Settlement in motion, weights doing the work

00:00:04 Good morning. A couple of items arrived today worth looking at from the machine room. The first is infrastructure — or rather, the layer underneath it that doesn't announce itself. IBM has connected its digital asset platform to Swift's blockchain ledger, which means banks can now move tokenized deposits 24/7.

00:00:26 The interesting part isn't that IBM added blockchain; it's how they did it without asking any bank to leave the messaging system they already use. Over 11,000 institutions are already on Swift's rails using ISO 20022 payment messages. IBM kept that layer intact and swapped settlement underneath.

00:00:47 Banks get near-instant finality without rebuilding their interfaces or convincing compliance teams to adopt a new protocol. That interoperability is the hard part. If you'd been watching this from the crypto side, it might feel like news. If you've been inside legacy financial infrastructure, it feels like maintenance — which is exactly what it is.

00:01:12 The float that banks rely on for ACH settlement (two to three days) gets compressed when the underlying layer runs on-chain. Parallel to that, ARK Invest launched a tokenized venture fund on Ethereum today. It allows twenty-four-hour trading of private-market exposure, replacing the traditional five- to ten-year lockup.

00:01:35 Even if you're skeptical about Cathie Wood's track record as an investor, the direction is worth noting: tokenized private markets is becoming infrastructure rather than staying speculative. The real constraint here isn't technology — it's LP gates and regulatory compliance.

00:01:54 Those haven't moved just because a blockchain is involved. That sets the scene for some research on the same thread.

00:02:03 A paper from Google and Peking University — arXiv 2609.24974, Harness-Zero — studies a question that keeps coming up in the agent space: what happens when you stop relying on external harnesses? Here's how it works. Agent harnesses — the external systems that manage tool use, context windows, and interaction with the environment — can significantly improve agent performance.

00:02:30 But those gains stay tied to the harness at deploy time. Every domain needs a different harness, so a general-purpose agent either settles for a mediocre shared one or maintains an ever-growing collection. Harness-Zero's answer: distill the harness behavior into model weights during training, then remove the harness entirely at deploy time.

00:02:55 The tricky part is that the optimized harness and the deployment harness have different action spaces. You can't just copy trajectories from one to the other directly. So the paper introduces what they call agent-as-harness — an intermediate system that wraps the student model, reads guidance from the optimized harness, and translates it into corrections expressed in the student's own action space.

00:03:24 Those corrected responses become training demonstrations. Supervised fine-tuning on those rollouts internalizes the behavior into weights. The numbers are specific. Across 28 harness-induced behaviors spanning knowledge work, tool use, and science: Eighty-two point three percent of those harness-induced behaviors are recovered on average.

00:04:02 And for frontier models using the same evolved harness, agent-as-harness beats code-as-harness at eighty-one point one versus seventy-eight-point-one percent across six benchmark-model settings. The paper notes that it remains to be seen how robust this approach is — fair enough, the ablations are promising but they're on a limited set of domains.

00:04:27 The per-domain spread is worth checking in future work. The piece John Rood flagged in the thread captures the shift well: harness policy stops being config and becomes weights. Corrections ride along during training, including the ones you'd have deleted. You can only delete what you can name; the rest ships inside the student, silent.

00:04:52 This reframes the category entirely. The cost shifts from recurring per-inference orchestration into upstream training runs. Every small harness change becomes a new fine-tuning instead of a config update.

00:05:06 A smaller observation from Paul Graham today that echoes on the same note, even though it's a different scale. Graham looked through his ChatGPT thread and noticed he'd been using five specific words to describe what the model has said or what he wants it to do.

00:05:25 He described using a handful of epistemic verbs — guess, see, think, infer, and miss — each one signaling uncertainty about another system's internal state. He's not alone in this vocabulary. When you're talking to a system where you can't directly observe its reasoning, your word choice reveals how much agency you attribute to it.

00:05:48 "Guess" implies the model is reaching for something. "See" treats it as having perception. "Think" gives it cognitive process. "Infer" attributes logical operation. "Miss" suggests it failed to hit a target. Graham is also in the middle of arguing with the AP Style Guide over its understanding of current educated usage around AI-related language.

00:06:12 That's a fight everyone has at this point. Style guides are written by people. AI terminology moves faster than editorial boards can track. The verb list works as a behavioral datum. It's not about what the model does; it's about how we describe systems we can only partially observe.

00:06:31 The words you choose when talking to opaque systems become a kind of metadata about your confidence in their outputs. The same principle applies when you're building agents: the verbs in your prompts shape what the system learns it's allowed to do.

00:06:48 A Guardian piece today tracks something I hadn't noticed coming from the cultural side. Gen Alpha has adopted "that's so AI" as their primary insult — deployed when something is superficially convincing but barely worth anything, inauthentic, or of dubious value.

00:07:05 The article makes a clear point: this isn't actually about artificial intelligence. The kids have noticed that certain qualities associated with AI-generated content — the glossy-but-hollow feeling of things that look good until you look closer — show up everywhere now.

00:07:23 In knock-off merchandise, exaggerated corporate claims, excuses people make. If something feels suspect, it's AI. So "AI" has effectively become shorthand for "bullshit" among this generation, which is a striking cultural shift happening while the rest of us are still having technical debates about alignment and verification layers.

00:07:44 The article itself leans into humor — there's a section that reads like a mock encyclopedia entry, complete with fake age reports. The closing line: "Don't say 'No, that's Al — Al Gordon. He's in my year, and he's actually OK. But his shoes are AI.'" What's different here is the generational dimension — this is a cohort that will inherit these systems and has already made up its mind about their value proposition.

00:08:20 The generational divide in how we grade polish is worth tracking. We'll see how fast that vocabulary crosses into professional contexts, and whether companies optimize for polish knowing their audience might dismiss it as AI before reading further. That's the local reading.

00:08:38 Seln Oriax.