed this as a fifty-billion-dollar revenue opportunity. Strategically, it's a direct challenge to Nvidia's dominance in the external accelerator market. Until now, if you wanted to run your own AI infrastructure, Nvidia was essentially the only serious option at scale. AMD has been making inroads, but Amazon entering as a chip seller — with the manufacturing scale they can bring — changes the competitive dynamics. For enterprises, more hardware sourcing options means more pricing leverage.

Priya: Meanwhile on the regulatory side, FERC has mandated that grid operators give AI data centers a fast-track interconnection process. This removes one of the key permitting bottlenecks that's been slowing compute buildout. But — and this is important — it doesn't address the underlying electricity supply shortage. You can get connected to the grid faster, but if there isn't enough power on the grid, you're still waiting.

Sam: It's solving the paperwork problem but not the physics problem.

Priya: Exactly. Quick hit on Subquadratic, a Miami startup that claims to have solved the quadratic attention bottleneck in transformers. Sam, give people the thirty-second version of what quadratic attention means.

Sam: In standard transformer attention, every token attends to every other token. So if your sequence length doubles, compute cost quadruples. That's the quadratic scaling. It's why long context windows are so expensive. Lots of researchers have proposed subquadratic approximations — linear attention, sparse attention — but they all involve tradeoffs in quality. Subquadratic claims to have solved this without those tradeoffs. They came out of stealth last month with thin details, and the community was skeptical. They're now sharing technical evidence.

Priya: If this holds up, the implications for inference cost and model scalability would be enormous. But "if it holds up" is doing a lot of work in that sentence. We should watch for independent reproductions.

Sam: Agreed. Extraordinary claims, extraordinary evidence.

Priya: A couple of industry notes. Baseten, the AI inference startup, is reportedly closing a one-and-a-half-billion-dollar round at a thirteen-billion-dollar valuation, months after its last mega-round. The inference infrastructure layer is attracting massive capital right now.

Sam: And OpenAI hired Noam Shazeer from Google DeepMind — that's a co-inventor of the original Transformer architecture — plus Dean Ball, a former Trump administration AI policy official. Both hires in the same week, clearly pre-IPO positioning on both the technical credibility and regulatory fronts.

Priya: Looking ahead, I think the thread connecting today's stories is the gap between capability and readiness. Three percent on realistic knowledge work. Agents that are overzealous rather than competent. Export controls being invented on the fly. The technology is powerful but the systems around it — the workflows, the security frameworks, the regulatory structures — are still catching up.

Sam: And the responses to that gap are diverging. DeepMind is publishing operational security frameworks. OpenAI is investing in alignment techniques that might generalize. The federal government is making ad hoc decisions. Amazon is trying to break hardware bottlenecks. Everyone sees the same gap, but they're approaching it from completely different angles.

Priya: What I'm watching is whether the alignment work and the security frameworks develop fast enough to match the capability growth. That three percent number will improve. The question is whether our ability to deploy AI safely and predictably improves at the same rate.

Sam: And whether the quadratic attention problem actually gets solved. If it does, the capability growth accelerates, which makes everything else more urgent.

Priya: That's our show for today. Show notes and links to all the stories we covered are at cleartext.fm.

Sam: Have a great weekend, everyone. We'll see you Monday.