@ScottMoura@paulg It’s interesting how we have to start challenging our assumption that getting great PhD students has anything to do with great letters - high quality string outputs is not the essence of a successful researcher.
@jetdillo I don’t think we are looking at “What went wrong with conversational AI”. We are learning the vulnerabilities and safeguards. Remember when nobody knew about cross site scripting?. These “attacks” are just the AI-era Bobby Tables: https://t.co/ufXIPixWfg
From the Wet Pavement Causes Rain Dept:
How surprising it is that a language model parrots out edgy and tortured answers to questions on self consciousness!
They’re so uncannily similar to sci-fi tropes we’ve come to expect! And written so much about in blogs and fiction!
We (the Real World Reinforcement Learning group at Microsoft Research) are looking to hire, even in a year of serious economic uncertainty: https://t.co/LxeJxvSY7b , including intern, postdoc, and researcher positions. Please apply by Dec. 9th if this matches your interest.
The Transformer is a magnificient neural network architecture because it is a general-purpose differentiable computer. It is simultaneously:
1) expressive (in the forward pass)
2) optimizable (via backpropagation+gradient descent)
3) efficient (high parallelism compute graph)
When people fully absorb the impact of this stunning new brain study @AllenInstitute, a few years hence, they will wonder what neural net researchers were even trying to do in the early 2020s, oversimplifying so much vital neural complexity.
great🧵on 1000s of types of neurons
@judsonalthoff@Haleon_health@Microsoft 🙌 to the SeeingAI team!
Wondering what wonderful things they’ll invent with the combination of mobile HW features, large language models, and multimodal audio capabilities.
Affordability and easy access to assistive technologies is a game changer, phones+AI enables that.
@mikeymckay If you’re into Winograd- have you seen Understanding Computers and Cognition? I find it is an underrated building block, with essential elements of AI agent design, and a nexus between the Silicon Valley and Flores-Maturana realms.
Models and APIs that are based on natural language, images and sounds are a first step to phenotropic programming and interoperability.
Next tie them RL to agents with episodic Winograd’s speech-act state machines and things start getting interesting.
@mikeymckay But for example take the work @natfriedman was doing. With gpt3+ you could do a natural language interface to a travel bookings site. That is infinitely more robust *for computer to computer interoperability* than some dumb json-based static APIs to find and book travel.
@profgalloway This is great only for those in privilege, regardless of their intent towards others.
I’d love to see a meta-study on who brings this topic up again and again, and the arguments and counter-arguments brought forth over time?
This morning we kicked off our new project "Adapting digital community scorecard for Strengthening the Voice of Women Garment Workers (SVW)”. @CAREinCambodia, #YCC, #ACT, #CWCC, #CWPD#worker-voice
@chrimbs@WardCunningham TDD enables a form of PP that has a rhythm and structure - it makes it natural to have a dialogue where you and I discuss what we are seeing versus what we have learned in our career. That way it balanced power and creates a playful focus on the now.