Dev Breakfast · 2026-10-09

Someone spent $24,000 worth of tokens porting tsc wholesale into Rust. The npm package is called tsc-rs, the command-line flags match tsc, and the author admits they haven't read a single line of the code. What interests me more is that the previous round burned over $400,000 and wrote 1.3 million lines of Rust only to stall at 84% compatibility — before trying to replace tsc, read the Known problems section first.

$24,000 in Tokens: An LLM Ported tsc into Rust

Someone used an LLM to port the TypeScript compiler, type checker, and LSP wholesale to Rust. The project is called ts-rust, the npm package name is tsc-rs, and the command-line arguments match tsc; for now only Linux x64 and macOS arm64 builds are provided. The author says the token bill comes to about $24,000, having burned roughly 925% to 983% of a Claude subscription's weekly quota over two weeks; an earlier attempt with a different batch of models spent more than $400,000 and 1.3 million lines of Rust without getting past 84% compatibility. The author explicitly states they haven't read a single line of the code — this is an early version. Before using it to replace tsc, read the Known problems section first, and don't stop at the “100% compatible” line.

Sources:

• GitHub - pingdotgg/ts-rust

DeepSeek 4.1 Flash: KV Cache Shrunk 437×, Under $1 per Day of Sessions

Someone spent a month writing code heavily with DeepSeek 4.1 Flash across a dozen-plus projects, and the conclusion: unless you're staring at the model name, you can't tell it from Opus — chatting, working, and speed all leave little to complain about. He doesn't care that there's no 4.1 Pro; he just uses it as a frontier model. On price, he subscribed to OpenCode Go at $10/month, and DeepSeek is effectively unlimited; the expected cost of a single session rarely exceeds $1, and sometimes a full day stays under that. The same job handed to Opus costs $1 a run; DeepSeek costs $0.003. His approach: dirty work, exploratory UI clicking-around tests, tidying desktop files — all dumped on the cheap side; only the occasional key task gets Opus 5.5 for a final code review pass to catch edge cases, and whatever it finds gets handed back to DeepSeek to fix. Calling in Opus or GLM isn't about quality either — it's about getting a second pair of eyes on the problem.

What holds up that price is the cache. DeepSeek shrank its KV cache roughly 437× versus its own V1, and keeping that cache in GPU memory was always one of the biggest costs of long sessions. He finds using Claude almost wasteful — not just in money; doing cache well means less water and electricity. He also admits he doesn't take sides on data ownership: a Chinese team mined Claude's training data, and Anthropic also took from others, but most developers don't think that way — they just want every dollar to go as far as possible.

So why didn't the industry blow up? His own explanation: these wins get ignored by the industry, because the industry only respects expensive things — if you don't spend big money it doesn't feel worth it, and FAANG just wants to buy the highest intelligence at the highest price. He also talks people out of self-hosting: on the economics of 4.1 Flash, self-hosting to save money never pays for itself; if it's privacy you're after, wait — cache optimizations like these will trickle down to local. The original post doesn't say whether the industry actually panicked later.

Sources:

• hackernews · jonotime

Opus 5.5 on One Prompt for Six Hours: 55 Invisible Cities

When we covered Haiku 5.5 a few issues ago, the anchor was the $0.01 cache read row — whether a cheap model is good enough, do the math before you pick. Today's item is the other side of the same release wave: someone gave Opus 5.5 a single prompt, asking it to use three.js to visualize all 55 cities from Calvino's Invisible Cities, explicitly writing “you have 6 hours, use them until it becomes a masterpiece,” with no follow-up questions or interruptions throughout. The same author cited another case: using Opus 5.5 for an interactive camera lab, one run of 1 hour 26 minutes, API bill $25.66.

Put these two together and the selection logic is pretty clear: Haiku 5.5 handles unit cost, Opus 5.5 handles the ceiling of capability — long-running, complex visualization, one-shot design work. It's not either/or; it's layering by task duration and complexity. The short-loop, high-frequency, bill-sensitive part stands — the math from previous issues isn't overturned. The heavy work that really needs to run for hours and hold its own without drifting is where Opus 5.5 comes in. $25.66 for 86 minutes — at that price you'd better be clear which part of the job it's doing for you.

Sources:

• I gave Opus 5.5 one prompt and six hours to visualize Invisible Cities

Whistle, a 16.9 MB Speech Model: First Token in 11 ms, on CPU

Whistle is an open-source speech recognition model that is a single 16.9 MB file, has no dependency on any library, and runs directly on CPU. It shares the same C++ engine and the same quantization as its sibling Needle, and one binary turns a chunk of audio into a tool call. It supports seven languages — English, German, French, Spanish, Italian, Dutch, and Polish — up to 30 seconds at a time, with 11 ms first-token latency, and it can also output per-word timestamps and frame-level speech embeddings. If you're building on-device apps, this means a device without a GPU can transcribe locally, with audio never leaving the device. The trade-offs are stated plainly: a 30-second cap, seven languages, an 8,192 vocabulary — long audio and Chinese are on you to solve. Anyone looking at on-device speech should read its limitations before deciding.

Sources:

• hackernews · gmays

That Star 116 Light-Years Out: 74 Analyses from Claude Code

An author said on Reddit that while using Claude Code to analyze NASA TESS telescope data, they noticed a star about 116 light-years away dimming by roughly 0.05% every 3.18 days for about two hours. The same dip lines up with independent observations in 2020 and 2018; if it really is a planet, it would be about 1.4 times Earth's size. They stress this is only a candidate planet, unconfirmed. On method, they handed downloading and parsing the data, writing the search code, fitting the transits, and ruling out false positives all to the agent, while they themselves chose the question and set the pass criteria — producing 74 analyses and over 1,000 scripts in two weeks, plus read-only cross-checks with Codex and separate sessions.

What I care more about is that line — “don't ask it whether this is a planet; keep giving it chances to prove it isn't” — that's the right posture for using an agent.

Sources:

• hackernews · pyduan

---

The author says it's an early version and hasn't read a single line of the code — would you dare wire tsc-rs into your existing build pipeline to try it, or would you honestly wait for compatibility data? See you tomorrow at 8am.

This issue picked 5 items out of 53 from the past 24 hours across X / Hacker News / GitHub Trending (sampled hourly all day, fact-checked, and then selected in the morning). Content is LLM-assisted, every item comes with an original source link, and important decisions should be cross-verified.