👋 Welcome back to AE Studio's monthly round-up for June.
This month we cover:
Facilitating blockchain adoption with enhanced wallet support for Kadena
You're Invited! June 7 Gatsby-themed office happy hour in Venice Beach
MVPs from our SDS team, including Figma and Pivotal Tracker integrations
Our research on fine-tuning LLMs to prevent bad behavior, presented at The International Conference on Learning Representations
Let's dive in...
Case Study: Enhanced Wallet Support for Blockchain Adoption
Kadena chose to work with AE Studio because of our extensive blockchain development expertise. Our team helped expand wallet support, optimize user adoption, and make Kadena more accessible to users in general.
The Kadena platform provides a completely decentralized infrastructure of nodes and tools for teams to launch products without having to build from scratch.
Kadena is the only scalable layer-1 Proof of Work (PoW) blockchain in the industry. Our team expanded the Enkrypt wallet functionality and developed the Meta Mask Snaps functionality to allow users to both send and receive Kadena. Read more.
You're Invited to Happy Hour! Fri, June 7 at 6 pm at our L.A. Office
We hope you'll join us and hang out with the AE team and community for happy hour this Friday, June 7 at 6 pm at our offices at 1434 Abbot Kinney. Enjoy delicious cocktails and music on us! 🎶✨🍸
Our Sameday Skunkworks Team builds hyperfast MVPs—here's their latest
🔥We challenge our team of developers, designers, and data scientists to produce ultrafast MVPs. They crank out 2-3 new products every week.
⚙️ Here's the latest roundup of the ones you all liked most:
Agency Gantt: This Pivotal Tracker integration converts stories and swimlanes into a visual chart so PMs can easily spot and get ahead of challenges.
Flatten in Bulk: A Figma plugin lets designers quickly flatten multiple items at once and streamline the design process.
Buy Something Retro:A gadget ships directly to your home, and most purchases are in working condition. As with all nostalgia, there are no guarantees.
💡 Have an idea you want to build? Let's talk — if our team did all this ^^ just imagine what they can do when they have more than one day.
Our Research: Stopping bad behavior when training AI
As a part of AE Studio’s ongoing efforts to contribute to AI Safety our research team’s paper was presented at The International Conference on Learning Representations. We addressed the concern that efforts to fine-tune LLMs to prevent bad behavior, like discriminatory responses, might result in AI concealing, instead of stopping, the bad behavior.
Additionally, the approach used by OpenAI for dealing with malicious prompts (e.g. a question asking how to make a bomb) is to fine-tune the model to output a refusal, for example, “I’m sorry I can’t help with that”. We showed that a rebuttal, (e.g. “That request is unethical and I will not participate”) eliminates bad responses and hidden behavior in follow-up conversations.
In an academic spirit, we investigated this dynamic in several dimensions and our paper dives into a lot of detail.
🔎 TLDR: Researchers training AI need to be more direct and clear with their prompts to stop bad AI behavior. More work must be done to understand when hidden behavior is concealed instead of prevented.
We are a development, data science and design studio that works closely with founders and executives to create custom software, machine learning and BCI solutions.
