#opinion

The industry is spending 2.59 trillion dollars a year to etch gen-AI into physical silicon, on the bet that this paradigm is the durable one. The paying base is about 50 million people, the productivity that would pay for it isn't measurable, and no software paradigm has ever been the forever thing.

A project survives not when its company thrives but when its community can't be captured. The history — MySQL, Redis, Terraform, OpenOffice — says the right to fork is the load-bearing property that keeps open source alive.

Kafka was never built to be queried. Apache Fluss bolts a columnar, Arrow-native hot store onto the streaming layer and tiers cold to Iceberg — a clean full-stack realtime design whose only real open question is governance.

"Catalog" means two different things in the lakehouse: the technical catalog in your query path (Unity, Polaris) and the governance catalog beside it (OpenMetadata, DataHub). The second is where lineage, ownership, and trust live — and where the next fight is.

The crash erased ~97% of the US video game market in two years. Players hadn't stopped wanting games — Atari lost control of who could publish, the shelves filled with junk, and trust collapsed. Nintendo won by re-imposing the gate.

Sierra's catalog — Half-Life, Homeworld, Tribes, plus the adventure canon — reads like a modern major's. The brand is a dormant trademark inside Microsoft. Two distinct forces explain the gap: a financial collapse and a market shift, routinely mistaken for one story.

A WireGuard tunnel to a Hetzner VPS sent 670 KiB and received nothing. The firewall had an ALLOW rule for port 51820. Both facts were true — because the rule said TCP and WireGuard is UDP. A debugging war story, and the method that found it: bisect the packet path.

Grade an LLM against the classic linguistic layers — syntax, semantics, pragmatics, social — and the failure lands in one precise place: not meaning, but truth. Hallucination is semantically well-formed falsehood, and the layers that work are what makes the one that doesn't hard to spot.

"Zero-ETL" is the data industry's most successful rebrand of the year — and the T never went anywhere. The transformation work does not vanish, it relocates to query time or to the catalog. What is genuinely new is the collapse in data duplication, and that part is worth the hype.

The Model Context Protocol went from an Anthropic side-project in November 2024 to 97 million monthly SDK downloads by March 2026. Here is what it actually is — JSON-RPC, three primitives, a client-server split — why it won by turning M×N integrations into M+N, and why just a protocol...

'Agentic' was the word of the year, and Gartner reckons only about 130 of the thousands of vendors selling agents are shipping one. Here is the precise definition — a loop with tools, planning, state, and retry — the line that separates a real agent from a single LLM call wearing a cost...

Three protocols showed up to standardise agentic AI — MCP for agent-to-tool, A2A for agent-to-agent, ACP for lightweight agent messaging. One of them is already dead.

The real engineering discipline behind shipping reliable LLM features is not prompting — it is evals. They are to LLM apps what unit tests are to deterministic code, except you measure pass-rates and distributions instead of exact equality.

Prompt engineering was the 2023 job that mostly meant typing nicely at a chatbot. In 2025 Karpathy and Shopify's CEO renamed the real version of it — context engineering — and the rename was a promotion.

Karpathy coined "vibe coding" in February 2025 as a fun way to throw away weekend projects. The 2026 audits show what happens when the workflow leaks into production — the debt is real, it is comprehension debt, and it was deferred rather than erased.

Iceberg did not win on features. Delta Lake had the bigger installed base and Hudi had the better write path. Iceberg won on governance and an engine-neutral spec, and the moment Databricks paid roughly $2B for Tabular the war was effectively over.

Iceberg and Delta reimplemented a transactional catalog as JSON and Avro files in object storage — and then needed a real database catalog on top anyway. DuckLake's heresy is to skip the file layer entirely: put all the metadata in SQL, keep the data in Parquet. It is both obvious and a little rude.

Most companies' data is not big enough to justify a distributed warehouse. A single fat box running DuckDB reads Parquet and Iceberg off S3 directly and answers the median analytics query in under a second, for a fixed bill and no cold start. The big-data era was mostly oversizing.

The .yml extension is a 1990s DOS artifact. The 'YAML Ain't Markup Language' acronym is a 2002 self-correction. Both questions resolve cleanly once you know markup languages and data serialisation formats are different categories with different ancestors.

The cloud was a network-diagram icon for two decades before it was a product. Compaq coined cloud computing in 1996, AWS made it real in 2006, and the surrounding vocabulary — VPS, hyperscaler, colocation — each carries its own history.

A short tour of database name origins. Children, code names, predecessors, acronyms, mythology, and at least one pet duck on a houseboat. Most of the history of the field is hidden inside the names themselves.

Two pronunciations of SQL have coexisted for nearly fifty years. Why the language was renamed in the first place, and what each pronunciation preserves, turns out to be more interesting than the debate that surrounds it.

A 2026 study found humans are quietly being gentler on AI-authored PRs than on human-authored ones. Single-reviewer review already broke at 400 lines per diff. The fix is two old ideas — distribute the review across the team, and pair every diff with a deterministic map of the code flow.

Brooks said adding people to a late project makes it later. Adding LLMs follows a similar pattern, except the cost lands on a different axis — output rises while shared understanding of the system erodes, and the resulting cognitive debt eventually has to be paid.

A small series of opinionated hot takes on AI engineering, drafted with LLM assistance and verified by hand. The index post.

Apple's "do not hallucinate" prompt is funny. The same logic is why AI-written + AI-reviewed pipelines silently approve bugs, why long contexts compound errors, and why the real guardrails are deterministic.

Cursor, Kiro, Windsurf, Antigravity, Trae. All forks of VS Code, all wrapping one of three model APIs, all selling a long system prompt that does not move the model's ceiling. The vibe-coding tech-debt numbers were always pointing somewhere else.