AI summary by TwinMind
Distinction between small & large models will go away – Sholto Douglas & Trenton Bricken
Dwarkesh Patel
5m 42s
Future Model Architecture Predictions
- Distinction between small/large models may disappear
- Long context could eliminate need for fine-tuning
- Current landscape: different model tiers + fine-tuned specialized models
- Future vision: dynamic compute bundles + infinite context for specialization
Long Horizon Task Performance vs Context Windows
- AI progress bottleneck: inability to perform tasks over long horizons (hours/weeks/months)
- AI agents haven't taken off due to reliability issues, not context limitations
- Problem: insufficient "nines of reliability" for chaining tasks successfully
- Current models (GPT-4, Gemini Ultra) insufficient for agent reliability
- Next model scale increment may provide extra reliability "nine" needed
- Context windows necessary but not limiting factor currently
Agent Architecture vs Monolithic Models
- Question: Multiple model copies communicating vs single large adaptive model
- Two factors favoring monolithic approach:
- Long context enables ingesting information no human can process
- Models are general-purpose (same GPT-4 for all tasks)
- Eliminates specialization problem (front-end vs back-end engineers)
- Future AI firm potentially = single model vs multiple connected agents
Near-Term Agent Implementation
- Short-term: agent-based architecture preferred
- Humans need isolated, reliable, trustworthy components
- Reliability challenges with task chaining
- Success probability multiplication: 99.99% → 99.9% per task dramatically reduces overall success
- Need better understanding of success rates over different time horizons
Long Horizon Task Evaluation
- Important to measure success rates across time resolutions (minutes/hours/days)
- Enables assessment of job/task family automation potential
- Better than current benchmarks for judging economic impact
- Context window evolution: 100K introduced <1 year ago, surprising progress rate
End-to-End Training Scenarios
- Hypothetical: AI firm trained on profit signals or client satisfaction
- Architecture firm example: success measured by client blueprint approval
- Internal agents: sales, design, editing specialists
- Reinforcement learning dream: sparse reward signals sufficient
- Reality: requires "incredible amount of care and diligence" from humans
- RL reward dependency: model must generate some reward to learn
- Future models will achieve sufficient reliability for reward generation
