Top Stories

The dead economy theory

1031 points · owenmcgrann.com

A long-form essay arguing that the headline economic numbers no longer reflect what most people actually experience — that GDP, employment and market indices have decoupled from the lived reality of paychecks, housing, and small business health. The piece struck a nerve on HN, where commenters traded their own anecdotes about a “vibes-cession” that refuses to show up in official statistics. Whether you buy the thesis or not, it’s a useful read for understanding the mood music behind a lot of current tech and policy debates.

SQLite is all you need for durable workflows

555 points · obeli.sk

Yet another entry in the “SQLite is shockingly enough” genre, this one making the case that you can build Temporal-style durable workflow engines on top of plain SQLite — no Postgres, no Kafka, no dedicated orchestrator. The author walks through how WAL mode, deterministic replay and a careful schema get you most of what frameworks like Temporal, Cadence and Restate provide. HN engineers loved the simplicity but pushed back on multi-writer and scaling limits, which makes the comment thread almost as useful as the post.

Notes from the Mistral AI Now Summit

386 points · koenvangilst.nl

A developer’s first-hand notes from Mistral’s flagship event, covering new model announcements, the European sovereign-AI pitch, and the agentic tooling Mistral is positioning against OpenAI and Anthropic. The write-up gets into the model lineup, pricing, and the political subtext of a French-led “alternative” stack. For anyone tracking the trio of Anthropic/OpenAI/Mistral, this is the most readable summary so far of where Mistral thinks it sits heading into the second half of 2026.

Is AI causing a repeat of frontend’s lost decade?

359 points · mastrojs.github.io

The author argues that LLM-driven coding is pushing the frontend ecosystem back into the same complexity spiral that defined the 2014–2020 era — towers of dependencies, framework churn, and apps that ship 5MB of JavaScript to render a paragraph of text. The twist: this time the bloat is being generated by AI rather than by overzealous humans. The HN thread got into whether agentic coding tools have any incentive to optimize for simplicity, or whether they’ll structurally produce overengineered code forever.

It’s hard to justify buying a Framework 12

311 points · jeffgeerling.com

Jeff Geerling’s measured take on Framework’s small-form-factor laptop, concluding that the repairability story doesn’t quite outweigh the price, weight, and performance trade-offs versus a current MacBook Air or used ThinkPad. It’s interesting both as a product review and as a referendum on whether the right-to-repair-flagship-laptop category has a viable mass market yet. HN, predictably, is split between Framework loyalists and people pointing out the still-large gap to mainstream ultraportables.

MCP is dead?

252 points · quandri.io

A provocative engineering-blog title arguing that the Model Context Protocol — Anthropic’s standard for connecting LLMs to external tools and data — is already losing momentum to alternatives baked directly into the major model providers. The author lays out where MCP got traction, where it stalled, and what the practical migration paths look like. HN was, unsurprisingly, all over this one, with the comments providing a useful reality check on a standard that’s barely a year old.

Bijou64: A variable-length integer encoding

232 points · inkandswitch.com

Ink & Switch publishes a new variable-length integer encoding aimed at local-first and CRDT-heavy workloads, claiming better small-number density than varint while staying byte-aligned and easy to skip over. The write-up walks through the format, the trade-offs versus LEB128/SQLite varints, and the real workloads that motivated it. A nice example of low-level systems work tied to a clear application — exactly the kind of post HN’s systems crowd eats up.

Liquid AI reveals 8B-A1B MoE trained on 38T tokens

192 points · liquid.ai

Liquid AI announces LFM-2.5, an 8B-active / 1B-shared mixture-of-experts model trained on a striking 38 trillion tokens, with benchmarks the team claims are competitive with much larger dense models. The release continues the trend of small-but-aggressively-trained MoEs nibbling at the frontier-model lead. HN’s interest is partly technical (the MoE routing details) and partly competitive — Liquid is one of the few non-US/non-China labs people still actively root for.