Updated
Updated · O'Reilly Media · Aug 5
Post-Training Turned LLMs Into 1 Billion-User AI Products
Updated
Updated · O'Reilly Media · Aug 5

Post-Training Turned LLMs Into 1 Billion-User AI Products

3 articles · Updated · O'Reilly Media · Aug 5

Summary

  • About 1 billion people now use AI products because post-training turned raw LLMs from brittle text completers into systems that can chat, follow instructions, use tools and reason through tasks.
  • GPT-3’s 175 billion-parameter base model showed why that mattered: tiny prompt changes could derail outputs, because pretraining optimized next-token prediction on internet text rather than robust conversation behavior.
  • Post-training adds that behavior through curated demonstrations, human preference data and programmatic scoring, shaping usability, safety and capability instead of the broader knowledge learned during pretraining.
  • The compute gap is also narrowing: standard alignment runs may use tens to hundreds of GPUs for days, while reasoning-focused post-training can consume roughly half a model’s total compute.
  • That shift means progress increasingly depends on specifying and rewarding good behavior, with verifier-based reinforcement learning already pushing models toward superhuman performance in math, coding and some scientific tasks.

Insights

Could better rewards and verifiers matter more than bigger models in the next wave of AI reasoning and tool use?
If pretraining gave AI knowledge, why does OpenAI say post-training is what truly made ChatGPT useful and safe?
Is post-training now more important than pretraining in AI progress, especially as reasoning-focused training rivals pretraining compute?