one year on
Phind launches GPT-4-powered 'Expert' search mode for developers
The developer-focused search engine integrates GPT-4 to answer technical questions with code examples and citations, claiming reduced hallucination versus the default model.
Phind, a developer-focused search engine, today launched a new ‘Expert’ mode powered by GPT-4. The feature uses generative AI to browse the web and answer technical questions with code examples and citations, aiming to reduce hallucination and keep it up to date.
rushingcreek says GPT-4 in Expert mode generates step-by-step instructions over 90% of the time, compared to the default model. The founder demonstrated improved answers for topics like RLHF and running Alpaca on llama.cpp. Users can enable the toggle at phind.com, though the feature has been quietly available on Discord for a few days.
Reactions from Hacker News were mixed. Some reported replacing most Google searches with Phind, while others noted it still confidently produced incorrect answers, especially for niche or version-specific queries. One user described the experience as ‘code reviewing a really apologetic and endlessly patient junior developer.’ The community debate centered on whether GPT-4’s accuracy improvements justify the slower speed, and whether LLM-powered search can ever overcome the hallucination problem inherent to the underlying technology.
announced the launch on HN, noting GPT-4 is 'significantly more concise and systematic' than the default model, and that Phind feeds relevant websites into the model to reduce hallucination.
Claimed to have replaced 90% of Google searches with Phind, but did not see a clear improvement with Expert mode.
Praised Phind's handling of a niche Swift/concurrency question, saying it pulled in relevant sources.
Reported that Phind gave poor answers to a pathfinding puzzle.
Described iterative refinement with Phind as 'like code reviewing a really apologetic and endlessly patient junior developer.'
One year later — open only if you can handle spoilers
Phind continued to iterate, later adding support for Claude and Gemini models. The company eventually raised a Series A and remains a niche player in the AI search space, still contending with the same accuracy issues identified in this HN thread.
The Weekly Replay · free by email
This week, one year ago — every Sunday.
One email each Sunday: the week's replayed AI news, with the one-year-later annotations included. Written like it's breaking — dated like it isn't.
Free · double opt-in · unsubscribe anytime · privacy