alt Hacker NewsDe mortuis and all that, but someone has pointed out the poor track record of his later initiatives and I think the reasons for those failures are an important part of the picture. Look at his last few blog posts, outlining a radical alternative to LLMs. Read closely, and at the heart of that idea we find "assume enterprise search". But the willingness to consider radical alternatives is part of the right way to think.
Replies
coldtea • today at 1:04 PM
>But the willingness to consider radical alternatives is part of the right way to think.
That sounds liked LinkedIn speak.
➕ show 2 replies
I don't really understand what your comment is trying to say, but I read through the last few pages of his blog and all the posts after the first one from 2026 announcing his return seem AI generated. I think it's a shame, as I loved his writing style in "Accidental Empires" and in his earlier blog posts and columns. It's strange to me that someone who had spent most of his life writing either didn't notice or didn't care about the drop in quality from using LLMs to write.
> Look at his last few blog posts, outlining a radical alternative to LLMs. Read closely, and at the heart of that idea we find "assume enterprise search".
I found the patent for his company's technology (https://patents.google.com/patent/US20250259069A1). I always have trouble reading patents, but as far as I can tell it's basically a wrapper around commodity LLMs. The reliability improvements come from prompting the LLM to only source its data from an external database that you provide rather than from whatever it was trained on, and from prompting multiple LLMs simultaneously then doing semantic analysis on the responses to make sure they all agree. I imagine that doing this would probably reduce hallucinations, but there's no hard constraint to prevent them entirely and I'm not sure what the moat is.