GPT-x, is it really worth it?
summary
Shows that GPT-3 runs with 175 billion parameters trained on ~300 billion tokens and that training is estimated at $10M, while single queries on Davinci cost about $0.01 and long chats quickly balloon to ~$0.50 per session. Demonstrates real failures: confident fabricated facts (Leonardo murdered), repetitive persona shifts, and confused comparisons (cars vs airplanes). Builds a practical fix path by fine‑tuning smaller models—distilled GPT‑2 works on a 24GB GPU—but concludes raw GPT‑3 needs wrapper controls to handle hallucinations and bias.
AI-written summary · about
chapters (12)
- 0:00Introduction to GPT-x
- 4:48How GPT models work
- 6:55Zero-shot learning in GPT
- 8:27Context and examples
- 16:35Cost of using GPT-3
- 20:38Knowledge limitations
- 24:13False information risks
- 26:45Context length issues
- 27:50Cost of long conversations
- 31:09Bias in GPT responses
- 34:32Why biases occur
- 39:36Conclusion and recommendations
related talks
39mCollect Data to Fine-Tune Your Large Language Model
52mBeyond the AI Hype: What's Real, What's Next
47mThe Rise of AI: A Reality Check on Generative AI
38mIntroducing the New Colleague - ChatGPT
46mConversations with ChatGPT: illusion or reality?
41mInside GPT – The Maths Behind the Magic
52mUnleash the Power of Large Language Models (LLMs)
38mArtificial Intelligence: How Much Will It Cost You?