Towards a Pragmatic Understanding of Large Language Models
summary
Argues that modern LLMs are fundamentally a next-token predictor whose surprising breadth comes from scale, and proposes treating them as function approximators searched in latent space. Gives three concrete claims: fluency is decoupled from expertise, models do not care so humans must retain semantic control over output. Describes a pragmatic benchmark that runs dialogue games to test conversational competence, cites transformer/BERT breakthroughs (2018), mentions Open AAI, Llama by Meta and Almo/Elen AI releases, and recommends design fixes: friction, local/private deployment, and licensing to pay creators.
AI-written summary · about
chapters (8)
- 0:00Introduction and motivation
- 7:31History of language models
- 11:26Current LLM limitations
- 15:24Panel discussion on AI definitions
- 27:32Positive and negative AI impacts
- 36:48KDE’s role in AI development
- 40:01Audience Q&A on data and resources
- 50:01Closing remarks
related talks
01h 11mThe Move Toward AGI: Why Large Language Models Surprised Almost Everyone... | Theoretically Speaking
49mA Practical Approach to Using LLMs: Peculiarities and Pitfalls
30mLarge Language Models and Prompt Engineering for Developers
33mOpening Keynote: LLMs, between fantasy and reality - the poin...
40mThe Future of Language Models
43mBeyond the Hype: A Realistic Look at Large Language Models
40mFrom Simulated Subjectivity to Collective Consciousness in Large Language Models
44mBuilding projects with language models