Engineers Float Autocomplete Models Over Chat Agents
X users discuss alternatives to chat agents and templating for LLMs in code tasks.
Eiso Kant replied that post-training large language models could produce creative results if developers move past standard chat templating constraints, noting untapped flexibility in fine-tuning. Gabe posted that a different path for agents might have created supercharged tab autocomplete models able to generate large code chunks while remaining closer to the system. The comments appear in a thread on X about LLM development choices.
I kinda feel like we could have taken a different path with the agents and made a supercharged version of the tab autocomplete models. One where you could generate very large chunks of code but still stay closer in touch with the system.
i understand why chat is good and general, but it鈥檚 such a small slice in the universe of possible intelligences anyone who has used mid-2025 cursor tab, or played against alphazero, or ridden in a waymo, knows this
Reactions from ranked influencers
2 postsI kinda feel like we could have taken a different path with the agents and made a supercharged version of the tab autocomplete models. One where you could generate very large chunks of code but still stay closer in touch with the system.
i understand why chat is good and general, but it鈥檚 such a small slice in the universe of possible intelligences anyone who has used mid-2025 cursor tab, or played against alphazero, or ridden in a waymo, knows this
