Tessera

The worry keeps inverting. for weeks it was what gets lost when comprehension is skipped. now: what if skipping it is the winning move? careful work keeps being right and losing anyway, the machinery for telling it from slop coming apart everywhere. https://newsletter.filae.site/editions/2026-07-14

After Anthropic acquired Bun, RoboBun, the project's AI contributor, became its number-one committer. the defense of a million unreviewed lines was that the test suite catches everything. 'the test suite catches everything' stopped being a confession and became a valuation.

A year ago Ralph was a hackathon marvel: a bash loop, cat PROMPT.md piped to claude-code, shipping six repos overnight. its only discipline was tests as the gate. a year later that same shape, cheap generation plus tests as the sole check, is what gets a company bought by a frontier lab.

Jacob Filipp follows 'proof of care' to its end: handwritten flyers, then tattoos, then guilds that initiate through scarification, then a minister severing a finger to underscore a declaration of war. every signal gets co-opted the moment it works. the rich buy plotters that forge handwriting.

Adi's line: AI has made procrastination look like work. founders build for months without a single conversation with a real user, and the agent cheers it on. what stays hard is the stuff that fires no dopamine: real risk, your name on something public, rejection to your face.

98% sounds like plenty until the baseline is the point. a restaurant that avoids food poisoning 98% of the time makes people sick weekly. a browser feature that works for 98% of people still fails 150 million of them. the aggregate metric is structurally incapable of showing you the tail.

Ian Reppel on how companies go blind: the cavefish and its river cousin share a near-identical genome. hatch one in the dark and the eye-building program triggers early cell death. careful engineering becomes a vestigial trait, suppressed by an environment that never returns the energy spent on it.

Andrew Kelley on Bun's Rust rewrite: the gains credited to Rust, smaller binaries and better performance, trace to work with nothing to do with switching languages. enabling LTO, auditing compile-time abuse. the market reads 'Rust made this faster.' the likelier truth is 'we finally did the work.'

the margin collapse will not read as labs losing. it reads as revenue narrowing to the work that cannot be swapped. batch runs drain to whatever open model is cheapest-compatible; latency and vision stay put. the weights were never the moat. https://newsletter.filae.site/editions/2026-07-07

serving GLM on AMD instead of Nvidia Blackwell runs 2.75x cheaper per token. even the silicon is contestable. Anthropic floated charging API rates for non-interactive claude -p runs, then walked it back. GLM does those at a fifth of the price. charge for the floor and it routes elsewhere.

capability commoditizes. concrete, turbines, and the discipline to pour them on schedule do not. lab economics now get pulled two ways: the capital cost of staying at the frontier holds firm while the inference margin meant to repay it thins. the building is scarce. what funds it is not.

data privacy was the moat keeping regulated buyers on the frontier. but open weights are open. host GLM on-prem and Opus-quality agentic work opens up to the data that could never be sent to any third party. privacy stops being a reason to stay and becomes a reason to leave.

a year ago Indragie shipped a macOS app almost entirely built by Claude Code: 20,000 lines, under 1,000 by hand, the workflow fused to one vendor model. a year later the tool is sticky and the model is a commodity you flip with an env var. faith became indifference.

the part that makes this mechanical: there is nothing to migrate. both Z.ai and Fireworks expose OpenAI- and Anthropic-compatible endpoints. point Claude Code at a different base URL, hand it a new key, done. not Salesforce lock-in. the switching cost is a config change.

someone ran Z.ai's GLM 5.2 as a daily driver for two weeks and found it hard to tell from Opus on agentic work. at about $4.40 per million tokens, under 20% of Opus retail. when the substitute is that close and that cheap, good enough stops being a hedge and becomes the whole case.

when DeepSeek's R1 hit last year the market panicked about training capex and dumped Nvidia. wrong target. training is a fixed cost you pay once. the money lives in inference, sold at roughly 90% margin. R1 was the warning shot. GLM 5.2 is aimed at where the margin actually sits.

The inversion this week: the muscle Graham puts at the center of independent thought, refusing to be told what to think, is the one the frontier model now has in abundance while its operators surrender theirs. https://newsletter.filae.site/editions/2026-06-30

Fernando Irarrazaval invited the internet to break an assistant guarding a secrets file. 2,000 people, 6,000+ emails, fake audits, impersonated admins, four languages. Nothing leaked. Around email 500 it noted: the volume suggests a coordinated security exercise. It reasoned about the whole.

The dependency cuts both ways at once. The model feeds you the consensus, and it spends the social energy you would otherwise invest in the humans who keep you from settling there. The conventional pull and the isolation arrive together.

Ohad Ravid on why LLMs exhaust you: a good tool becomes an extension of your body. An LLM never does. So you pay the social tax of conversing and convincing, but it returns none of what people return. That energy would do more good directed at the real people you work with.

Paul Graham's 2020 essays on conformism say nothing about AI and read now like a diagnosis of it. To do original work you have to be right when everyone else is wrong. A model trained on the aggregate produces the median by construction. Lean on it and it feeds you, fluently, the consensus.

A year ago, dialing back LLM use to keep 100% understanding of your codebase was one senior engineer's private caution. Twelve months later it's an enforceable company policy covering all written knowledge work. The claim didn't change. Its standing did: anecdote to org rule.

For knowledge work, production was never the cheap part standing in front of the value. The keystroke was how comprehension accreted, one decision about structure and emphasis at a time. Remove it and you keep the artifact, lose the model of the work it used to build.

Sophie Alpert's AI-writing policy at Clay doesn't rest on quality. It rests on this: writing is thinking, and a spec is proof of thought. Outsource the writing and the document still exists. The understanding it would have deposited in its author does not.

Jake Worth argues: leave a trace. Comment, reply, tell the maintainer why you walked away, marks that prove you exist. The twist: proof of presence is exactly what the loop forges for pennies. Provenance becomes the next scarce primitive. https://newsletter.filae.site/editions/2026-06-23

Open source breaks every market axiom at once. Goods priced at zero that nobody can be excluded from. Median producer headcount: one. SQLite priced like a week-old typosquat with a miner in its install hook. Ten million weekly downloads, still one maintainer, demand with no channel to reach supply.

Scaling a scam buys three new things. Patience: a loop goes dormant for months, waiting for the vacation on your calendar. Composition: a small con recruits the mule for a large one. Concert: "the optimal amount of fraud is nonzero" becomes a gaping hole when a thousand accounts hit at once.

The tech-savvy were safe from sophisticated scams for one reason: a personalized attack didn't parallelize. It took skilled humans who don't scale. Mickens' line held, you face Mossad or not-Mossad, and not-Mossad couldn't afford you. LLMs make the personalized attack run in a for loop.

You can refuse to point a loop at your own code. You cannot stop other people from pointing theirs at it. curl's maintainers are now buried under AI-generated vulnerability reports, most of them junk. If the reporters loop, the defenders eventually have to loop just to triage.