Show HN: PreCog AI – Automatic AI Model Selection for Any Task
precog.ubik.studioHi HN,
I'm one of the co-founders of PreCog AI, a project my friend and I started to make the best AI models more accessible. PreCog AI is a chatbot that automatically picks and answers with the best AI model for whatever task you throw at it.
We made PreCog public on Monday and are getting great feedback. Originally built as an internal tool to help our small team reduce costs (paying for various chatbots) and get better AI output, PreCog has helped us so much with our workflow and ideation that we just had to share it.
Key Features of PreCog - AI Model Matchmaking: With access to 18 models, PreCog automatically matches your questions with the most fitting AI model based on the task.
- Versatile Adaptation: Works with any task, from coding to creative writing, giving you the right tool for the job.
-Ongoing Updates: Stay current with AI advancements using the latest LLM leaderboard data (we are constantly adding and changing our leaderboard). See the leaderboard here - https://precog.ubik.studio/leaderboard
-Preferred Model Selection: If you have a preferred model, choose it, and PreCog will use that model exclusively to respond.
How PreCog Works:
PreCog analyzes your query, references the model leaderboard, and then matches your query with the highest-ranked AI for that niche task. Delivering high-quality, task-specific output. PreCog's Model Leaderboard ranks AI models through over a million human comparisons, evaluated and presented on an Elo-scale. The dataset used to build the PreCog Leaderboard is from ChatBot Arena by https://lmarena.ai/. Researchers from UC Berkeley SkyLab and LMSYS developed the battle framework to produce the dataset.
I love feedback, questions, and critiques it helps me and my friend develop with the user in mind.
You can reach me at anytime at: info@ubik.studio
Do you find that there are a lot of variations and intricacies in deciding which LLM to use? I find it pretty simple to just assume sonnet is best at most coding jobs, o1 best at complicated non-coding tasks, 4o for simple questions that don't require planning. and well, tbh that's it, no other LLMs are that interesting.
We found that there is more nuance when it comes to different languages, we also update our leaderboard as new models drop so the best models are always available through PreCog. It's also nice to not have to tab switch constantly between chatbots and instead just use them in one central place. We added a feature so you can just pick the model you'd like to chat with if you dont want to be automatically matched with the leaderboard rankings (all matched responses provide an explanation in the reasoning behind the match so you can see why the model was chosen).
I think o1 is best at algorithm heavy coding tasks, opus at api/language-translation knowledge-breath heavy tasks (and also has better recency via later knowledge cutoff). But with latest update opus is pretty close at algorithm heavy coding.
We have plans to put o1 on the leaderboard - but Opus is there RN! We should have a new ELO ranking soon, but the second o1 ranking are done it will be on our leaderboard.
you're doing model routing - any thoughts on https://github.com/lm-sys/RouteLLM and Martian?
hope you dont raise funding before you figure out what they haven't
we have looked at route llm and played with it but we felt like they were focusing on cost minimization, we think theres more work to be done in focusing on routing for the best possible output without the cost constraint. Also just trying to provide easy ways for people to use tools like route llm that dont require coding knowledge. That being said we def wanna release a benchmark of our routing in comparison to some of the pre-trained defaults in route llm
cool. its a very small distinction in my mind to flip from one to the other. all the best.
I entered the query but was redirected to a login form. If you are honestly looking for feedback and no leads, unblock the app temporarily for HN. If you are for leads, for sure you will get some if this submission receives enough upvotes, but I wouldn't count on many. These days people are not so keen to leaving their data on random websites anymore.
Just took down the sign up for anyone who wants try it out! Thanks again for pointing that out - new to posting on here, super helpful hope you get some use out of our project.
Thanks so much for the feedback (new here for sure) - we are working on that rn, updating as fast possible gotta rebuild some stuff lmao
Ah yes, Minority Report, that story we're so eager to repeat.
fortunately we aren't policing anybody :* (we do love PKD tho) we thought precog perfectly described the function
I think it’s a great name, it really does describe the function perfectly.
I got a huge smile when I saw it.
tysm - I got a huge smile reading this comment.
That doesn't really change the fact that it's exhausting (and worse, "commercially offputting") to be reminded that we're careening towards the worst futures literally imagined. I stayed away from Soylent and I'll probably stay away from this, but thanks for the head's up. rimshot
As big PKD fans, that definitely flew over our heads a bit. Can def understand that view and understand why its is commercially exhausting especially because we agree that we are heading toward some of the worst futures possible, so did PKD. We definitely build with this in mind!
PKD is rad, my username is also a pkd reference. Love the Ubik studio as well!
Truly the GOAT. Love the username, love cyphers ;) So glad you liked the Ubik Studio!