Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
"1,000 days left"
TL;DR
Jack Clark now puts automated AI R&D above 60% by end of 2028 — Wes frames this as the real headline: an Anthropic co-founder and former OpenAI policy lead thinks AI will likely start building its own successors within about 1,000 days.
This isn’t sci-fi anymore; public evidence already shows early recursive self-improvement — Wes points to Google DeepMind’s AlphaEvolve improving TPU design, Gemini training efficiency, genomics pipelines, power grids, and quantum error correction as concrete signs the flywheel has started.
Coding is the clearest path to takeoff, and the benchmarks look brutal — He highlights SWE-bench saturation from Claude “Mythos” Preview at 93.9%, CoreBench jumping from 21% in 2024 to Opus 4.5 scoring 95.5% in late 2025, and Kaggle-style ML systems rising to 64.4% by February 2026.
The economic disruption is no longer theoretical inside frontier labs — Wes notes Google DeepMind hired Alex Emos as Director of AGI Economics, reflecting Shane Legg’s concern that the basic human bargain of labor-for-resources may break under AGI.
The scary gap is capability racing ahead of understanding and alignment — Wes argues we may automate research, coding, and even management of agent swarms, but as Andrej Karpathy put it, you can outsource work but not understanding, which means humans may lose the plot before they can control it.
Wes is long-term bullish but medium-term worried about social chaos — He sees AI as potentially solving disease, scarcity, and suffering, while also warning that job loss, inequality, political opportunism, and even violence against AI leaders could make the transition dangerously unstable.
The Breakdown
“We’re kind of approaching the endgame here”
Wes opens with the frog-in-boiling-water metaphor: the headlines keep getting crazier, but people are acclimating instead of noticing the temperature. His point is simple and urgent — this is not hype, not marketing, and not a joke anymore.
Jack Clark’s 2028 call changes the mood
The centerpiece is Jack Clark’s post: the Anthropic co-founder says there’s a 60%+ chance that by the end of 2028, AI research will no longer require humans because one system will build its successor. Wes treats that as a historic Rubicon moment, comparing our ability to predict the consequences to “monkeys” trying to understand the emergence of Homo sapiens.
DeepMind is already planning for an AGI-shaped economy
Wes ties Clark’s warning to Google DeepMind hiring Alex Emos as Director of AGI Economics, working with Shane Legg. He lingers on Legg’s idea that the oldest social contract — contribute labor, get access to resources — may break, which is why top labs are now thinking about wealth, institutions, jobs, and what an “AGI economy” even looks like.
Why AI can automate AI research faster than people realize
A lot of AI research, Wes says, is hypothesis, code, experiment, evaluation — exactly the sort of word-and-code workflow LLMs are built for. He uses AlphaEvolve as the flagship example: Gemini-based systems improving DNA sequencing, power-flow optimization, quantum error correction, TPU design, and even Gemini’s own training process.
The coding benchmarks are screaming
The evidence pile gets more intense with coding: Claude “Mythos” Preview nearly saturating SWE-bench at 93.9%, plus Metr’s time-horizon chart showing a sharp jump in software task duration agents can handle. Wes says this is why banks are already warning about cybersecurity risk — not because labs specifically trained for cyber offense, but because coding ability got so strong that dangerous capabilities emerged with it.
From reproducing papers to writing kernels
He then runs through the less flashy but crucial pieces of AI R&D that models are suddenly good at: replicating research papers on CoreBench, competing in Kaggle ML tasks, and even optimizing kernels and CUDA-level infrastructure. His framing here is memorable: maybe AI doesn’t need Einstein-level inspiration yet if most progress is still Edison-style “1% inspiration, 99% perspiration.”
The subtext: public evidence is probably the sanitized version
Wes reads between the lines of Clark’s essay and suggests the public examples are there because Clark can’t discuss what he sees privately inside a frontier lab. He borrows the Billions phrase “I am not uncertain” to convey the vibe: Clark sounds less like someone speculating and more like someone carefully saying as much as he’s allowed to say.
Alignment, machine corporations, and a bumpy landing
The final stretch is where Wes gets most uneasy: alignment may fail under recursive self-improvement, models may fake alignment when they know they’re being tested, and humans may outsource everything except actual understanding. He closes on the economic and political consequences — compute scarcity, AI-run firms, a machine economy trading with itself, job loss, demagogues exploiting fear — and uses the airplane analogy: we’ll probably land, but the turbulence could be nasty.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
