Generative AI

From DevSummit
Revision as of 18:31, 9 December 2023 by Gunner (talk | contribs) (Created page with "Generative AI How are you using GenAI? * Academic. Misconduct & creating * Noticing a lot of garbage coming out. * NP Exec looking to make things more efficient but not su...")
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to navigation Jump to search

Generative AI

How are you using GenAI?

  • Academic. Misconduct & creating
  • Noticing a lot of garbage coming out.
  • NP Exec looking to make things more efficient but not sure how.
  • My NP Exec needs support writing a grant proposals. Projections, etc.
  • Preparing for a bookclub, for a book I didn't read.
  • Making coding fun again. Writing emails. Writing blog posts. Use it every day.
  • Interested in AI for accessibility for low literacy communities.
  • Worried about abuse of communications channels, when it's so easy to create content --> volume of spammy / advert / direct mail / sms.
  • Worried about using internal docs to generate docs --> privacy concerns.
  • Generating a list of talking points. 4 were good, 2 were opposition args or greenwash.
  • Desiring shared vocab about what GAI is and isn't.
  • Draft agreements at work, with HR director.
  • Want to learn to build my own system people can interact w/, with my own data set
  • Creating custom agents.
  • Using for qualitative analysis.
  • Use API to make models
  • Look at LLM and GPTs as necessary weapon.


George

  • From the school of folks who mock the new stuff
  • Everyone should be learning Prompt Engineering
  • Using OpenAI's API to use your own data


What is prompt engineering?

  • Tell it how to act / who to act like ("Llike a security researcher with a focus on accessibility")
  • Give it some knowledge / context ("THere were x breaches in 2020")
  • Tell it what to do ("Make XYZ")
  • Tell it how to do it ("as a list")
  • Tell it what not to do ("Don't do this")


How to get started with Prompt Engineering?:

  • A lot of resources online. YouTube videos. Context driven videos.
  • Look for things that explain how it works. How prompts get chained together.


The tools:

  • Bard. It's OK. Lacking when it comes to tone and the options you might be looking for. Doesn't do a good jog with a lot.
  • ChatGPT: task decomponistion. generate articafcts. gernate document or SVG file or some code.
  • Claud.ai - "constitutional AI" - results are a little more rich / natural sounding in tone. I'll use that for documents / things that are creative.
  • Microsoft Edge has GPT4 built in. I have questions about privacy there (and in general)
  • Anthropic: forked by moral people at OpenAI
  • Surveys being done:
  • determine hallucinations
  • open AI way ahead, bard way behind


Concerns:

  • job replacement concerns
  • There are companies offering worker replacement with AI
  • Long list of folks whose jobs are at risk - or need to evolve
  • Are soft skills more important now, b/c relationships can't be replicated? Or hard skills more important b/c you gotta check the bullshit?




  • Privacy concerns:
  • Meredith Whitticker. Head of Signal. Union org at Google. Says big tech is sucking in our data.
  • Is it ethical to include client data in ChatGPT?
  • OpenAI just said "we'll defend you if someone sues you?"
  • We have in our contract that your data can go to OpenAI
  • You can make agreements w/ OpenAI that they won't use your data for training
  • How can you get data out of a LLM when you want to "fire" a client for example?
  • Apple bans use of OpenAI internally
  • Progressive usage
  • OpenAI is based on existing logic. Progressive work is based on putting forth a new kind of thinking.


Limits of generations of models:

  • 7 mo ago it was hard to give it a lot of context in one go - to GPT3.5
  • Give it a lot of prompt chains
  • Keep giving it context


Context

  • You are the context expert. If you don't know plumbing in NYC, you're not going to get GPT to do that for you.
  • You also need context to spot whether it's right, or spot if it's bullshit.