AI/LLM

Large Language Model AI (LLMs)

I often use these for answering questions (usually technical, occasionally political).

commercial services

• ChatGPT (by OpenAI, which is neither open nor AI)
• Copilot (Microsoft)
• FastGPT: seems to be Kagi's own model, not just a front-end to some other LLM
  • Kagi's Docs: Kagi AI
  • 2023-03-16 Kagi's approach to AI in search
• Perplexity
  • What software is needed for training a large language model (LLM)?

ideas

• an LLM which summarizes the most important news items that have happened since the last time you checked.
  • The number of items and the amount of detail could be customized.
  • Certain subjects could be prioritized or filtered out (e.g. I wouldn't want to hear anything about sports or celeb news unless it had implications outside of those fields).
• LLMs could make federated search (of which I think YaCy is currently the only exemplar) actually workable/useful:
  • You type in your search -- either as keywords or as a natural-language query.
  • A search is done (however LLMs do it currently -- I don't know if they fine-tune the search parameters or just pass them on to a search engine).
  • The LLM then processes/glosses the top results to look for pages that seem to be most applicable to the request. If nothing passes a threshold test, it looks at the next bunch of results. If it can't find anything likely, it summarizes what it did find and notes that it may not be what you're looking for.
• I'd like to see an LLM browser plugin that can find tabs based on natural-language queries -- "where's the one with that medical form?" "Have I stashed any tabs about...[subject]?". Similarly, an app or library which can take a URL, load the corresponding web page, and produce a summary and keywords would also be hella useful.
• (long-standing wish which now seems within reach) I need an LLM which can make phone calls to do basic adulting tasks, e.g. making appointments, renewing prescriptions, getting information ("is the car ready for pickup?").

Links

• 2024-11-11 AI’s math problem: FrontierMath benchmark shows how far technology still has to go
• 2024-11-10 Evaluating the World Model Implicit in a Generative Model (paper)
• 2024-11-05 Despite its impressive output, generative AI doesn’t have a coherent understanding of the world
• 2024-11-04 When Can LLMs Actually Correct Their Own Mistakes? A Critical Survey of Self-Correction of LLMs
• 2024-10-07 GSM-Symbolic: Understanding the Limitations of Mathematical Reasoning in Large Language Models (paper by Apple researchers; via)
• 2024-09-21 Hype, Sustainability, and the Price of the Bigger-is-Better Paradigm in AI (paper)
• 2024-08 LLMs cannot find reasoning errors, but can correct them given the error location (paper)
• 2024-07-27 Evaluating LLMs at Detecting Errors in LLM Responses (paper) - introduces the ReaLMistake benchmark
  • ReaLMistake code - but the data is password-protected.
• 2024-03-14 Large Language Models Cannot Self-Correct Reasoning Yet (paper)
• 2023-05-24 Worried About Sending Your Data to a Chatbot? 'PrivateGPT' Is Here

Code Sources

• Auto-GPT «chains together LLM "thoughts", to autonomously achieve whatever goal you set.»
  • Downside: not fully autonomous; still depends on OpenAI
• privateGPT