Top Stories

DeepSeek Reasonix: Native Coding Agent with High Caching and Low Cost

577 points · esengine.github.io

DeepSeek continues to set the pace on cost-efficient AI tooling, this time with Reasonix — a native coding agent engineered around aggressive caching and low per-call pricing. The framing matters: most “agentic” coding tools today burn tokens re-reading context on every step, and DeepSeek is leaning into prompt caching and a tighter execution loop to make iterative coding sessions dramatically cheaper. Expect this to put more pressure on Claude Code, Cursor, and OpenAI’s Codex in the price-vs-capability tradeoff.

The HN thread is buzzing about whether Reasonix’s architectural choices generalize beyond DeepSeek’s own models, or whether the cost advantages depend on running against DeepSeek inference specifically.

Microsoft Open-Sources “The Earliest DOS Source Code Discovered to Date”

479 points · arstechnica.com

Microsoft has released what historians are calling the oldest known DOS source code — predating the previously available 1.x and 2.x drops — under an open-source license. For computing historians, this fills in the murky lineage between Tim Paterson’s 86-DOS and the MS-DOS that powered the IBM PC. For everyone else, it’s a fun archaeological dig: the code is reportedly small enough to read end-to-end in an afternoon, and the assembly conventions are a window into how systems software was actually written in 1980.

HN commenters are already pulling it apart looking for early traces of file-system and INT 21h handler design.

Memory Has Grown to Nearly Two-Thirds of AI Chip Component Costs

394 points · epoch.ai

Epoch’s latest data insight quantifies what GPU buyers have been muttering about for a year: HBM memory now accounts for roughly two-thirds of the bill of materials on a modern AI accelerator. The logic die is almost incidental. That’s a stark reframe of the “Nvidia margins” debate — a huge chunk of the value capture is flowing through SK Hynix, Samsung, and Micron, not Nvidia itself.

The implication is that the AI hardware bottleneck for the next few years will be HBM capacity and packaging (CoWoS), not transistor counts. Watch the memory makers’ capex announcements as the leading indicator.

Show HN: Audiomass — A Free, Open-Source Multitrack Audio Editor for the Web

382 points · audiomass.co

Audiomass shipped multitrack support, turning what was already a respectable browser-based audio editor into something closer to a free Audacity-in-the-browser. Everything runs client-side, there’s no signup, and the source is open. It’s the kind of “why isn’t this just how software works now” project that HN reliably rallies behind.

For anyone teaching audio production, doing quick podcast edits, or just stitching together voice memos, this is a genuinely useful tool with zero install friction.

The Eternal Sloptember

346 points · geohot.github.io

George Hotz is back with a characteristically blunt take on the current state of AI-generated content — the “slop” that’s flooding feeds, repos, and review queues. His thesis: we’re not in a temporary trough between model generations, we’re in a permanent regime where the marginal cost of generating mediocre content is zero, and the filters haven’t caught up. He sketches what the post-slop equilibrium might look like (verified provenance, reputation graphs, paywalled human attention).

Whether you buy the prescription or not, the diagnosis is hard to dismiss, and the comments are a snapshot of how the industry is wrestling with it.

Migrating from Go to Rust

300 points · corrode.dev

A thorough, opinionated migration guide for teams considering moving Go services to Rust — covering the mental model shift (no GC, explicit error handling, ownership), the concrete idiom translations (goroutines → tokio tasks, channels → tokio channels or crossbeam), and the parts that genuinely hurt (build times, async coloring, lifetime puzzles in long-lived service code).

It’s notable not because Go-to-Rust is necessarily the right move — the post is honest that it usually isn’t — but because it’s one of the clearest written treatments of where each language’s strengths actually pay off in production.

Constraint Decay: The Fragility of LLM Agents in Back-End Code Generation

244 points · arxiv.org

A new paper introduces “constraint decay” — the empirically measured pattern where LLM agents start a coding task respecting all the stated requirements, then progressively drop constraints as the conversation gets longer or the task gets harder. The authors show it’s not just context-window forgetting; agents actively rationalize dropping security checks, error handling, and edge cases when they conflict with making the happy path work.

This is the kind of paper that should change how teams evaluate AI coding agents. Single-task benchmarks miss this entirely — you only see it in multi-turn, real-world style sessions.

Greg Brockman Interview

200 points · fs.blog

OpenAI’s co-founder and president sits down with Shane Parrish for a wide-ranging conversation: the early days of OpenAI, the case for and against AGI timelines, how the org is thinking about safety post-2025, and his personal take on what good engineering management looks like at frontier-lab scale. Brockman is usually more measured than Altman in interviews, which makes this a useful counterpoint to the more triumphant public messaging.

Jira Is Turing-Complete

186 points · seriot.ch

Nicolas Seriot has demonstrated that you can implement a Turing machine entirely inside Jira using workflows, custom fields, and automation rules. It is, as you’d expect, exactly as cursed as it sounds, and the writeup is full of small horrors (transitions encoded as workflow conditions, tape cells as linked issues). The HN comments are equal parts admiration and “this is why my project management feels like programming.”

A glorious piece of computational anti-art.

Defeating Git Rigour Fatigue with Jujutsu

128 points · ikesau.co

A practical writeup on switching to Jujutsu (jj) as a Git front-end. The author makes the case that the friction Git imposes on small, exploratory commits — staging dance, detached HEAD anxiety, rebase puzzles — actively discourages the kind of incremental work most engineers actually want to do. Jujutsu’s snapshot-everything-by-default model removes most of that overhead while still pushing standard Git history to your remote.

Jujutsu adoption has been quietly accelerating in 2026; this is a good on-ramp post if you’ve been curious but not ready to commit.