John Bohannon is both a journalist (Science, Wired, The Guardian) and an AI practitioner, specifically in NLP, as Director of Science at Primer.
As I’ve begun making more videos hosting interesting people I thought I would sit down with John to tap his experience on the process of interviewing. I also wanted to check if AI is going to soon replace the journalist-interviewer in which case I could free myself to go bark up some other tree. Spoiler: the answer is a resounding “hell no, we’re not even close yet!”.
So here’s the show:
The notes below give an outline of some of the topics we cover:
Reverse of Turing Test
An “Interview” is an operation to extract information from a person
Several modes of professional interviewing
Need an answer to a specific question
I know a question I’m driving at and want to explore where it may lead
Cadaver artwork project
“Where are you getting these bodies?”
“Are you getting these bodies from the Chinese government?”
“Are you getting these bodies from prisons and insane asylums?”
More to the art of interviewing than “what to say next?”
GPT is trained to figure out the next thing to say.
Transformers and attention windows:
How good would an AI interviewer need to be to be able to setup the next interview? Realize the next person to talk to is X and not Y.
OpenAI is actively working on something akin to a longer-term memory
Who are you if not your memory?
What’s hard about this
How do you figure out what’s relevant?
How do you decide the right strategy of which questions to ask, who to follow-up with for more clues? How do you frame the follow-ups to get the answers you want?
GPTs have no “goals”
Reinforcement Learning - crystalized sense of goals
Sensors, processing, and prediction
Two different worlds - synthesize these to
Say-Can at Google: Say Can
Adobe - audio AI
Whisper - speech-to-text from OpenAI
Video is a data paradise.
Text summarization
“Nut graph” in journalism - 2nd or 3rd paragraph that summarizes the whole article
Bert changed everything - 2020 got crazy good
No ranking of sentences/grammar
InstructGPT - RLHF - take user feedback and shape the way the model outputs text
Cambrian explosion of text2text models - how cheap and efficient can you make that?
Eleuther - open source version of GPT-3 called GPT-Neo
Deepmind consistently coming out with paradigm shifting AI - Flamingo - computer vision married to NLP
Old methods of authenticity signaling - power get consolidated to existing platforms
Give up on truth or give up on anonymity - both have huge consequences
Manufacturing consent
Driver’s license for the Internet
Early Internet was like a sparse suburban childhood
Stretching the tether from one person to the other - comments on a YouTube video - you’re the worst version of yourself
AI attempting to host a journalistic interview
“Make a chatbot that’s worth talking to”
The passive data stream
Gathering the facts: important across crisis management, triage situations, emergency healthcare, national security
Who, what, where, when, why?
Automated simple journalist - reads the work of other journalists or reporting of any kind on events - figure out what are called “casualty events” (people getting hurt or killed)
Inferences are expensive to run real-time stream on everything daily
Intelligence gathering - a lot of people are going to have to change jobs or adapt
AIs looking for problems
An attempt at an AI-driven media company