jdilla.xyz

Open AI Dev Day reflections

2023-11-07

I'm most excited about

  • Longer context windows
  • Lower prices
  • Custom GPTs. I've got ideas that range from silly to serious that I can't wait to try out
  • Text to speech – can't wait to try it out

Questions

  • How should I be thinking about documents at the assistant level vs. the thread level?
  • I assume the assistant level is more “core” knowledge… but how will this impact performance?
  • Should I be managing the interplay here between “core” things and stuff that just passes through the threads?
  • When would I choose to use the Chat API instead of the Assistant API?

Observations

  • Between Chat / Assistant / GPTs there are a lot of different levels to play in
  • The value layers (to me) seem to be 1) what unique data can you bring to the model from the real world and 2) what can you make it easy for the model to do in the real world (although this is less well understood so far).
  • If I'm understanding custom actions correctly (and it's possible I'm not), then I believe that custom actions can only be leveraged by the CustomGPT that creates them. So as an example, only developers with access to the Instacart API could make ChefGPT that creates orders for Instacart. If this is the case, I think this is mistake If I have a service that can do interesting things in the real world (e.g., Instacart), I want to be a tool that can be leveraged by ~all custom GPTs, not just the customGPT that I create. I suspect that this will get changed over time.

I wish

  • Assistants could leverage the web browsing tool

    • The mental health of Georgia Farmers

    2023-11-07

    From State Affairs:

    The Georgia Rural Health Innovation Center at Mercer University School of Medicine did an online statewide survey between January and April 2022 . The survey drew 1,651 people in the agriculture industry. It looked at their mental health and well-being. Here’s what the survey found:

    • 29% of farmers said they thought of dying by suicide at least once a month.
    • 42% of farmers have had suicidal thoughts at least once in the last year.
    • 47% said they experienced loneliness at least once a month.
    • 49% reported being sad or depressed at least once a month.
    • 39% said they felt hopeless at least once a month.

    Truly shocking numbers.

    Recap Roswell covered in Appen Media

    2023-11-01

    My side project, Recap Roswell was covered in Appen Media, our local news outlet this week:

    “I feel like Roswell, Georgia, the United States would be healthier if we have more of those institutions, or micro institutions, where people are connecting and going through the process of deciding what they think they should do together and then, trying to make that happen in the real world,” Dillard said.

    In May, Dillard built Recap Roswell from scratch using ChatGPT, an AI-powered language model developed by the nonprofit organization OpenAI.

    For each post, he creates a transcript from the video recording of the latest council meeting. Then, he runs the transcript through the OpenAI API, based on a chain of different requests which synthesizes it down to a quick summary.

    “You now are taking hundreds of hours of time and condensing it to something that someone, like yourself, can get through in 15 or 20 minutes,” Dillard said.

    I appreciate them taking the time to discuss the project with me!

    The rate of reports to child protective services

    2023-10-25

    I found these to be astoundingly high, so much so that I am wondering if there is an error in the data. From the National Library of Medicine:

    37.4% of all children experience a child protective services investigation by age 18 years. Consistent with previous literature, we found a higher rate for African American children (53.0%) and the lowest rate for Asians/Pacific Islanders (10.2%).

    I have no expertise in this area, but the point of the paper is trying to differentiate between first time investigations and follow up investigations, since presumably children that are investigated once are more likely to be investigated again.

    In 2014, 4.57% of all US children had a maltreatment investigation... Of these, about half (2.39%) had no previous investigation in the 2003–2014 database. After adjusting for database-first-time investigation rates as described in the preceding section, we estimate that 2.09% had a true first-time investigation.

    It occurs to me that even though I'm a parent, I'm not sure what exactly would cause a CPS investigation and what wouldn't.

    If this data is correct, I'm torn between being sad for a whole lot of families and wondering if we're over investigating.