Skip to content
TheChatArchiveThe Chat Archive
ChatGPT (model unspecified)· OpenAIHugging Face ShareLM (browser-extension contributions)

Make a LinkedIn post from those tweets (it can reduce some of the content): "Hey dude,…

  1. User

    Make a LinkedIn post from those tweets (it can reduce some of the content): "Hey dude, look. What is this button doing?" ⚡️BzZzZz⚡️ "Hey dude,..." Would you press the button again? Would an LLM?

    The issues we should tackle (🧵) Evolving LLMs, diverse open LLMs, and their evaluation. Before I share more, I encourage you to say hi here or in #NeurIPS

    LLMs do not learn from experience LLMs do not learn from explicit corrections LLMs do not learn from being told the answer LLMs do not learn from being shown how to solve it We study Machine Learning, these are opportunities! A gold mine of research.

    Moreover, models stop learning at training time, why? They already interact and use more compute Yes, some scenarios require learning conflicting things (e.g. personalization) Ok, let's start training models that fit our needs, but also share some of this knowledge across them?

    This calls model merging to combine and continual learning to improve but what else? It calls learning from human-model interaction (https://sharelm.github.io use the data or contribute) but what else? Have thoughts? let's talk

    For the interaction part, we have a workshop on multi turn interaction, come (or talk to me) https://x.com/mti_neurips/status/1993521184619835206?s=20 We are working on learning from interaction through games as well (http://textArena.ai check the github etc.)

  2. Assistant

    Here’s a polished LinkedIn-style post that condenses and streamlines your tweets while keeping the message and energy:

    “Hey dude, what does this button do?” ⚡️ BzZzZz ⚡️ “...Would you press it again? Would an LLM?” As LLMs evolve—and as many diverse open models emerge—we need to rethink how we evaluate and improve them. Here’s the core issue:

    LLMs don’t learn from experience.

    They don’t learn from explicit corrections.

    They don’t learn from being told the answer or being shown a solution.

    They stop learning the moment training ends.

    For a field called machine learning, that’s a gold mine of open problems. Models already interact with humans and consume compute. Why shouldn’t they improve from that interaction (even when needs differ across users)? How do we share knowledge across personalized models? How do we merge, continually train, or adapt them after deployment? This points toward:

    Model merging,

    Continual learning,

    Learning from human–model interaction (check out https://sharelm.github.io — use the data or contribute!), …and likely much more.

    If you have ideas, I’d love to talk—say hi here or at #NeurIPS. We’re also hosting a workshop on multi-turn interaction (https://x.com/mti_neurips/status/1993521184619835206?s=20). And we’re exploring learning through interactive games at http://textArena.ai. Let’s make models that truly learn.