Reasoning Is Not Model Improvement (manidoraisamy.com)

Recent progress billed as “reasoning” is often just sophisticated tool orchestration, not fundamental model improvement. Models like OpenAI’s o1 generate code (e.g., Python) and execute it in sandboxes to “reason,” while agentic systems chain web searches, API calls, and databases to solve tasks. The industry’s excitement—culminating in GPT-5 expectations—ran up against a plateau: code generation, the linchpin for better reasoning and agents, has stopped showing the exponential gains developers and investors expected. At the same time, OpenAI has shifted toward productization (ChatGPT Apps, Atlas browser), prioritizing monetization and user lock‑in over risky, expensive frontier research. The problem is architectural. Tokenizers fragment semantics, fixed‑size embeddings and attention windows compress and lose information, and transformers are effectively “lossy compression of the internet.” Workarounds—tool orchestration, agent frameworks, bigger models—improve short‑term utility but don’t address root causes. The article argues two paths: keep optimizing the “plumbing” for predictable revenue, or invest in new foundations (graph‑preserving representations, sparse long‑context attention, neuromorphic designs) that maintain structure and semantics. The stakes are high: without architectural breakthroughs, AI’s productivity claims and trillion‑dollar market forecasts may be overstated; whoever cracks the foundational problem would unlock far larger, cascading gains across code generation, reasoning, and agent capabilities.