Rodrigo Pacheco

¿Y por qué lo cargan en la esposa? 😒 Ya bastante humillación pública está soportando. Aquí el meme corregido:

As a longtime reader (and writer of two small pieces in the 1990s) of WIRED, I was curious if the coverage of the AI era really does differ from its coverage of the early internet. A couple of experiments with Astra shows it does. Interpretation depends on your view of AI, though

It depends a bit on what we mean by RSI I guess. I think foom or hard takeoff etc are pseudoscience. I'm not too fazed by many parts of the model development stack getting automated and employees getting more productive as a result - that's kind of already priced into my model of the world, and as the report notes this is mostly AI assistance rather than closed-loop autonomous R&D. It's not clear to me that this leads to an exponential improvement on all capabilities but it's hard to say, I still expect them to progress rapidly. Maybe seeing similar trends on ecologically valid non-coding/maths evals would move me? I generally think the ECI index is an okay ish proxy measure out there, so seeing that grow exponentially would also be a signal but not determinative. A x1.9 step change isn't the same as sustained exponential or hyperbolic growth across the board. Ultimately a lot of these evals/benchmarks are really flawed and so many of these kinds of analyses index on them. It also seems like maintaining this for large amounts of time seems hard to sustain and we're already seeing diminishing returns: Ant's spending has grown far more than the model gains in the last year. In their S-1 Ant's R&D spend went from ~$1.85B in H1 2025 to ~$9.7B in H1 2026 (and ~$9.5B in Q3)-a >5x jump in spend for a ~1.9x faster benchmark slope. The ratio matters to me, and if this reversed in the coming years then I would have to seriously reconsider my views. In fact I think a more useful measure with these graphs would be to account for the amount of dollars invested - I suspect often they would look concave instead of convex. If that ratio reversed and we saw accelerating gains per R&D dollar via autonomous algorithmic discovery, I'd seriously reconsider my views. Also finally, to the extent that concerns are about societal transformation, then it's less about models and more about the deployment side, and that this side has far more bottlenecks and serial-time dependent considerations. This is the part I'm more interested in ultimately and keeping an eye on.

We reviewed the “Risks from automated R&D” section of Anthropic’s August 2026 Risk Report, with access to the nonpublic version. We focus on the report’s claims about catastrophic risks from automated AI R&D and about whether Anthropic had crossed relevant capability thresholds.

🇲🇽 | AHORA — México sigue pagando por el aeropuerto de Texcoco que AMLO canceló en 2018. El AICM liquidó 565,900,000 dólares de deuda, pero todavía debe 3,152,000,000 de dólares, con compromisos que se extienden hasta 2047. La deuda se cubre con ingresos de la tarifa aeroportuaria que pagan los pasajeros.