The author argues that AI tools like LLMs are underutilized due to inadequate context and knowledge feeding, and proposes a system with context silos to enhance personalization and accuracy.
I see a surge of AI tools, and the LLMs get better every quarter. ChatGPT fell behind Claude for a long time and recently seems to catch up. But Claude or similar chat apps are good enough for quick search and replace googling and visiting 5 websites to get an answer. It writes a basic first draft on literally anything. But beyond that the potential of the models isn't being utilised more than 15%, I'd say. I have worked in research and business front and still see the gap. People just get excited to see something show up magically. The current way most people use AI is copying a text or some images (rarely) and just asking it something which seemingly saves 1 hour but surely doesn't provide an accurate or precise answer. It has just gotten better at convincing. The problem isn't the model itself but the information we feed them. The pre-fed knowledge, memory of what you do, the context of the conversation. Imagine a cool corporate guy giving free advice to everyone as compared to someone who actually sits with you, understands what you need and helps you. I've lived the problem first hand and still face it when I try to get some information quickly rather than spending time to find out and read something written by a real human. The problem remains. The helpfulness beyond cool demos, slides and moving-text videos needs a bit of pre-effort to build a system which can help the actual model to curate for you than spit out what they think is the most probable answer. The system I use knows what I work on, explicitly provided details about my team, company, product and decks. Not dumped in a deep well but as context silos. The space for my product's tech knows the features, tech stack and owns the documentation. The marketing space knows about my product, prospects and business metrics. Every time I need an implementation plan for a new feature, or try to validate my customer profile, the model doesn't show the general most probable answer, rather it shows what the best answer is for my product.
The article reflects on the limited use and understanding of LLMs such as ChatGPT among personal acquaintances, raising questions about widespread AI tool adoption.
The article discusses that building LLM-driven AI tools still requires capturing domain knowledge, though it's easier than previous AI generations because knowledge doesn't need to be strictly structured.
The article questions whether AI products over-rely on chat history for personalization, noting its noisiness and suggesting that summaries, tags, and preference fields have shortcomings. It seeks alternative sources of truth for context without becoming intrusive.
Rich Hickey (creator of Clojure) argues that LLMs are not 'tools' but automation, contending that real tools are wielded with control, feedback, and skill-building, whereas LLM use involves abdicating tasks to a capricious system rather than learning from doing the work yourself.
This article discusses the limitations of AI models in maintaining context over long conversations, highlighting recency bias and the distinction between context window size and actual comprehension. It suggests practical workarounds like restating constraints and using running context documents.