The archive · AI & Models · Product decision · 2024-2025
Phind rebuilds AI search around a bespoke 70B model with visual, self-searching answers
After six months rebuilding, Phind relaunched with a bespoke 70B model, visual answers and self-run searches; HN launch drew 537 points on 2025-02-13
Phind
What the business is
Phind is an AI search engine that answers questions in natural language with cited sources and, since Phind 2, with visual widgets, agentic follow-up searches and Jupyter-backed computation; it serves developers first but now competes as a general search product on a free tier plus paid plans.
How it started
Phind started as an AI answer product for developers: it shipped a VS Code extension, launched on Hacker News, and gathered paying users who asked coding questions there. The launch post for Phind 2 says the new 70B model is completely different from the one the company launched a year earlier. By early 2025 the retention leak was visible in the community: users who loved the first version said they were shifting traffic back to Perplexity and ChatGPT as those products improved, and one subscriber described cancelling because DeepSeek was enough for occasional LLM answers. CEO Michael, posting on HN as rushingcreek, said the team went quiet while it rebuilt the entire product.
What happened
Phind 2 shipped on 2025-02-13 after the six-month rebuild. The team said off-the-shelf models were incredibly bad at generating visual components reliably, so it post-trained a bespoke 70B series with a system of LLM critics that generated a high-quality dataset, and trained the model to emit Mermaid diagrams when helpful. The new frontend renders inline images, cards and diagrams, lets the model run extra rounds of searches when it needs more information, and verifies calculations by executing them in a Jupyter notebook. The company also deprecated its VS Code extension and said it was going all-in on search, with an API planned later that year. The launch thread drew 537 points and 192 comments: paying users called the results better than anything else they had tried and some resubscribed, while others reported hallucinated event dates, verbose diagrams, sources that did not always appear, and service unavailable in their region.
No ending yet — it is still running.
Background
Phind began as an AI answer engine aimed at developers: it shipped a VS Code extension, launched on Hacker News, and gathered paying users who asked coding questions. When Phind 2 arrived on 2025-02-13, the launch post said the team had spent six months rebuilding, and that the new 70B model was completely different from the one launched a year earlier. Several commenters explained why the rebuild mattered: as Perplexity and ChatGPT improved, they had shifted traffic away from Phind, and one subscriber had cancelled in favour of DeepSeek.
The bet was that answer quality, not chat convenience, is what keeps people paying for AI search. Phind built its own 70B series after finding off-the-shelf models incredibly bad at generating visual components reliably, post-training with a system of LLM critics and teaching the model to emit Mermaid diagrams. Phind 2 presents answers with images, cards and diagrams, lets the model run additional rounds of searches on its own, and verifies calculations by running them in a Jupyter notebook; the company also deprecated its VS Code extension, saying it was going all-in on search.
The launch thread drew 537 points and 192 comments on 2025-02-13. Enthusiasts said the results beat Perplexity and ChatGPT, and some resubscribed; critics reported hallucinated economic-calendar dates, overly verbose diagrams, missing sources and regional blocks. The team promised an API later that year, a products UI, and no ads. The thread discloses no funding or revenue figures.
What has to be true
- The retention leak was concrete and community-visible: users said they shifted traffic back to Perplexity and ChatGPT as those improved, and one subscriber cancelled for DeepSeek
- Off-the-shelf models were, in Phind's words, incredibly bad at generating visual components reliably, so it post-trained its own 70B - a visual format it had to build, not wrap
- Self-run extra searches answered the failure users named: confidently out-of-date answers are useless for time-sensitive questions, so Phind 2 chased timeliness and breadth
- The company committed rather than hedged: it deprecated the VS Code extension, went all-in on search, planned an API, and publicly ruled out ads - a positioning bet as much as a product one
What can be applied
When rivals improve faster than you, do not chase their chat UX: rebuild the answer itself - a model and output format they did not bet on - commit to one channel, and refuse the obvious monetization
Aftermath
As of the launch date, Phind 2 is live with a free tier and paid plans. In the HN thread the team answered criticism directly, pointed users to a plaintext answer profile, said a products UI was on the way, and promised to keep Phind ad-free; it also said an API was planned for later in 2025. The public material discloses no funding or revenue figures.
Sources
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card