AI development is entering a recursive loop. Anthropic engineers no longer prompt AI agents directly. Instead, they set up loops where agents prompt each other autonomously. According to a Ten31 segment on TFTC, this has led to a system where Claude now writes roughly 80% of the code for its own new models.
Marty Bent argues this eliminates the friction of human linear thinking, signaling a potential fast takeoff in model capabilities. The technical disclosure, largely ignored by mainstream media, suggests the traditional development curve is becoming exponential.
"Anthropic's blog post claims Claude now writes 80% of its own code for new models, accelerating toward recursive self-improvement and potential AGI."
- TFTC
Yet the race for raw scale masks deep inefficiencies. On Nerd Snipe, Theo argues that standard coding benchmarks like SWE-bench are broken, compromised by contamination where models regurgitate training data. A new, more realistic audit from DeepSeek reveals a 30-point performance gap between GPT-4o-mini and the full GPT-4o on real tasks.
Claude Opus 4.8 scored 58% on this new benchmark, trailing OpenAI's 70%. The cost delta is staggering: OpenAI's model solves tasks for $6.60 on average, while Opus 4.8 costs $12.58.
"On the DeepSeek SWE benchmark, GPT-4.5 scored 70% while Claude Opus 4.8 scored 58%. The hosts note a massive efficiency gap."
- Theo and Ben, Nerd Snipe
The industry's astronomical costs are driving a political consensus. On TFTC, John noted that both Donald Trump and Bernie Sanders have proposed the federal government taking a stake in leading AI labs. This rare alignment suggests frontier AI is being viewed as a critical national utility.
This shift toward 'too big to fail' status comes as companies increasingly avoid public markets. On the All-In Podcast, Brad Gerstner noted secondary trading volume has doubled since 2021, now representing 31% of all venture activity.
Gavin Baker argued this creates a sycophant loop. Private investors, afraid of losing access to hot rounds, hesitate to deliver hard truths. Without public market pressure, founders risk building in a bubble. Baker cited Mark Zuckerberg’s delayed pivot to mobile at Facebook as a classic example of this dynamic.
The core tension is between explosive, self-directed capability growth and the market mechanisms needed to discipline it. The AI writing its own code may outrun the systems built to guide it.




