The Worm That No Computer Scientist Can Crack
Plus, more links to make you a little bit smarter today.
Making a new video every week until I make $5 million - Week 15
Using Heuristics to Dominate Coding Interviews
If you’re a self-taught programmer like me, you’re probably wondering how people with Computer Science degrees can so easily come up with answers in high-stress technical interviews. I looked back at the field where I got my degree – Finance – and realized they were probably approaching technical interviews the exact same way.
The Strength to Remember and the Strength to Forget: James Baldwin on What Makes a Hero
“Let everything happen to you,” wrote Rilke, “Beauty and terror.”
The UX of LEGO Interface Panels
Piloting an ocean exploration ship or Martian research shuttle is serious business. Let's hope the control panel is up to scratch. Two studs wide and angled at 45°, the ubiquitous "2x2 decorated slope" is a LEGO minifigure's interface to the world.
The VIX as Stochastic Volatility for Corporate Bonds
Classic stochastic volatility models assume volatility is unobservable. We use the Volatility Index: S&P 500 VIX to observe it, to easier fit the model. We apply it to corporate bonds. We fit autoregression for corporate rates and for risk spreads between these rates and Treasury rates. Next, we divide residuals by VIX. Our main idea is such division makes residuals closer to the ideal case of a Gaussian white noise. This is remarkable, since these residuals and VIX come from separate market segments. Similarly, we model corporate bond returns as a linear function of rates and rate changes. Our article has two main parts: Moody's AAA and BAA spreads; Bank of America investment-grade and high-yield rates, spreads, and returns. We analyze long-term stability of these models.
The Worm That No Computer Scientist Can Crack
One of the simplest, most over-studied organisms in the world is the C. elegans nematode. For 13 years, a project called OpenWorm has tried—and utterly failed—to simulate it.
Think, Prune, Train, Improve: Scaling Reasoning without Scaling Models
Large language models (LLMs) have demonstrated strong capabilities in programming and mathematical reasoning tasks, but are constrained by limited high-quality training data. Synthetic data can be leveraged to enhance fine-tuning outcomes, but several factors influence this process, including model size, synthetic data volume, pruning strategy, and number of fine-tuning rounds. We explore these axes and investigate which conditions enable model self-improvement. We introduce the Think, Prune, Train process, a scalable framework that iteratively fine-tunes models on their own reasoning traces, using ground-truth pruning to ensure high-quality training data. This approach yields improved performance: on GSM8K, Gemma2-2B achieves a Pass@1 of 57.6% (from 41.9%), Gemma2-9B reaches 82%, matching LLaMA-3.1-70B, and LLaMA-3.1-70B attains 91%, even surpassing GPT-4o, demonstrating the effectiveness of self-generated reasoning and systematic data selection for improving LLM capabilities.