The second wave of AI coding is here

Abstraction: Next-gen AI coding agents targeting functional correctness via process data

Key points:

• Second wave goes beyond syntax-correct autocomplete to functional correctness — code that does what you wanted, not just code that runs
• Startups Cosine (Genie), Poolside, Zencoder, and Merly are building training data that captures the process of coding, not just finished code
• Poolside uses RLCE (reinforcement learning from code execution), analogous to RLHF, letting models self-improve via trial-and-error like AlphaZero
• Cosine's Genie tops SWE-Bench; Poolside building its own LLM from scratch to avoid irrelevant internet pretraining
• Merly's dissenting approach: skip LLMs entirely, train on intermediate representations (like machine code) arguing LLMs are fundamentally illogical
• Google reports >25% of new code AI-generated; Yang Li (Cosine) predicts software team sizes of 10-20 replacing hundreds; "era of adding thousands of engineers is over"

Connections: Agentic Coding · Reinforcement Learning · Anthropic · Openai · Github

Source: https://www.technologyreview.com/2025/01/20/1110180/the-second-wave-of-ai-coding-is-here/