Mikael Yemane·1d agoSeconds to Minutes Cut My Model’s Latency Effect by 99.7%Three measured leaks in scikit-learn and pandas, why the defaults still make sense, and seven rules for catching them.
Mikael Yemane·4d agoYour accuracy is flat and your probabilities are wrongAlert thresholds for calibration error and drift belong just above a noise floor set by window size, while feature staleness gets a hard…
Mikael Yemane·Nov 12Is Spark Still Relevant in 2026? DuckDB Was 7x Faster. Amazon’s Data Complicates It.Across Amazon’s Redshift sample, the 0.03% of giant queries took about 6% of query time. A head-to-head benchmark on 30 million rows, the…A response icon1A response icon1
Mikael Yemane·Nov 9Plan to serve 10 trillion tokens and Chinchilla’s 70B becomes 30BAt Chinchilla’s budget, Hoffmann’s published constants put the optimum at 59 tokens per parameter, or 93 with rounded exponents, and a…
Mikael Yemane·Nov 2A √2 in the attention scale sat silent for 47,213 stepsA stale config key ran every attention head hotter than designed from step one. Why it took that long to matter is a variance calculation.
Mikael Yemane·Oct 29Python for Machine Learning: The Honest Pros and Cons, and Why It Still WinsA 100x interpreter penalty, a global lock that only became optional in 3.14, 3 GB of CUDA wheels, and why none of it decides which language…
Mikael Yemane·Oct 26Without bias correction, Adam’s first step is 3.16× too largeThe second moment shrinks faster than the first, so an uncorrected Adam starts big. Correct m with β₂ by mistake and step one grows to…A response icon1A response icon1
Mikael Yemane·Oct 22Why MLOps Matters More Than Your ModelWhy ML systems fail without a single error in the logs, and the MLOps practices that catch it, with a 20-line demo you can run.
Mikael Yemane·Oct 19One 128K-token conversation outweighs the INT4 Llama-3.1‑70B serving itServing cost per token is set by two ceilings on the batch. The latency target picks the ceiling that binds.
Mikael Yemane·Oct 7The AI bubble debate is watching the wrong numberOn Google’s own reported numbers, revenue per API token could fall somewhere between 1.2× and 3.4× a year before depreciation outruns it.