Hacker News

Top Story

The classic SaaS playbook of “growth first, margins later” is hitting a structural wall as expensive LLM inference costs compress AI gross margins down to an average of 52%. This fundamental shift transforms software from a near-zero marginal cost miracle into a hardware-style business model where ongoing compute costs require rigorous input cost management and usage-based pricing.

Front Page Highlights

AMD Acquires Taalas to Boost Inference by Etching Models in Silicon · The Register In a bid to challenge Nvidia’s dominance, AMD has acquired Taalas, a startup that etches deep learning model weights directly into silicon as “model-specific integrated circuits” (MSICs). While their 6nm test chip achieves a staggering 16,960 tokens per second on Llama 3.1 8B, the obvious catch is that the chip’s weights are permanently baked in—any major model update requires an expensive physical re-spin of the hardware. The HN crowd is keenly debating whether this represents the ultimate performance tier for stable production models or an impractical dead end in an industry where models change on a monthly basis.

How to Make a Nintendo 64 Game in 2026 · Phoboslab The developer behind the C port of the Impact engine details the brutal, beautiful reality of writing a native 3D first-person shooter, Xibalba 64, for original N64 hardware. The post is a masterclass in wrestling with systems constraints, including navigating the console’s tiny 4 KB texture memory, managing atrocious bus latencies, and implementing custom raycasting for fast visibility checks. The community is applauding the sheer dedication required to get a custom, homebrew 3D title running at a locked, tear-free 60 frames per second.

TIME Is Serving AI Bots a Different Website, with Ads Built In · Vincent Schmalbach A fascinating technical audit reveals that TIME.com is serving custom-tailored Markdown files to assistant bots like ClaudeBot and OAI-SearchBot while serving standard, heavy HTML to humans. Hidden inside this bot-specific Markdown are sponsored FAQs and ad-tech impression tags that humans never see, essentially treating LLM tokens as the new ad impressions. This gives the technical community a jarring first look at how publishers are actively refactoring the web to monetize the machine-to-machine economy.

The DISTINCT in Your COUNT · Boring SQL This database deep dive explains why adding a single DISTINCT keyword inside a PostgreSQL COUNT() aggregate instantly kills the parallel query planner and forces a single-core sequential sort. Because parallel aggregation relies on partial states and a combine function, counting distinct values is inherently hard to distribute since workers would have to exchange entire sets of unique values to prevent double-counting. The post provides an elegant performance workaround, showing that pushing the deduplication into a parallelizable GROUP BY subquery makes the query run over 3x faster on large datasets.

Humans missed 1 in 3 threats approving AI agent commands across 40k game runs · ScaleX Using stats gathered from a browser game simulating human-in-the-loop approvals for AI coding agents, this post highlights the systemic failure of command-approval security models. Under pressure and fatigue, developers missed a whopping 33.4% of exfiltration attacks and were particularly blind to malicious code hiding behind benign script names like npm run analyze. The comments reflect a strong consensus that treating the human as the last line of defense for a stateful agent is a fundamentally broken paradigm.

Unearthing my 1996 windowed OS in machine code for Am29000 homebrew computer · Nanochess A beautiful retrocomputing retrospective follows the author’s journey writing a custom Am29000 emulator to boot “Windows Fénix,” a cooperative multitasking windowed OS he hand-coded in machine language at age 18. After a grueling multi-day debugging marathon, he traced a fatal file-allocation crash to a single sign-extension instruction bug (EXHWS) in his emulator. HN is collectively swooning over the sheer tenacity it took to port a C-based Type 1 font rasterizer manually into Big-Endian assembly in the mid-90s.

Show HN & Launches

The community’s favorite launch is Show HN: Pokémon Emerald Ported to Raspberry Pi Pico 2, which runs a GBA decompilation natively on a $6 microcontroller. The developer bypassed emulators by compiling directly to Cortex-M33 and using the second core as a highly optimized software rasterizer to push 60 fps over HDMI. Also trending is Show HN: Wallfacer – A terminal session manager for Claude Code, and more, a SQLite-backed local TUI and CLI to name, tag, and group scattered agent sessions across Claude, Cursor, Kiro, and Codex.

Discussion & Debate

In Born Against, or why hobby programming communities are against LLM usage, the author analyzes the growing hostility toward AI assistant tools within highly technical, niche circles like OSDev, chess engines, and demoscenes. Commenters are having a highly introspective debate on craftsmanship, arguing that while LLMs serve as great force multipliers in commercial software, using them to bypass learning in a hobbyist community completely robs programmers of the hard-fought “craft” of mastering a domain.

🎧 If you’re heading out or coding, I could spin up an audio briefing of today’s digest so you can catch up on the go.