Elicit

??

If you’re a researcher and you haven’t tried Elicit - particularly when you need to understand an unfamiliar field rapidly - you’re in for a treat! Apart from ChatGPT, this is the only AI tool I pay for.

I had the pleasure of recently presenting a very well attended presentation on using contemporary AI tools to speed up your research at American Psychological Association's 2024 conference in Seattle. It was literally wall to wall packed (and yes, I used AI to blur faces in the photo). All those who attended were welcomed to write a paper with me using AI on a variety of topics (who will be reaching out to shortly!). The whole presentation is available here: https://osf.io/bnefc Shoutout to R Discovery, Elicit, LM Studio for being the most popular products people talked to me about. Elicit especially was loved.

"Be Human" is one of Elicit's core values - and this is one of the surprising ways it pops up in our day to day. We have a blast celebrating each person's individual superpowers and quirks, and see them as a key part of what makes our team uniquely strong. Want to join a company that values authenticity? You can learn all about the values that shape us at elicit.com/careers We can't wait to create your user guide next!

What’s a human-centric startup tradition you’ve instituted? One of my clients, Elicit, puts a doc together for each of their employees’ birthdays and has the entire team add quotes, stories, and nice words about them. They call it the birthday person’s “user guide”. This user guide is then reviewed live during their Friday morning team sync. This week Elicit celebrated Ben Rachbach's birthday. It was so special to see the team’s genuine appreciation for someone like Ben. He was all smiles and it was awesome to also see his gratitude on full display for all the nice, thoughtful words that were written about him. I’ve had the pleasure of working directly with Ben as he’s very involved with helping Elicit’s enterprise customers be successful in cutting down the manual work that’s typically involved with systematic literature reviews. Every client he comes in touch with loves him and he takes so much personal pride in giving them AI superpowers that actually work ?? After every customer call we’re on together he’s also hungry for feedback and has a growth mindset that is rarely seen. There’s so much more I can say but I want to add that he’s also just generally a good human, which was echo’ed consistently across the team within his “user guide”. Happy birthday Ben it’s been a pleasure collaborating with you, Jungwon Byun, Andreas Stuhlmüller, Kevin Bird, Sarah Park, James Brady, and the entire Elicit team! What would your user guide say? #fridayperspective #startuptraditions

"AI tools that actually work" is about the best possible compliment you can get as an AI tool right now

?? We’ve substantially improved Elicit’s answers across the app: - Summaries of top papers now have 78% fewer errors - Abstract summaries better follow your query with 88% fewer errors - Standard custom columns now have 37% fewer errors than before All the features above can be used an UNLIMITED number of times for free, thanks to the pricing changes we rolled out in the past few weeks. We’re excited to see these new improvements speed up your work and help you find research insights faster!

We invest a lot in evaluations at Elicit. We want to be known for our accuracy and overcome the limited truthfulness of LLMs. So we’ve built a suite of tools to understand LLM quality for research tasks demanding high accuracy. We’ve tried every single option there is - hiring contractors, working with expert users, using existing benchmarks, creating our own benchmarks, etc. etc. Now, models are capable enough that they can evaluate other models. But how do you trust the evaluators? In this post, étienne Fortier-Dubois lays out the trust-scalability-flexibility tradeoff with evals and how we managed to scale up evaluation to support constantly improving models, over a dozen different tasks, and hundreds of thousands of LLM outputs per day.

There's no short-cut to ensuring that Elicit gives accurate answers to questions, but there are better and worse strategies. In this post, we describe a few approaches we've tried and what we currently do to make Elicit the most trustworthy tool out there.

Finally out - Elicit's new search is much better for technical searches and nuanced queries E.g. "arguments against universal basic income" surfaces papers that actually argue against UBI, not just vaguely related papers