AI summary by TwinMind
Why The Next AI Breakthroughs Will Be In Reasoning, Not Scaling
Y Combinator
35m 17s
AGI Timeline and Sam Altman's Predictions
- Sam Altman essay predicts AGI/ASI within "thousands of days" (4-15 years direct estimate)
- Same ideas Sam discussed in 2015 when starting OpenAI - sounded crazy then, plausible now
- Essay describes techno-optimist future: space colonies, climate fixes, abundant energy
- Core belief: AGI will accelerate scientific progress across all fields by being better at science than humans
- YC funded $600B+ worth of companies, sees AI developments firsthand
O1 Model Architecture and Capabilities
- Built on reinforcement learning techniques from OpenAI's DOTA work (beat professional players)
- Combines GPT generative approach with reward functions for factual correctness
- Chain of thought reasoning trained on massive dataset of task breakdown examples
- Current version is O1 preview; full O1 model coming soon with huge performance jump
- O2 and O3 models "not far behind" according to Sam
- Two parallel research directions: scaling base models + O1 reasoning approach
Chip Design Breakthrough - Diode Computer Demo
- Company automates PCB design through 4 major steps:
- System design/architecture
- Component selection (resistors, sensors, processing units)
- Schematic layout/placement
- Routing (NP-complete problem due to layer interference)
- Previous GPT-4 version: handled constraints, automated schematic design, simple routing
- O1 breakthrough: can now do system design and component selection
- Reads datasheets, selects correct components
- Takes high-level request ("wearable heart rate monitor with accelerometer")
- Outputs complete working PCB design
- Uses different models for different tasks: GPT-4 mini for PDF extraction, O1 for reasoning
- Same prompts that failed on GPT-4 worked immediately on O1
CAD Design Revolution - Camper Demo
- "Dev for CAD" - natural language to CAD design conversion
- Example: "Design 5 airfoils optimized for 50mph, minimum drag/lift ratio 15, 5° angle of attack"
- O1 writes/solves partial differential equations (Navier-Stokes) for airfoil design
- Runs multiple simulations simultaneously
- Built as desktop executable that controls SolidWorks UI automatically
- Represents step beyond O1 paper examples (which focused on verification/simulation)
Scaling Laws and Compute Investment
- Sam wants 4 orders of magnitude increase (up from previous 2 orders)
- Trillion dollar spend on training discussed
- O1 enables inference-time compute scaling - can spend more money/time for iteratively better results
- Models are "worst they'll ever be right now" - weekly improvements visible
- Chain of thought allows feedback on reasoning steps, not just final outputs
Customer Support Automation - Gigaml Case Study
- Indian founding team, pivoted from college applications → fine-tuning services → AI customer support
- Zepto customer: 30,000 tickets/day, 1000+ human agents, high turnover (few months)
- Previous implementation: GPT + rules, 70% error rate on complex cases, 0% on edge cases
- O1 + hardcore evals approach: 5% error rate overall, 85% accuracy on complex edge cases
- Massive improvement from 0% to 85% accuracy on difficult cases that consume most time/cost
- Demonstrates O1's ability to handle "messy" problems that rules-based systems couldn't solve
Startup Strategy and Moats in AI Era
- Key differentiators: proprietary data, comprehensive evals, UI/UX, integrations, distribution
- 10,000+ test cases on proprietary data not available online
- Enterprise sales to get "undercover" access to arcane/complex job functions
- Value in final 10% accuracy improvement - strongest technical teams capture most value
- Classic software moats still apply: switching costs, brand, data network effects
- Companies should target customers willing to pay premium for 100% accuracy
Technical Implementation Insights
- O1 uses fake reasoning traces in ChatGPT to protect competitive advantage
- Real chain of thought reasoning hidden from API users
- Need for interpretability and editability in future versions (O2/O3)
- Ability to edit individual reasoning steps would be next major unlock
- Jake Heller's prescription still applies: break tasks into steps + comprehensive evals
Industries Unlocked by O1
- Mechanical/electrical/chemical/bioengineering applications
- Physical world improvements through better math/physics capabilities
- Phone-based AI agents now working reliably
- Scientific research acceleration across all fields
- Complex problem-solving in regulated industries (aerospace, medical devices)
Potential Disruption Concerns
- AI coding agents may face challenges as O1 excels at programming problems
- Companies that invested in chain-of-thought infrastructure may lose advantage
- Need for abundance-focused applications to counter AI fear in society
- Race between technological progress and public acceptance
