Ishan Anand is chief AI officer at Insight Sciences, co-founder and former chief technology officer of Layer0, and creator of Spreadsheets Are All You Need. He makes language-model mechanics accessible to nonspecialists and investigates whether AI-generated personas can reliably predict human preferences.
From internet infrastructure to interpretable AI
Anand studied mathematics and electrical engineering and computer science at MIT and worked at Digidesign/Avid before co-founding Moovweb, which became Layer0. As chief technology officer, he helped develop infrastructure for delivering fast, dynamic web applications. Limelight Networks acquired Layer0 in 2021; Anand subsequently served as vice president of product at Edgio.
His best-known educational project, Spreadsheets Are All You Need, implements GPT-2 Small entirely with Excel functions, exposing tokenization, embeddings, attention, neural-network layers, and next-token prediction to cell-by-cell inspection. He subsequently built a roughly 600-line vanilla JavaScript implementation that lets web developers load model weights locally and step through transformer inference using ordinary browser debugging tools.
- Make model behavior observable. Anand uses familiar tools to connect architectural details with practical consequences: tokenization changes what models perceive, attention supplies context, and architectural differences can affect prompting. He favors testing prompting strategies against evaluations instead of relying on intuition.
- Demonstrate activation steering directly. In his spreadsheet-based GPT-2 demonstration, Anand traces intermediate predictions through the residual stream and injects an independently identified Jedi-associated feature into model activations, changing a predicted object into a lightsaber. The example illustrates activation steering without attributing the underlying sparse-autoencoder research to him.
- Treat synthetic personas as forecasts, not respondents. At Insight Sciences, Anand develops synthetic personas for market research while emphasizing that repeated model-generated answers do not create additional human observations or statistical significance. Missing context can introduce confounding assumptions; answer order and prompt wording can distort results; and models generally predict stated attitudes more reliably than real-world behavior.
- Validate entire distributions against human evidence. His approach emphasizes calibration against human ground truth, comparisons between simulated and observed response distributions, sensitivity testing, and measurement of uncertainty in the original survey data. He also sees potential in generative agent-based modeling, where grounded personas interact within larger simulations. His spreadsheet and browser experiments embody that idea: functional software designed to make sophisticated systems understandable.
