6 ms·

Hi HN, We're Jed and Angie, and we're building LazyWeb (https://lazyweb.ai https://lazyweb.ai), a search engine with a simple chat interface that's free from a

Hi HN, We're Jed and Angie, and we're building LazyWeb (https://lazyweb.ai https://lazyweb.ai), a search engine with a simple chat interface that's free from ad-tech and tracking. It's like messaging a smart friend that answers questions and sends you useful links. I'm a reluctant expert in how much damage ad-tech has done to media. I'm a programmer, but I also trained and worked as a journalist and editor, and built a SaaS startup serving media companies. So I've watched first-hand how the media industry has not-so-slowly starved, including seeing friends and customers lose their jobs and businesses. In its place, ad-tech, SEO spam and clickbait have taken over the Internet and dominate search. HN has posts every day about Google search getting worse, and ad-tech more intrusive. We're not under any illusions how hard this problem is to fix. But we think we can make a dent in it by building a search application based on ethics and trust, and focused religiously on users, not advertisers. Google is an ad-tech company, not a search company. More people work on ads than search there. We want to deliver results without commercial influence, filter bubbles or censorship. We're using APIs rather than trying to duplicate Google. We use NLP models based on transfer learning (GPT-2 for conversation models + some Amazon Lex for routing) to detect utterance intent and extract entities from queries (eg programming, knowledge graph or restaurant searches). We use a combination of routing logic and deep learning models (trained on public search query data) to predict the best sources for that intent / entity type / topic. Then we query the APIs for those directly (ideally for a direct match), or spider their search-result pages. We have a database of around 15K sites. We rank the results for topic match, content quality etc using models trained on example content. We use a BERT-style approach to extract key information or a direct answer within the top-ranked results. Examples: Computation queries => Wolfram Alpha or direct calculation. Knowledge graph => Wikipedia or DuckDuckGo Instant Answers. Programming => github, stackoverflow etc. Web searches from Bing or other web search APIs are 50%+ of results. We are building some specialized indexes (eg direct navigation). We are "broad but shallow" to start. We're in the OpenAI GPT-3 beta and excited to experiment more. We will make money three ways. 1. If a user buys something after searching, we may make a small commission from anonymous referral links. We will share revenue 50/50 with content producers used in the search. 2. We will have some paid Pro and Business plans in future. 3. We will provide a chat plugin that businesses can offer for search and navigation on public or internal sites. Try it out and see how the results work for you! Some fun features to try include recent events (try "is elon musk taking dogecoin to the moon"), direct navigation (try "go hn bitcoin" or "go youtube aws lambda"), or question answering ("how many baseballs fit in the moon"). You can also change how results display. Try visual results with feed or card grid views. If (like me) you love HN's minimalism, or miss Google circa 1999, try the "Hacker" or "Goggles" view on desktop (top right). Also try Reader Mode for a proxied ad-free experience. LazyWeb is in early alpha testing. It's free and open for anyone to use, and improving fast. We'd love your feedback, suggestions, bugs, improvements and anything else, and we're here to answer any questions you have!

vkkhare 5y ago
How do you train newer models then? From what I read you use public datasets to train your models but what about in future? You would need some kind of data collection mechanism? Gpt-2 and gpt-3 are great but the datasets they are trained would soon get old.
- jedwhite 5y ago
  Hey thanks. We will not log searches or collect personal data. There are public sources of search data that we can use with transfer learning against large scale language models like GPT-3, and that are updated regularly. Transfer learning works well without needing massive data sets with this sort of data (phrases mapped to intents). Having said that, the app tracks the intents and topic profiles of searches (not the search itself, just for example FoodPlaceSearchIntent) and whether the execution was likely a good result or not based on signals (like whether the search was likely repeated or rephrased - again without recording the actual search), and the models learn from that. We're adding signals including anonymized upvote/downvote as well. Approaches like differential privacy are something we want to pursue more in future. We are still very early days!
  - vkkhare 5y ago
    Makes sense though I wonder what would be the original source of that data (someone like google/microsoft must be logging user data and then making some parts of it anonymized and public). Maybe also look into on-device learning, it can be efficiently hooked up with differential privacy and give more specific results.
    - jedwhite 5y ago
      Yes, ironically, SEO industry resources can be helpful, and we used them in putting together training data. If you're interested there are some good simple free ones to get started also, like these from Mondovo: https://www.mondovo.com/keywords/ https://www.mondovo.com/keywords/ Brave browser uses aggregated search history data that's been anonymized, but we're not trying to personalize results (we're looking for objectively true, rather than "true for you"), so we're not trying to replicate ad-industry style personalization. A good set of labelled data matching intents to phrases helped us build some models simply that are surprisngly good at picking intents :)

vkkhare 5y ago

How do you train newer models then? From what I read you use public datasets to train your models but what about in future? You would need some kind of data collection mechanism? Gpt-2 and gpt-3 are great but the datasets they are trained would soon get old.

jedwhite 5y ago
Hey thanks. We will not log searches or collect personal data. There are public sources of search data that we can use with transfer learning against large scale language models like GPT-3, and that are updated regularly. Transfer learning works well without needing massive data sets with this sort of data (phrases mapped to intents). Having said that, the app tracks the intents and topic profiles of searches (not the search itself, just for example FoodPlaceSearchIntent) and whether the execution was likely a good result or not based on signals (like whether the search was likely repeated or rephrased - again without recording the actual search), and the models learn from that. We're adding signals including anonymized upvote/downvote as well. Approaches like differential privacy are something we want to pursue more in future. We are still very early days!
- vkkhare 5y ago
  Makes sense though I wonder what would be the original source of that data (someone like google/microsoft must be logging user data and then making some parts of it anonymized and public). Maybe also look into on-device learning, it can be efficiently hooked up with differential privacy and give more specific results.
  - jedwhite 5y ago
    Yes, ironically, SEO industry resources can be helpful, and we used them in putting together training data. If you're interested there are some good simple free ones to get started also, like these from Mondovo: https://www.mondovo.com/keywords/ https://www.mondovo.com/keywords/ Brave browser uses aggregated search history data that's been anonymized, but we're not trying to personalize results (we're looking for objectively true, rather than "true for you"), so we're not trying to replicate ad-industry style personalization. A good set of labelled data matching intents to phrases helped us build some models simply that are surprisngly good at picking intents :)