Phind

Phind built an AI search and reasoning system that combined language models with web and codebase context to answer technical questions and help implement solutions. Founded in 2022 by Michael Royzen and Justin Wei, with Royzen as CEO, it joined Y Combinator’s Summer 2022 batch. Its consequential decisions included adopting GPT-4 when its own models fell short, returning to specialized model training, and expanding from developer assistance into generated interactive interfaces.[2][1][6]

The company’s search architecture grew out of Royzen’s earlier question-answering experiments, beginning in 2020, before Phind’s founding.[1]

Royzen credited Hugging Face researcher Yacine Jernite’s summer-2020 long-form question-answering demo with starting this work. He recalled that the notebook used BART and supported retrieval through an Elasticsearch index of Wikipedia or a dense index using Facebook’s FAISS. He then experimented with a larger BART model fine-tuned on Reddit’s ELI5 questions and answers.[1]

In fall 2020, Royzen filtered Common Crawl to roughly the top 10,000 websites and built an Elasticsearch index of approximately 350 million webpages, according to his account. Retrieved passages supplied the context for generated answers. After shelving the project for almost a year, he returned in fall 2021 using BigScience’s T0-3B, again fine-tuned on ELI5. He presented this first iteration of Hello on Hacker News in January 2022.[1]

The team focused on programmers in June 2022. Royzen said it was more interested in reasoning, debugging, and complex problem solving than in summarizing basic search results. After entering YC, it abandoned its cumbersome internal index and used Bing as the raw search-data source, launching as Hello Cognition. He recalled approximately 10,000 searches per day for that summer’s programmer-focused product.[1]

Royzen dated the rebrand to Phind to late December 2022. He credited Paul Graham with suggesting the spelling while discussing alternatives to the team’s sayhello.so domain; the founders subsequently purchased the domain and changed the name.[1]

Phind adopted OpenAI models in February 2023 because its own models were inadequate, Royzen said. In April, it offered GPT-4 answers free without requiring a login. He described the decision as expensive but effective at showing users what Phind could do. He also said OpenAI helped the company configure Cloudflare bot controls after others built unofficial APIs to exploit that free access.[1][4]

The August 7, 2023 Phind V2 launch, signed Michael and Justin, made GPT-4 the default answering engine and introduced a VS Code extension that retrieved relevant codebase context. Earlier versions searched the web for every input; V2 could choose when to search, ask clarifying questions, enter a debugging mode, or call itself recursively for problems requiring multiple searches or reasoning steps. The older Basic Search mode remained available.[4]

Pair-programmer conversations exposed another context problem: users sometimes built threads of 60 or 100 messages while working from an idea toward functioning code. Royzen said requests to pin important messages led to a prioritization feature that helped retain selected context. He favored supporting the developer’s broader planning and implementation process through both the web interface and an IDE extension, rather than owning an IDE.[1]

One concrete use case came from Hacker News user CharlesW, who reported rebuilding a Vue 2 and PHP application using Vue 3, Vite, and TypeScript with Phind’s help. The user estimated 50% time savings on the tasks where it was used, while also reporting confidently wrong answers that required correcting assumptions, missing information, or mistakes.[4]

Phind’s CodeLlama-34B-v2 model card describes further training of Phind-CodeLlama-34B-v1 on 1.5 billion proprietary programming-data tokens. The data consisted of instruction-answer pairs rather than code-completion examples. Phind used native fine-tuning without LoRA, DeepSpeed ZeRO 3, and Flash Attention 2, reporting 15 hours of training on 32 A100-80GB GPUs with 4,096-token sequences. Its 73.8% HumanEval pass@1 result was explicitly self-reported.[3]

A separate model summary in the November 2023 Latent Space interview described a CodeLlama-34B-based model with more than 70 billion additional code and reasoning tokens, a 16,000-token context window, and a reported 74.7% HumanEval score. Those figures describe the interview’s model summary rather than the public v2 card’s training configuration and result.[1][3]

Royzen described evaluating real-world answers with GPT-4 scoring and user feedback, alongside separate checks of retrieval context and answer correctness. Bad context and a bad answer suggested retrieval needed improvement; good context and a bad answer pointed toward prompting or model limitations. Phind measured hallucination indirectly through these checks. Royzen also described exploring reinforcement learning for correctness to reduce inconsistent answers, rather than presenting it as a completed capability.[1]

On February 14, 2025, Y Combinator announced Phind 2, a separate generation from August 2023’s V2. The announcement reported a six-month rebuild following more than 130 million cumulative searches. It added inline images, diagrams, cards, and other widgets, along with multiple rounds of additional searching when more information was needed. The accompanying demonstration showed Python calculations and graphs within chat, including coin-flip simulations and simple perceptron training.[5]

The subsequent Phind 3 launch, signed Michael, replaced prebuilt widgets with generated React mini-apps and tools whose schemas the system designed itself. The launch explained the limitation through a travel example: a fixed Expedia widget could display cash fares but could not accommodate a request comparing cash and points fares. Generated interfaces were intended to adapt to such requests, with examples including apartment filters and maps, adjustable recipes, and an interactive quicksort walkthrough.[6]

Phind 3 introduced Phind Fast, based on GLM-4.5-Air, and Phind Large, based on GLM 4.6, with custom Eagle3 heads for faster inference. Phind claimed maximum speeds of 300 and 200 tokens per second respectively, without specifying hardware or workload conditions. In response to users who preferred the previous experience, Michael directed them to old.phind.com.[6]

Royzen said Paul Graham invested during the Summer 2022 YC batch, after which the rest of the round came together quickly. Graham introduced the team to Ron Conway, whose email to Jensen Huang led to NVIDIA assistance with GPU options. Royzen also credited NVIDIA with implementing streaming generation for Phind’s then-used T5 models in FasterTransformer.[1]

By August 2023, Phind described itself as venture-funded and offered paid plans for higher GPT-4 usage limits. On December 3, 2025, Axios reported that Phind raised $10.4 million.[4][7]