Key Takeaways

  • Nate Silver pays $100 to $200 a month for maximum subscription tiers on Claude and ChatGPT to serve as assistant coders on his quantitative models.
  • Algorithmic tasks that previously took Silver two and a half days to write now finish in 15 minutes, producing scripts that run 10 times faster.
  • Large language models fail when asked to build complex software architectures in a single prompt; their value comes from auditing existing 10,000-word programs.
  • Unchecked AI output introduces silent logic errors, including a soccer rating model where predictions for 2026 looped backwards to alter ratings from 1900.

Models That Took Days Now Run in Fifteen Minutes

When Joe Weisenthal asked whether artificial intelligence had objectively simplified forecasting workflows, Nate Silver did not hesitate. "Yeah, for sure," Silver answered. “Look, I get to hire two world-class programmers, coders, for whatever I pay for the max plans, right? 100, 200 bucks a month.” (play from 46:16)

For Silver, the payoff is immediate. Building quantitative models in previous election cycles meant grinding through tedious script writing. “And last year, when I built our model, it took me 2 and 1/2 days to code them up, and like, some of the code was buggy, we skipped a couple of tie breakers, it was very slow, right?” (play from 47:20) Silver said. “You know, and now, you have GPT or Claude do it, and they do it in in 15 minutes while you're at the gym, right? And it's code is 10 times faster.” (play from 47:30)

Silver still formulates the mathematical relationships, defines the probability distributions, and evaluates the data inputs. What the models eliminate is the friction of manual programming syntax. When an analytical script skips edge cases like tie-breaker rules or chokes on inefficient execution loops, Claude and ChatGPT clean the code and optimize runtime in minutes.

Feed Ten Thousand Words to Audit, Never One-Shot

Many founders expect an AI model to design an entire system from scratch. Silver takes the opposite approach. He insists that complex forecasting architectures cannot be generated by typing a single instruction into a chat window.

“I have these complex projects that you definitely couldn't one shot, or even come close to it,” (play from 46:31) Silver explained. “But I know this really well, and I've coded these things myself, and I know when you feed GPT or Claude a 10,000-word program and say, carefully audit this, it's quite sharp about it.” (play from 46:37)

The workflow requires human domain expertise before machine intervention. Silver writes the underlying logic himself or designs the complete structure. Only then does he pass the massive 10,000-word script to Claude or GPT. By constraining the AI to an audit role, he avoids the common trap where an assistant hallucinates arbitrary dependencies or misinterprets domain logic. The AI acts as a sharp reviewer catching syntax bugs, optimizing matrix operations, and handling tedious conditional statements.

Spotting the Bugs That Loop Backward in Time

The danger of trusting AI code lies in silent logic failures. When a human programmer writes broken code, the script usually crashes with an error trace. AI models, by contrast, frequently write clean, elegant code that executes without errors while distorting the underlying numbers.

Silver experienced this failure mode firsthand while constructing sports analytics tools. “Now, again, there are dangers,” (play from 47:47) Silver warned. “Like, they can introduce bugs and kind of hide them from you. We built a soccer model, too, and like in one version of soccer model, it had the ratings like looping back around, so like the 2026 rating would inform like the 1900 rating, and it got it it it would do things that were was quite weird.” (play from 47:49)

A script that leaks future information into historical baselines looks brilliant in backtesting while failing in production. The code does not throw an exception; it just distorts reality. If you do not understand the underlying statistical theory, you will deploy a broken model without realizing why it failed. Silver catches these loops only because he knows what the statistical outputs should look like at each stage.

What to Do With This

Take your slowest, most complex internal script, paste the entire file into Claude or ChatGPT with a strict prompt to audit performance, and ask it to optimize execution time without altering your data schemas. Before running the optimized code on live data, write unit tests checking historical boundaries to ensure future outputs do not bleed into past inputs.