Blog, page 3 - Quesma

Outside of the bubble, AI is Black Mirror

A reality check on AI enthusiasm: how a simple chart generated vitriolic reactions outside the tech bubble.

Nano Banana Pro: raw intelligence with tool use

Finally, an AI that can draw a map or create an infographic. The capability of leveraging tools pushed the frontier of image generation.

A postmortem on our $2.5 million database gateway: lessons from pilot purgatory

We had a great team, $2.5 million, and a validated problem. A year later, we sold our IP for parts. Here’s what we learned about urgency and co-foundership.

HN

The security paradox of local LLMs

Local LLMs prioritize privacy over security. Our research reveals a 95% backdoor injection success rate.

AI for coding is still playing Go, not StarCraft

AI excels at clean algorithms but fails at messy, real-world codebases. The solution lies not in Go-like intelligence, but StarCraft-like complexity.

HN

CompileBench: can AI compile 22-year-old code?

We tested 19 LLMs on their ability to handle real-world software engineering tasks like compiling old code and cross-compiling. See how Anthropic, OpenAI, and Google models stack up in our new benchmark – CompileBench.

HN

Tau² benchmark: how a prompt rewrite boosted GPT-5-mini by 22%

We expected small models to be fast, but our benchmarks revealed a common reliability trap. Here’s our deep dive on finding and fixing it.

Featured

60-second Linux analysis - with Nix and LLMs

A one-line command to diagnose server health. Uses Nix to fetch tools without sudo and an LLM to summarize the output. No installation required.

Tau²: from LLM benchmark to blueprint for testing AI agents

Deep dive into the Tau² benchmark that goes beyond LLM evaluation to reveal innovative methodologies for testing AI agentic systems in realistic scenarios. Learn how this framework can transform how we test AI-powered software.

Observability in Go: what real engineers are saying in 2025

Learn how Go practitioners ship telemetry in 2025 – what works, what hurts, and the tools, workflows, and guardrails they rely on for metrics, traces, and logs.

Sandboxing AI-generated code: why we moved from WebR to AWS Lambda

Why we moved our AI chart generator from in-browser WebR (WASM) to AWS Lambda. A case study on the real-world trade-offs of running AI-generated R and ggplot2 code.