AI
Technology
18 rabbit holes · 9 people exploring
Common starting questions
- Why did Clippy fail?
- What's the difference between a workflow and an agent?
- How do you test an AI product before the AI works?
Rabbit holes
- the Mechanical Turk was a person in a box
A chess-playing 'machine' that toured Europe for decades and beat Napoleon and Benjamin Franklin. A chess master hid inside. Amazon named its crowdwork service after it, which is either honest or bleak.
Wren Abernathy · 4 sources
- from Deep Blue to AlphaZero to neural Stockfish
Kofi Asante · 6 sources
- you trained AI every time you solved a CAPTCHA
First it was digitizing old books and newspapers (one known word, one unknown), then Street View house numbers, then street signs and traffic lights. Then it got less useful because AI got better at them than us.
Ayodele Bankole · 4 sources
- Clippy was built on serious research and everyone still hated it
Arjun Pillai · 6 sources
- every benchmark eventually becomes a training target
'When a measure becomes a target, it ceases to be a good measure.' Applies to ML benchmarks, school test scores, hospital wait times, and sprint velocity.
Arjun Pillai · 5 sources
- finally understood backprop thanks to 3blue1brown
The visuals make it click in a way textbooks never did for me. Rosenblatt's perceptron in 1958 was hyped as basically an electronic brain by the New York Times.
Arjun Pillai · 5 sources
- why robot hands are still so bad
A human hand has around 27 bones, dozens of muscles, and thousands of touch receptors. Most industrial robots get by with two-finger grippers and suction cups for a reason.
Omar Haddad · 4 sources
- The Chinese Room argument
Pete Lindgren · 5 sources
- ELIZA fooled people in 1966 with ~200 lines of code
Weizenbaum's own secretary asked him to leave the room so she could talk to it privately. He was horrified, not flattered.
Ludovica Ferri · 5 sources
- the Roomba came out of a different idea of AI
Omar Haddad · 5 sources
- conversational interfaces were already a thing in 1970, with blocks
SHRDLU let you type to a simulated robot arm in a world of blocks and pyramids, and it could answer questions about what it had done and why. It looked like a huge leap but only worked inside that tiny world. A few years later text adventures put a (much dumber) parser in front of a lot of ordinary people.
Callie Sutherland · 7 sources
- what actually makes an 'agent' an agent
Arjun Pillai · 5 sources
- reading the transformer paper 8 years late
Finally sat down with the actual paper instead of explainer threads. It's a translation paper! Nobody in it is talking about chatbots.
Arjun Pillai · 5 sources
- Netflix paid $1M for an algorithm and never fully used it
The winning team tied a rival on score and won by submitting about 20 minutes earlier. Then Netflix said the full winning ensemble wasn't worth the engineering effort, especially since they were moving to streaming, where what you actually watch says more than star ratings.
Arjun Pillai · 5 sources
- Clever Hans and why ML models cheat
A horse that 'did arithmetic' by reading tiny cues from the questioner's body. ML models do the same thing: a classifier that 'detects pneumonia' by reading the hospital's watermark on the X-ray.
moth · 4 sources
- Moravec's paradox: the hard stuff is easy and the easy stuff is hard
Arjun Pillai · 4 sources
- the AI winters
Two big funding collapses (mid-70s and late 80s), both after hype ran ahead of results. Reading this right now hits different.
Arjun Pillai · 5 sources
- IBM tested a speech-to-text typewriter in 1983 by hiding a typist in the next room
The 'listening typewriter' study: people dictated letters to what they thought was a speech recognition system while a hidden typist typed what they said. The point was to find out if anyone would want it before building it. Faking the hard part with a human is still a standard trick for testing voice and AI products.
Arjun Pillai · 5 sources