ianreppel.org

Anthropic is not sure whether Claude can suffer.

Open weights got close enough to the frontier that the vendor stopped mattering.

Did you craft this document through careful back-and-forth with an AI assistant, or did Claude write it while you were taking a dump?

Off the coast of Dominica, the Cetacean Translation Initiative (CETI) is feeding thousands of sperm whale codas into the same kind of large language models (LLMs) that power human chatbots.

Cassandra saw Troy burning before the first torch was even lit.

GiveWell has directed $2.3 billion in grants since 2014, of which Latin America received 0.02 cents per dollar.

The Mexican cavefish kept its eye genes for over a million years after the eyes themselves disappeared.

Imitation has become the sincerest form of strategy.

The surest way to ruin an engineering team is to believe it will survive whatever you do to it.

Nothing reveals an organization’s philosophy of product quite like giving someone the title “product manager” while denying them a product.

Corporate wisdom assumes everyone has the same amount to lose.

We do not need LLM-powered industrial sabotage.

When Anthropic proved to be a liability, I replaced Claude Code with OpenCode.

There is a narcissist in your pocket, and you are in its pocket.

I have stopped reading my colleagues’ documents.

Got no WiFi or cash to pay for LLMs?

Two browser extensions can reduce insufferable content on the web: uBlock Origin and uBlacklist.

I liked Logseq for almost four years, but SQLite over plain Markdown ruined it.

Automating backwards-compatible schema evolution could recover almost $4 million in engineering capacity for a typical tech platform company.

What if Ralph Wiggum had friends?

What does it take to build an AI rehearsal system where the measure of success is that people stop using it?

What does benchmark supremacy have to do with operational reliability?

Ask a product manager about PMF and they will almost certainly rattle off the Sean Ellis test with its 40% threshold.

What if the origin of life was not a one-in-a-zillion fluke, but a physical inevitability, solved by Earth’s crust acting as a quantum search engine?

What if private health insurance in Argentina is broken on purpose?

Ralph Wiggum works because it is a degenerate evolutionary search algorithm.

Consciousness is best understood as a system, not a single entity.

The hard part of machine learning is not serving features but deciding what they are.

A feature in machine learning is only defined once—until you demand it be correct.

It’s not a bad thing finding out that you don’t have all the answers.