The coding reality check
Real talk: everyone is lowkey freaking out that AI will eventually build a smarter version of itself and leave us all in the dust. But a fresh safety report from Google DeepMind on their new Gemini 4 Argon model says the vibes are actually pretty chill. Even though AI is currently churning out more than 75% of the code engineers at DeepMind ship, the lab says there is zero proof that AI is actually speeding up the overall development of new AI models.
The “intelligence” index
Google researchers built an internal index called GCI v1 to track AI progress like an IQ test. The data shows that frontier models have been improving at a super steady, straight-line pace of roughly 0.1 points per month. While the new Argon model is a bit higher, researcher Haydn Belfield says that’s just down to Google dumping more compute power and resources into it, not some rogue AI breakthrough.
Humans are still the main characters
Don’t start prepping for a sci-fi takeover just yet. While AI handles a ton of execution tasks, about 88% of code changes still require a human to steer the ship or manually review the work. The consensus among senior leaders is that AI is a total W at doing what it’s told, but it’s still majorly struggling with the “long-horizon research taste”—aka deciding what ideas are actually worth pursuing.
Why it matters
This report pushes back against the “recursive self-improvement” narrative that has had Wall Street and AI CEOs spiraling. While some industry heavyweights are worried that AI acceleration will outrun safety guardrails, Google’s internal data suggests we aren't at the point where AI is building its own brain—at least not yet. Stay skeptical, keep your bags packed for whatever comes next, and remember: none of this is financial advice.
