Read on Twitter
We just pushed a major upgrade to @BloomBotAI, our @LangChainAI implementation of our open-source Tutor-GPT codebase
It's no longer just a reading and writing tutor...
It's your personal learning companion 🤓
github.com/plastic-labs/t…
It's no longer just a reading and writing tutor...
It's your personal learning companion 🤓
github.com/plastic-labs/t…
We received feedback on the bot being repetitive as a result of it's task-focused prompting
It was task-focused because we didn't know how to prompt for *objective discovery*
We'd been thinking about this for a while, but the answer was right in front of us the whole time...
It was task-focused because we didn't know how to prompt for *objective discovery*
We'd been thinking about this for a while, but the answer was right in front of us the whole time...
We were using very education-focused tokens in our prompts
Our dialogue was very pedagogical by design, and that was likely leading us to a point in the model's latent space that contained token distributions very biased towards traditional education
Our dialogue was very pedagogical by design, and that was likely leading us to a point in the model's latent space that contained token distributions very biased towards traditional education
Through this lens, it makes sense that its behavior reflected that of a well-trained educator
But we want to give students superpowers, not embody all the teachers they've likely had...
We had to re-think our prompting
But we want to give students superpowers, not embody all the teachers they've likely had...
We had to re-think our prompting
Leaning into the theory of mind capabilities, we made the (counter-intuitive?) decision to pay *more respect to the model's intellectual capabilities*
We are letting the model *write its own system prompts* based on what IT thinks the user needs
(user, not student - less bias)
We are letting the model *write its own system prompts* based on what IT thinks the user needs
(user, not student - less bias)
This made all the difference
GPT-4 is able to both make a reasonable prediction as to why someone said something AND list out things it'd need to improve its prediction
It's constantly updating its prompts - it never has to "decide" you have a specific objective
(v bayesian)
GPT-4 is able to both make a reasonable prediction as to why someone said something AND list out things it'd need to improve its prediction
It's constantly updating its prompts - it never has to "decide" you have a specific objective
(v bayesian)
Even more impressive - its ability to determine whether to act on its prediction or gather more information from you
It can literally 💃 catch a vibe 💃
It can literally 💃 catch a vibe 💃
Join our Discord and try out Bloom yourself: discord.gg/bloombotai
• • •
Missing some Tweet in this thread? You can try to
force a refresh
