Gojiberries (Page 6)

Sign in Subscribe

More issues

Using Reward Models to Select Pre-Training Data

Large code models currently treat all code as equal during pre-training, then use RLHF to fix the resulting problems. This is backwards. We should allocate pre-training compute toward good code from the start. The approach is straightforward: score code quality using verifiable signals, then weight training examples proportionally to their

No Comparative Advantage

The case for free trade is based on the principle of comparative advantage. For example, Indians “pay” for Snapchat (and, let’s say, allow it to repatriate the profits) because Americans are better at building it, and Indians are better at producing something else. The example doesn’t land. India

Diabetes as a Dummy Variable

“Adults with diabetes are nearly twice as likely to have heart disease or stroke as adults without diabetes.” NIH These alarming numbers are used to justify more pre-emptive testing, e.g., Calcium score for the heart, more aggressive use of statin therapy, etc., among diabetics. The analyses that underlie the

What do we learn from bestseller regressions?

In The Bestseller Code, Archer and Jockers learn about the attributes of bestsellers by regressing whether or not a book is a bestseller on features of the book’s text. They find that topics like “human closeness” are prognostic of a book’s success. So is the author’s ability

Reading Readers' Needs

With Gaurav Gandhi The books people pick to learn from often fail them. All too often, the book has intimidating jargon, or the examples are irrelevant to the reader, or there is too much duplication. Or the content is just wrong. Readers make poor choices about books because finding a

Building Together, Separately: Challenges of Software Development

With Atul Dhingra Microservices, Macroproblems A single page on Doordash can make upward of 1000 gRPC calls (see the interview). For many engineers, upward of a thousand network calls nicely illustrate the chaos and inefficiency unleashed by microservices. Engineers implicitly diff 1000+ gRPC calls with the orders of magnitude fewer

Unbiased Hiring

When hiring for technical positions, people often use high-precision heuristics conditional on relevance to decide who to interview. Otherwise, the interviewing burden would be too much. The advent of GPT makes use of high-precision heuristics, yet more important as interviewing and resumes are less trustworthy now. The heuristics people often

Not so Prompt: Prompt Optimization as Model Selection

Here's a framework for prompt optimization: Defining Success: Metrics and Evaluation Criteria Before collecting any data, establish what success looks like for your specific use case. Choose a primary metric that directly reflects business value—accuracy for classification, F1 for imbalanced datasets, BLEU/ROUGE for generation tasks, or

The Welfare Costs of Low-Friction Idea Production

Humans are naturally curious, and curiosity often takes us into domains where our understanding is shallow. Large language models (LLMs) have radically lowered the marginal cost of experimentation in such domains. Minutes of prompting can yield a plausible business strategy, technical architecture, or research framework—outputs that look like the