Most People Get This Wrong: Vector Search
Most people pick the wrong one. What's yours?
Your question gets turned into a vector by an embedding model, and that vector is what the database searches.
Did you get it right?
#VectorSearch#Embeddings#RAG#AIEdShorts
Most People Get This Wrong: RLHF
Most people pick the wrong one. What's yours?
People rank two replies, a reward model learns those rankings, and that signal trains the answering model.
Did you get it right?
#RLHF#AIAlignment#LLM#MachineLearning#AIEdShorts
Most People Get This Wrong: RLHF
Most people get this wrong π
People rank two replies, a reward model learns those rankings, and that signal trains the answering
Pause and pick A, B, C or D, then drop your answer below π
#RLHF#AIAlignment#LLM#MachineLearning#AIEdShorts
Both AIs Said London. Only One Got Punished.
Two AIs gave the same wrong answer. Only one got punished.
One was 55% sure of LONDON, the other 99%. The right answer was PARIS. AI measures surprise with the natural log: the less it expected the right answer, the bigger the loss, 0.80 versus 4.61. Training is just shrinking that number.
Did you know AI learns by measuring its own surprise?
#NaturalLogarithm #AI #MachineLearning #DeepLearning #TechExplained
Watch a Giant AI Shrink Onto Your Phone.
Your phone can't run a giant AI. So a giant AI teaches a smaller one.
The big teacher shows the small student how to spot spam and turn speech into text, until the student can do both jobs alone, on your phone. That's knowledge distillation.
Which on-phone job do you use more: spam filtering or speech-to-text?
#KnowledgeDistillation #AI #MachineLearning #DeepLearning #TechExplained
Tesla's global fleet has just crossed 17 billion miles driven on FSD (Supervised).
Last week, Tesla updated this FSD mile counter to reflect growing usage. Its fleet now drives 1,000 miles every 1.8 seconds on FSD, or 18.2 billion miles annually, up 35% since July 2026.
Stop Trusting a Confident AI and watch this
This AI is 99% confident. And wrong.
The right answer got just 1% β so how surprised should the AI be? AI counts surprise with logarithms: every halving adds one loss point, so 50% confidence is a loss of 1.00 and 1% confidence is 6.64.
Training is just shrinking that number. Did you know AI learns by measuring its own surprise?
#LogLoss #AI #MachineLearning #DeepLearning #TechExplained