Last week, Anthropic published what felt like a “panic piece” regarding Chinese companies “stealing” their models. Well, kind of. They described in detail how DeepSeek (and two others) ran extensive conversations with Claude to train new models based on its outputs.
To some, it’s IP theft; to others, it is a fun and effective technique. It reminds me of an ambitious junior regularly consulting a mentor to receive distilled wisdom on how to handle specific challenges. Sounds like a good thing, right?
Escaping the News Cycle
This is the third Monday I’ve dedicated to “outside reading” recommendations. It started with an ad for Tipsheet (a superb AI marketing newsletter) and continued with Robin Sloan’s ephemeral literary take on our AI reality. This is becoming a fun series on “places outside my bubble that found me.”
Today, I want to urge everyone to escape the madness of the 24-hour news cycle by indulging in a different kind of distraction: doom-scrolling through the future.
The 2027 Roadmap
Just last week, I re-read situational-awareness.ai (often mirrored at ai-2027.com) right before the Anthropic news broke. This project is the work of Leopold Aschenbrenner, one of the giants of modern LLM research who “freaked out” early, left the industry (or was pushed, depending on who you ask), and started playing wargames to predict the future.
He developed detailed methodologies to predict how this “game” will reshape civilization. One of his key concepts is the geopolitical race. I strongly recommend reading it every now and then—even though it’s a heavy, unsettling lift. It provides a framework for monitoring the news as it unfolds.
The “Agent 2” Scenario
There is a pivotal chapter where “China” steals the weights of Agent 2. In Aschenbrenner’s simulation, this was scheduled for February 2027, roughly a year after the “real” acceleration of AI capabilities begins.
The logic? Chinese research lags behind, and stealing the weights is the only way to catch up at the moment US research achieves “escape velocity.” Agent 2 is the first model capable of autonomous replication—not yet super-intelligent, but capable of self-improvement.
Scalping vs. Stealing
The Anthropic news felt like a first taste of how this scenario might actually play out.
- The Rehearsal: Maybe you don’t need to steal the actual weights of the entire model to “scalp” its highest practical value.
- The Pivot: Or maybe this is just a dress rehearsal for the “real thing” in 2027.
- The Actor: And maybe it won’t be China; it could be any entity looking to leapfrog the developmental curve.
On the flip side, the Aschenbrenner scenario predicts that being even six months behind could be devastating once the inflection point of AI self-training begins. I’m not entirely convinced a six-month delay is catastrophic. China is remarkably resilient and self-sufficient. In many ways, hearing US researchers obsess over this “delay” scenario tells us more about the American psyche than it does about Chinese capabilities.
Leave a Reply