Andrew Reece | PhD Behavioral Data Scientist

Introducing the CANDOR corpus from @BetterUp: A 1TB, 850hr audio-video dataset of 1,656 unscripted conversations in America in 2020. CANDOR = Conversation: A Naturalistic Dataset of Online Recordings arxiv.org/abs/2203.00674

@hwchase17 for Chat Agents using retrieval_with_sources tools, is it possible to store sources in memory? Intermediate steps don't save into history and Agent Final Answer keeps dropping source info. Thanks!

CANDOR corpus is now available! It took years of hard work, but we hope it will be useful for researchers from many fields interested in conversation and social interaction. Dataset available for download (link in manuscript). Please share it widely. science.org/doi/10.1126/sc…

In sum, if it's going to be robot overlords I at least hope it's the one who doesn't have to Google my pleas

GPT-3 (first answer): def square_root_of_sum(foo, bar, baz): return math.sqrt(foo + bar + baz) Blender: 0 (but why) GPT-3: 5 (bad indentation forgiven)

I asked 5 hard questions to two modern language AIs - Meta's new BlenderBot (blenderbot.ai) and OpenAI's GPT-3. Warning: Gets a little computer nerdy, but there are riddles and metaphysics too! Answers and commentary in :thread:

@MeirSimchah Basically histograms of behaviors for good (blue) and bad (red) conversationalists. Y-axis = freq. a given behavior occurred, at different lvls of extremeness (x-axis bins). So high-right blue dot on vocal intensity means good Cs more frequently expressed high VI than bad Cs.👌

@_andrewreece I don't understand these charts yet... Can you clarify what's happening on the x-axis?

Introducing the CANDOR corpus from @BetterUp: A 1TB, 850hr audio-video dataset of 1,656 unscripted conversations in America in 2020. CANDOR = Conversation: A Naturalistic Dataset of Online Recordings arxiv.org/abs/2203.00674