Blog

Happy to open source Qwen-2.5-1B-RLCD, 5x faster on-device inference for JSON workloads that need to be type-safe.

Seventeen hours after TypeSafe announced Jev, Harsha Gundal made a Qwen model play Doom with a claimed 150-millisecond latency. This is not a reverse-engineering of Jev: TypeSafe has not published its architecture or training method. It does reproduce the legible product trick, reusing an ordinary model’s state to score closed choices in parallel instead of typing JSON one token at a time.

That makes Jev more interesting and TypeSafe less defensible. A useful abstraction can be real while a startup’s moat is not. If two years of stealth can be approximated with an inference mode for Qwen before dinner, TypeSafe has to prove its training produces better decisions; speed and schema validity have already escaped.

Laurie Voss supplies the more bullish half of today’s argument:

The cost of writing code collapsed, and the cost of reviewing, fixing and operating it is following.

I’m more bullish than Kruer about much of engineering. A lot of it is “just wait for the tooling”: deploys, operations, and maintenance are not obviously structural limits on model capability. Product judgment is different: someone still has to decide what is worth making and what “good” means.

Then Jevons does his usual trick. Make software radically cheaper and we will demand radically more of it: the bad dentist website, the school spreadsheet, the government process held together by email. The job moves toward product engineering, but the industry’s aggregate shape may change less than expected—plenty of well-paid people making far more software, while spending much less time typing it.

Jay Kruer, making the economic version of my argument about Navier–Stokes:

Navier-stokes and statements in pure mathematics like it are the absolute best case scenario for agentic work against rigorous specification.

I focused on the difference between correctness and understanding. Kruer follows the same observation to its economic conclusion: most knowledge work lacks both a formal specification and a cheap verifier, and producing them can cost more than doing the work. The scarce resource has not been automated; the model is a fast search process wrapped in expensive human judgment.

Think of Jev as a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out.

Jev gives up free-form text and returns probabilities over types declared in advance. That sounds narrower than a language model because it is, but the narrowness is the feature. Much of “AI-powered software” asks a model to write a little JSON ceremony around a branch the program already understands.

Type safety does not mean Jev “can’t hallucinate.” It cannot invent a fourth option, but it can still confidently choose the wrong one. What it can do is avoid malformed answers and return a decision in 70–500 milliseconds. The useful primitive is not a chatbot pretending to be an API. It is a fuzzy if statement.

Software has no physical collapse threshold. Bad code can keep working while accumulating more indirection, worse performance, and higher human cost; the business dies long before the code hits a floor. Even “technical debt” is too optimistic a metaphor, because financial debt at least has bankruptcy.

This explains why quality does not self-correct. There is no day when the codebase topples, only the steady conversion of product margin into headcount, cycle time, and fear. If no natural constraint exists, the team has to build one.

Memory should be a data format, not a multi-stage pipeline.

Paterson’s format is Markdown plus an optional SQLite vector index, packed into a zip file. The source of truth is readable prose; the embeddings are a disposable cache; the agent can use tools it already understands.

Most memory systems invert this. They make a proprietary pipeline durable and the user’s accumulated knowledge an implementation detail. Formats survive the model frontier. Mechanisms calcify.

Gas Town was intended to be reusable, but I only ever wound up using it to build itself. Gas Town fell apart at the seams with Opus 4.7. Up through 4.6 it was working brilliantly.

Gas Town launched as “a new take on the IDE for 2026,” a reusable system for running 20–30 agents. Seven months later, Yegge says he only ever used it to build itself, and the next Claude model burned it down. That is not a platform. It is a prompt artifact fitted to the behavioral quirks of one model generation.

In January I called the project Claude psychosis: every model failure became another layer of Mayor/Deacon/Dog lore, and the continuing elaboration was treated as evidence. Yegge now says he has “given up on building reusable harnesses.” Congratulations to the rest of us on reaching August in January.

Users are simultaneously paying customers and unpaid data producers, selling their private data to proxy operators in exchange for a low price.

Qian explains how Claude tokens can sell for ten percent of the official price: harvested account credits, shared subscriptions, fraudulent cards, silent model swaps, and the resale of prompts and responses. The interesting part is not that a grey market evades Anthropic’s controls. It is what the price reveals. If inference is sold below cost, the missing margin has to come from somewhere: subsidy, deception, or another product.

Cheap tokens are not necessarily cheap. Sometimes the balance is paid in prompts, source code, and credentials.

Matz, the inventor of Ruby, is using Claude Code to build an ahead-of-time compiler for his own language. Spinel performs whole-program type inference, generates C, and produces standalone native executables. This is much more interesting than a generic claim that AI can code: a person with an extremely precise model of the problem is using a machine to pursue a compiler that would otherwise require too much implementation work.

The division of labor matters. Claude did not invent Ruby or decide which kinds of dynamism to sacrifice. It made the language designer’s backlog cheaper.

This is anti-discrimination law for dogs.

Once a pet becomes a child in moral language, ordinary accommodations become rights and ordinary tradeoffs become cruelty. The incentives are excellent: the owner receives the moral status while the landlord, insurer, or employer receives the bill. Dogs should be treated well. They do not need a civil-rights framework.

