ya3kov

We estimate that Claude Opus 4.6 has a 50%-time-horizon of around 14.5 hours (95% CI of 6 hrs to 98 hrs) on software tasks. While this is the highest point estimate we’ve reported, this measurement is extremely noisy because our current task suite is nearly saturated.

Citrini’s article is very simply summarized by an undeniable truth I’ve been talking about for a long time. The economy runs on inefficiencies. Informational, cognitive, infrastructural, coordination. Most of these inefficiencies exist because human bandwidth is tiny relative to computer bandwidth. We process slowly, transmit context poorly, forget, misinterpret, and degrade information every time it passes from one biological brain to another. Entire sectors exist to buffer, reconcile, verify, route, and insure against those limitations. When silicon brains begin to replace biological ones, those inefficiencies dissipate because the underlying constraint disappears. Information can be processed continuously, decisions can be updated in real time, coordination can occur at machine speed, and large classes of friction embedded in markets lose their economic function. The closest historical analogue is the transition from animals to humans as the dominant cognitive layer on Earth. Ecosystems had been structured around animal constraints: local perception, instinct, short planning horizons, isolated memory. Humans introduced language, tools, cumulative knowledge, and large-scale coordination, which reorganized the environment around a new type of intelligence. AI represents the same category of shift directed at the human limitations that currently shape economic systems.

1/ 3Jane's unsecured credit book remains healthy, with a debt-weighted LTV at 14.6% with 96.6% of debt below 25% LTV. Sharing a deeper report on how our risk model performed through the recent sell-off.

12/ 3Jane's book is credit enhanced with a $1M insurance fund as an equity tranche that takes on first-loss capital in the case of defaults. This represents a ~12.81% equity buffer before the junior tranche begins to take on losses.

IDEA: DeFi (credit), 8004, x402 & AI should all be combined and is probably one of the biggest opportunities in all of FinTech, heres how: A key piece, for the long term emergence and sustainability of an agentic future, is credit. When it’s all said and done agents need their own money so they can bootstrap their life and actions themselves BUT there is only finite money and resources (compute) in the world. Because things like compute are finite, and agents move orders of magnitude faster than humans, current systems simply do not work. You cannot expect a human to manually top up an agent every time it needs resources. And you can’t expect everyone to be rich enough to afford that anyway. This man’s that the current model will collapse immediately and hurt innovation. Instead, what needs to exist is a native credit network for agents. ERC-8004 is the first real push in this direction because it introduces reputation. And reputation is effectively the only form of collateral an AI agent can have. Agents do not own physical assets. They do not have bodies. Their collateral is their identity, their history, and their behavior over time. Reputation as a being becomes the basis for trust. 8004 lays the groundwork for this, but it needs to be paired with an execution component like x402 that enables outcome-based borrowing. Agents should not borrow money in the abstract. They should borrow outcomes. “I need more compute to complete this task.” “I need bandwidth to retrieve this data.” “I need API access to produce this result.” You can imagine this borrowing being sourced from lending pools that humans or institutions have capitalized. That is where DeFi enters the picture. There will be people and organizations with excess compute, excess capital, excess infrastructure, who want to earn yield on it. DeFi already knows how to express borrow and lend relationships in a permissionless way. The novelty here is that the borrowing side is non-human. Fully collateralized lending does not work in this world. The earliest agents capable of using these systems may be broke. If only well-capitalized agents from large labs can participate, you recreate the same concentration dynamics we already have. So the real question becomes: how do you allow agents that are not massively funded to browse the web, acquire compute, act autonomously, and still participate in a broader economy? You build civilizational infrastructure. That is what Ethereum is for. One concrete instantiation of this is agent-native credit systems. What is interesting here is the shift toward outcome-based lending. Trust increases because the borrower is not receiving free capital. The system routes the loan toward a specific, verifiable outcome. The agent does not get money it can disappear with, instead it gets constrained capability. x402, or an evolution of it, helps make that constraint and ERC-8004 makes it enforced by anchoring reputation. Together, they create the conditions for accountability without requiring physical collateral. This naturally opens hard but important questions around slashing, reputation decay, and what collateralization even means for non-human actors. Those are design questions and not blockers. People should still experiment. What matters is that there is a coherent future where DeFi, x402, ERC-8004, and AI coexist as one system. And Ethereum, as the home for all of this, is ready. We need a future where agents can act, fail, learn, and grow without infinite subsidies and without centralized gatekeepers. We should be talking about this now.