This website uses cookies
Read our Privacy policy and Terms of use for more information.
WebDevelopment
Jun 9, 2026
•
6 min read
Vite gets 100 million downloads a week and still couldn't pay for itself. That's the real story behind the Cloudflare acquisition, and it tells you which risk to actually watch.
AWS_AI_Series
+1
9 min read
Build a Python benchmark that measures the real token cost of a repeated-context workload, then apply prompt caching, cross-Region inference, and batch inference and watch the bill drop.
AWS
Jun 8, 2026
Take the smallest Nova model, teach it one job, and watch it jump from 41% to 97% accuracy for $2.18 - then serve it at the cheap-model price.
Jun 7, 2026
We'll move one prompt out of your codebase into Bedrock Prompt Management, version it, then run an LLM-as-a-judge evaluation job that scores three models against your own dataset and tells you which one to ship. About 90 minutes, most of it waiting on the eval job.
AI Agents
Jun 6, 2026
Memory benchmarks are nearly saturated on recall. A new benchmark that penalizes stale memory shows the same systems falling apart, and getting worse the longer they run.
+2
8 min read
Build a support-assistant endpoint protected by Amazon Bedrock Guardrails, then run four attacks against it and watch each one get blocked or masked. Plan for 60 to 90 minutes if you have used boto3 before.
Jun 5, 2026
Build a queryable, citation-backed knowledge base over your own documents with Bedrock Knowledge Bases and S3 Vectors, entirely from Python. 60-90 minutes, under a dollar in AWS charges.
Jun 4, 2026
Episode 1 of the AWS AI Series: a Python CLI that talks to Amazon Nova 2 and Claude Sonnet 4.6 through one identical code path, streams tokens, and returns schema-validated JSON. About an hour, end to end.
DevOps
Jun 3, 2026
7 min read
Multi-AZ is sold as protection against correlated failure. Two Azure incidents this spring show the independence assumption only ever covered one peril, and that your stateful layer never recovers at the speed of your compute.
Software Engineering
May 30, 2026
Splitting a 2,000-line diff into five 400-line PRs makes each one easier to read. It does nothing about the only resource code review actually runs out of: a reviewer who is paying attention.
May 27, 2026
SubQ shipped a frontier-class subquadratic LLM with a 12M-token window and benchmarks that beat Opus 4.6. The interesting story isn't the architecture. It's what their own numbers reveal about the gap between context size and context use.
Tutorial
May 25, 2026
We'll build a SnapStart-enabled Java Lambda with CRaC priming hooks, deploy it with AWS SAM, and benchmark the p99.9 cold start down from 6.2s to 781ms. Plan for 90 minutes if you have written a Lambda in Java before.