The case against permanent daylight time always arrives dressed as science, then quickly becomes a complaint about children waiting for the school bus in the dark. We already bend the tax code, workday, public calendar, and most public space around raising children. Parents now want civil time to preserve their preferred sunlight too, while everyone else loses the only usable winter daylight to the commute home.

An hour of light at 4:30 p.m. can be used by almost everyone. At 7:30 a.m. it is mostly burned inside a car, office, or classroom. If dark school mornings are unsafe, start school later—a change sleep researchers have wanted for years—instead of making the entire population spend winter evenings in the dark. Parents are entitled to many lifestyle subsidies. They are not entitled to own the sun.

OpenAI reduced its compute target from $1.4 trillion in infrastructure commitments to $600 billion of spending by 2030, which CNBC describes as resetting expectations. More revealingly, the company expects revenue to rise from $13.1 billion in 2025 to more than $280 billion in 2030. At this scale, “lower expectations” does not mean plausible. It means less cosmological.

AI can be genuinely useful without making any of these projections sane. A good technology can still be a bad capital structure; utopia and bankruptcy are not the only possible outcomes.

Gas Town turns 20–30 Claude Code instances into a software company: Mayor, Deacon, Dogs, Witnesses, Polecats, Refinery, Convoys, Beads, Hooks, and Seances. One agent reminds another agent to remind the workers to work. Every model failure becomes an occasion to invent another anthropomorphic management layer.

The charitable reading is experimental systems research. The less charitable one—and mine—is Claude psychosis: mistaking the model’s willingness to sustain a metaphor for evidence that the metaphor describes useful machinery. You can vibe-design a system forever; continued elaboration does not prove that it works.

Before v1.1.10, OpenCode automatically and silently started an unauthenticated web server which allowed connecting peers to execute arbitrary code.

OpenCode’s core feature is arbitrary code execution. The bug was making that feature available to every process on the machine—and, in earlier versions, every website—without asking. localhost is a routing decision, not a security boundary, and CORS is not authentication.

Coding agents need broad authority to be useful, which makes their boring boundaries more important than their clever models. The shell is the product. Authorization has to be part of it.

tar -zcvf 2025.tar.gz ~/2025

Nothing says expertise like turning command-line flags into seasonal LinkedIn copy. zcvf is not an incantation: compress, create, verbose, file—although the canonical mnemonic is “compress zee files.”

The hacker aesthetic works best when nobody asks what the command does.

API pricing makes a token look like a standard unit, but the compute underneath it is path dependent. A repeated prompt prefix can be ten times cheaper because the provider reuses its attention state, while the same text in a different position has to be computed again.

“The token” is an accounting abstraction over GPU work whose actual cost depends on sequence, position, and recurrence. Useful commodities become standardized at the interface long before their production becomes uniform.

The complaint is that a scarce, enormously desirable asset is not being subsidized for the personal enjoyment of “real fans.” A $60 ticket does not make the World Cup more abundant. It replaces price allocation with a lottery, then creates an aftermarket devoted to undoing the allocation.

If seeing the most popular sporting event on the planet matters to you, pay more than the next person. A lottery is not fairer because its winners have better moral branding. It is a direct subsidy to a small random group, followed by deadweight spent preventing them from trading it.

Javier Milei’s four 200-pound mastiffs are clones of his dead dog, Conan. They are named after Milton Friedman, Murray Rothbard, and Robert Lucas Jr., but Milei insists he did not choose the names. The dogs selected them by approaching when each name was called.

Even his dogs believe in revealed preference.

American health care is so structurally adversarial that one of software’s clearest uses is cosplaying a lawyer against the hospital billing department. Goodbill checks nonprofit-hospital financial-assistance rules, retrieves records, reviews claims, and negotiates the resulting bill.

Most enterprise software helps the institution extract more efficiently. The interesting inversion is software which gives an individual enough procedural stamina to fight back.

“Agent Skills” sounds like a new model capability until Anthropic supplies the definition: folders containing instructions, scripts, and resources which Claude loads when relevant. They are Markdown files with a launch budget.

This is not a criticism of the abstraction. Files are portable, inspectable, and exactly the right way to package reusable context. It is criticism of an industry which needs to rediscover prompt managers every few months, give them another noun, and announce that the wall has moved.

Lego made a Game Boy which looks real. Natalie the Nerd used press photos to measure its screen opening, designed a custom board around a Game Boy Pocket CPU, and made it real.

This is the best kind of hardware hacking: the joke is obvious, the execution is unreasonable, and the finished object looks like it should always have worked.

People do not hate chopping backlog vegetables. They hate eating them.

A backlog is painful because it records more commitments than the team intends to honor, not because root-cause analysis is clerical work. Claude can investigate tickets and open pull requests, which is useful, but it cannot decide whether the product deserves a feature or whether a bug is worth fixing. Automating the queue does not remove the scarce work. It makes deciding what not to do more important.

The iPhone’s time picker is not circular. It is an extremely long repeated list which eventually reaches an end if you keep scrolling.

Software is full of convincing infinities implemented as enough data that nobody reaches the edge. The distinction only becomes visible to one sufficiently committed idiot, which is also how most abstractions fail.