Episode Summary
Executive Summary: The episode tests whether ChatGPT can help build a real battery-dispatch optimizer for wholesale power markets. Using a simplified CAISO/NP15 day-ahead battery scheduling problem, Duncan Campbell and Syed Madeni show that an AI-assisted workflow can produce working code, debug constraints, and generate outputs in hours instead of weeks—though it still needs human guidance, misses ancillary services and forecasting, and is not production-ready.
Main Topics: AI as a practical coding partner for energy problems (Priority: 5/5): The conversation frames ChatGPT not as a replacement for engineers, but as a fast collaborator that can help non-programmers and technical operators build useful software with heavy human steering. Battery dispatch optimization in wholesale markets (Priority: 5/5): Syed defines the real-world challenge: schedule a 25 MW, 4-hour battery in CAISO/NP15 to maximize revenue from energy markets, simplified from the much harder market-bidding problem. Prompting strategy: broad versus bottom-up (Priority: 4/5): Duncan compares a 'buckshot' prompt that produced a high-level outline to a narrower, stepwise approach that gradually taught the model the math, constraints, and code structure needed to make progress. Debugging and constraint definition remain human-heavy (Priority: 5/5): The model repeatedly mishandles obvious domain constraints—such as not charging and discharging simultaneously or defining a cycle correctly—showing that AI still needs domain experts for QA and correction. Limits of the demo: forecasting and ancillary services (Priority: 4/5): The prototype ends up using historical prices rather than forecasts and ignores ancillary services entirely, which Syed says are a major share of the complexity in real trading systems. Time savings and the future of energy software work (Priority: 5/5): The pair argue that AI can compress weeks of work into hours for analysts and engineers, especially in energy where many experts are strong at systems thinking but not programming. Best near-term use cases in energy (Priority: 4/5): Shayle closes by highlighting utility tariff parsing and similar document-heavy workflows as promising LLM applications because they require reading, comprehension, and structured extraction.
Key Arguments: ChatGPT works best when the user decomposes the problem into small, explicit steps rather than asking for a complete solution at once. AI can meaningfully accelerate battery optimization work, especially for analysts or operators who understand the domain but do not code. The model can generate structurally sound code, but it still misses critical domain-specific constraints unless they are explicitly stated. Forecasting market prices is a major missing piece in the demo; the prototype ultimately relies on historical prices rather than true day-ahead predictions. Ancillary services are materially harder than energy-only dispatch and were excluded, which significantly reduced the complexity of the task. The output is useful for pre-construction analysis and project evaluation, but far from ready for real-time production trading or automation. The episode suggests a broader productivity revolution in energy, where LLMs act as a force multiplier for small teams and individual contributors. A particularly strong LLM use case may be parsing complex utility tariffs and regulatory documents into structured, machine-readable data.
Data Points: Battery size: 25 MW - Syed’s simplified challenge asset: a battery to be optimized in CAISO/NP15 Battery duration: 4 hours - Defines the storage asset’s usable energy capacity for the scheduling problem Market: CAISO / NP15 trading hub - The wholesale market and simplified location used for the exercise Cycle limit: 300 annual cycles - The optimizer was constrained to a yearly cycle limit during the run Prompting time estimate: ~4 hours - Duncan’s rough estimate of the total time spent prompting, debugging, and iterating Traditional team estimate: Two engineers; best case two weeks, average one month, worst case one and a half months - Syed’s estimate for getting to a similar prototype without AI assistance Compensation reference: $20/month - Duncan referenced ChatGPT’s cost relative to hiring an engineer Efficiency reduction in problem complexity: ~90% - Syed said the challenge was simplified by removing bidding and ancillary services Complexity share of ancillary services: ~40% of the remaining complexity - Syed’s estimate of how much ancillary services add after simplifying the broader problem Interview horizon for disruption: 9 months into the ChatGPT era - Shayle notes the technology is still very early in development Market frequency note: Second-by-second / minute-block / hourly block - Syed explains ancillary services operate at higher temporal resolution than simple hourly energy optimization
Pivotal Quotes: "It felt very similar to like trading Slack messages back and forth with like an analyst on my team." — Duncan Campbell: Describing the experience of co-building the code with ChatGPT "We’re flying a plane that’s souping up the autopilot system, but I think we still need pilots." — Syed Madeni: On AI augmenting, not replacing, software/energy professionals "You want it to solve a big problem for you and you kind of have to walk it along the path to get there." — Shayle Khan: Summing up the practical prompting lesson from the experiment
Implications: LLMs are already useful for energy analytics, optimization scaffolding, and document-heavy workflows, but they remain tools for accelerated human work—not autonomous production systems. Expect major gains in productivity before fully novel capabilities.