Episode, Podcast
Why The Next AI Breakthroughs Will Be In Reasoning, Not Scaling
Y CombinatorJared, Harj, Sam Altman, Diana Alvear, Mark Mandelmann, Francesc Campoy Flores, Melanie Warrick, Kendrickson Jr., Harshi, Jake Heller
- Sam Altman predicts the arrival of Artificial General Intelligence (AGI) and Artificial Super Intelligence (ASI) within 4 to 15 years, a timeline described in his recent essay as "techno-optimist" and comparable to his 2015 forecasts which were previously dismissed.
- Altman originally posited that AGI would accelerate scientific progress by outperforming humans in fields like physics, a motivation that drove the development of OpenAI's O1 model.
- OpenAI's O1 model demonstrates a "step function" capability unlock in reasoning, specifically in engineering domains where previous models like GPT-4o failed.
- In chip design, O1 successfully automated the system design, component selection, and routing phases, solving the NP-complete routing problem that typically requires thousands of engineers at firms like NVIDIA and Intel.
- The startup Diode Computer utilized O1 to generate fully functional PCBs from high-level prompts (e.g., "wearable heart rate monitor"), a task that failed entirely when attempted with GPT-4o using identical prompts.
- In mechanical engineering, Camphor leverages O1 to generate and optimize CAD designs (e.g., airfoils) by writing and solving partial differential equations in real-time, functioning as a direct co-pilot to SolidWorks.
- The technical architecture of O1 is built on reinforcement learning techniques inspired by OpenAI's Dota 2 and AlphaGo projects, rather than standard next-token prediction.
- The model utilizes a "chain of thought" process where the model evaluates and corrects its own reasoning steps before generating a final answer, a method trained on a proprietary dataset of tasks broken down into steps.
- O1 represents a parallel research direction to raw scaling of base LLMs, focusing on "unhiding" intelligence through iterative reasoning and self-correction.
- GigaML, a Y Combinator-backed company, applied O1 to AI customer support, demonstrating the model's impact on complex edge cases.
- GigaML pivoted from fine-tuning open-source models to solving the "squishy" problems of customer support, where rules-based systems have a 70% error rate.
- Using O1 combined with a rigorous "10,000 test case" eval strategy, GigaML reduced error rates from 70% to 5% and achieved an 85% automation rate for a client (Zepto), handling 30,000 tickets daily.
- The improvement was significant enough to replace manual work on 30,000 tickets previously managed by over 1,000 agents, addressing the high-turnover nature of rote customer support jobs.
- The conversation identifies a strategic shift in startup moats in the era of advanced reasoning models.
- The primary moat is no longer just the prompt or the model, but the ownership of proprietary data required to create a dense set of evaluation cases (evals) for specific verticals.
- Founders are advised to target the "last 10%" of accuracy where human-level precision is critical, as general base models handle the first 90%, creating high-value opportunities for teams that can integrate these tools into existing workflows.
- Strong technical teams that can build robust UI/UX, integration layers, and high-stakes accuracy layers are positioned to capture the most value, rather than commoditized "wrapper" companies.
- Future iterations of the model (O2, O3) are expected to follow O1 closely, potentially introducing "directability" features that allow users to edit specific steps in the chain of thought.
- Currently, O1's reasoning process is opaque, though it appears to hide the chain of thought from users to prevent immediate cloning of the training data.
- The ability to intervene, edit, and rerun specific reasoning steps is identified as the next major unlock for productivity.
- Y Combinator observes a trend where AI capabilities are unlocking startups in hard tech and physical science sectors (mechanical, electrical, bio-engineering) rather than just consumer applications.
- There is a belief that the rapid improvement in model capabilities, moving from two to potentially four orders of magnitude in performance, will enable solutions for complex physical world problems like fusion energy and climate change.
- The consensus among speakers is that this acceleration is the "worst" these models have ever been, with week-to-week improvements making previously impossible tasks feasible immediately.