Last weekend I took part in a generative AI hackathon in london. Huge thanks to xx and  Victoria Stoyanova  and Sarah Drinkwater for organising it. (Sarah's writeup - https://betterprogramming.pub/how-to-run-a-generative-ai-hackathon-dc27f8d4fdd0). 

It was hosted by Entrepreneur First and sponsored by Amazon, the British Medical Journal, DeepMind, FacultyAI, Mind Foundry AI, OpenAI, the Royal Academy of Engineering, and WeTransfer.

It was the best hackathon I’ve participated in (closely followed by elife's innovation sprint from a few years back - http://partiallyattended.com/2018/05/18/elife_innovation_sprint_-_appstract.pub_/). 

In this post I want to dig into

•  What made the event great 

•  A reflection on the acceleration in capability that we have with technology compare to 10 years ago

•  Highlight the research and education hacks that were developed 

•  Give you a flavour of some of the other things that were built. 

Huge thanks go to the sponsors for making the event possible. 

Sahara and Victoria picked 120 people from the applicants. Everyone brought energy, creativity and created an inclusive space. That level of filtering was critical. 

I don’t know how many folk volunteers, but they were on hand, helpful, and made the event run smoothly. 

Streamlined comms to enable folk to help each other and to build psyche 

Big thanks to hacker one. That it was a 10 minute bike ride form where I live was perfect for me too. 

The pace of productivity that LLMs give you, and the landscape of platforms for quickly spinning up instances, is transformative (a lot of apps were run on Vercel, and I ran a few on glitch). What could have taken two or three days 10 years ago, now takes under a day. What would have been impossible to hack 10 years ago, now takes a few lines of python. 

https://valleyresearch.cc 

Write in a question, and this tool queries semantic scholar and summarises the paper using an LLM, but then also pulls out questions related to the question you asked, and lets you navigate through the literature based on these related questions. All of this is powered by the ability of large language models to extract meaningful summaries and relevant questions from the literature. 

This hack showed how LLMs could match someone with targeted advice or targeted types of therapy, with the aim of building a network of therapists. It's a two sided marketplace, but for getting people to the best mental health solution for them, at that moment. The ability to chat with the models was also showcased, allowing folk to have quick touchpoints with help. The thinking here is that the barrier to finding good mental health support is just too high, and this tool could help. 

You search for a topic, and it pulls information for you. You can then turn a complexity knob and it changes the level of complexity of the explanations presented. Aimed at students to allow them to find their own level with the material being taught. The pitch for this was brilliant, I think one of the best on the day. They talked a lot about the importance of tools like this for inclusiveness in the classroom. 

This team compared example treatment plans from doctors with best practice recommendations. The aim here is to help speed up the process through which health insurance companies validate these plans. At the moment these companies employ nurses, using up valuable nursing time. If you can make that bit of work more efficient you could free up nursing capacity. 

 https://panaceahealth.vercel.app/  

 

One of my colleagues, the very very talented Orlando, worked on this with some other folk from the hackathon. They tried to do a live demo, and it mostly didn't work at demo stage, but I have no doubt this was one of the most impactful hacks presented. As I understand it, they took data on rare diseases from BMJ Best practice, and created an artificial patient, THAT YOU CAN TALK TO! Doctors can now speak with a patient and get info in conversation, to help them try to diagnose the rare disease. This could be huge. 

llmsaregoinggreat.com 

I worked on this. It's a showcase of what could go wrong. The examples are mostly whimsical, but we got each of those examples up and running in well under an hour each. Working on this has significantly raised my concerns about this technology. The reason is that as humans we react viscerally to something that speaks to us (or chats to us), and it is now trivially easy to make these agents behave with malicious intent. 

I'm going to give a one liner for some of the other ideas. The thing to take away is that all of these were demoed working, all of them were built in under 24 hours. That is mind blowing. 

•  BrAIn ENO - mind mapping augmented. 

•  Spice-GPT - outputs LTSpice model of circuits to model, from text description to circuit description. 

•   Move2img - turn how people move into art. 

•  Dancing Spider -https://dancing-spiders-web.vercel.app - auto create short fun videos and graphics based on podcast content - helps with podcast discovery. 

•  Disinfo.ai - fact checking for a better world. - grammerly for bullshit.

•  Docq- put your company docs in there, then be able to use chat to find them - solves info retrieval in the enterprise. 

•  Lares - https://interconnected.org/home/2023/04/26/lares - giving LLMs working memory, and enabling them to control robots.

•  Quicklease - get the LLM to look for rental properties for you, and to email for you to book viewings. 

•  fAIght club - https://faight-club.vercel.app - a game - build you own monster, fight others. 

•  Dungeon hacker - help for dungeon masters to make dungeon mastering easier. 

•  ModeratorAI - help content creators understand feedback from their audience in real time. 

•  Examiner - generates exam questions on the back of uploaded PDFs - using langchain in the background. 

•  búho - help you understand leasing contracts, and ask question of the contract, like "am I allowed to have pets". 

•  customer.ai - build synthetic users based on product interviews. 

•  Algorthym -  LLM powered DJ.