← All drillsGenerative AI in Text, Image and Sound
What you'll be able to do- ✓Read the probabilities a language model gives its next token and change how it writes with temperature, top-k, top-p and a seed
- ✓Compare a bare instruction, a system message and worked examples on the same task
- ✓Recognise a confident invented answer and see what putting the evidence in the prompt does and does not fix
- ✓Follow an image model from random noise to a picture, and change the seed, the prompt and the guidance scale
- ✓Turn text into speech, transcribe it back, and measure the round trip with word error rate
A guided tour of the three kinds of generative AI behind today's assistants, image tools and voice features, with short working examples you can change, and an honest look at where each one fails.
⌁ Language models write text one token at a time, diffusion models make images by removing noise, and speech systems pair a speech-synthesis model with a recogniser such as OpenAI's Whisper. The same ideas sit behind assistants, image generators and voice and captioning features from OpenAI, Google, Microsoft and Stability AI.
Start this internshipCreate an account to unlock the 4 sections, the workbench, and askThili.
BeginSections
1. Why this course: from what you teach to how machines now write, draw and speak
🔒 locked2. Text: How a Model Writes, and How to Ask It
🔒 locked3. Image: From Noise to a Picture
🔒 locked4. Sound: Speech Out, Speech Back
🔒 lockedDig deeper
📄The Curious Case of Neural Text Degeneration (Holtzman et al., 2019): top-p sampling
paper📄Language Models are Unsupervised Multitask Learners (GPT-2, Radford et al., 2019)
paper📄Denoising Diffusion Probabilistic Models (Ho et al., 2020)
paper📄High-Resolution Image Synthesis with Latent Diffusion Models (Rombach et al., 2021)
paper📄SpeechT5: Unified-Modal Encoder-Decoder Pre-Training for Spoken Language Processing (2021)
paper📄Robust Speech Recognition via Large-Scale Weak Supervision (Whisper, Radford et al., 2022)
paper