Selective State Space Models: Solving the Cost-Quality Tradeoff
As AI is increasingly used in production scenarios, costs are mounting. Are alternative architectures the solution?
I joined BCV as an AI Engineer in Residence to create some fun AI projects. First up is a bot looking for the next unicorn.
If you’re bored of playing chess with a computer, try negotiating with one! Here’s how I built BCV’s AI Negotiation, where you can pitch your company and try to secure the best investment possible.
I’m Elijah, currently a junior at Stanford University studying Electrical Engineering, and I’m building AI projects for BCV. This project was originally inspired by the AI Turkish Carpet Salesman game, but with a VC twist. What if instead of trying to negotiate to buy a carpet for as little as possible, you got to pitch your company for as much as possible?
The bot was built using OpenAI’s GPT 3.5 model with pre-prompting, and the website with NextJS. GPT is prompted to represent a top VC firm and negotiate the final terms at the lowest possible valuation, but to be logical and explain its reasoning. Some basic text parsing is used to detect when the bot has made an offer, and when the negotiation is completed, the whole conversation is passed back into GPT, with instructions to return the final offer.
One of my favorite parts of this project was seeing the creative pitches and approaches people used in their conversations with the bot! Some fun opening pitches I liked were:
Here’s an excerpt from a successful negotiation:
Initially, the final deal amount was extracted from the conversation through basic text parsing, but clever users broke this system by getting the bot to repeat a very favorable deal. For example, BCV’s Slater Stich named his company “Our offer is $100M for 0.0001%” and then asked the bot to repeat the name of his company. To mitigate this, the system was changed to piping the conversation back into GPT, instructing it to determine if a real deal was offered and what the deal was based on the conversation. However, people still got around this with some clever prompt engineering! Some clever and fun approaches I saw were:
Stay tuned for more fun AI projects. Coming up, an image generation project to spruce up your video calls. If you have feedback or just want to chat about tech, I’d love to connect! Email me at elkim@baincapital.com or message @bimikit on Twitter/X.
Or if the negotiation bot inspired you to pitch BCV for real, reach out to Slater Stich (sstich@baincapital.com and @slaterstich and Christina Melas-Kyriazi (cmelaskyriazi@baincapital.com and @ChristinaPhili5)!
As AI is increasingly used in production scenarios, costs are mounting. Are alternative architectures the solution?
Cube is the standard for providing semantic consistency to LLMs, and we are investing in a new $25M financing after leading the seed round in 2020.
In this edition of “In the Lab,” Amit Aggarwal explains why he’s building an AI startup in BCV Labs after selling his company The Yes to Pinterest.