Dogs and the Bouba-Kiki Effect & Hybrid Reasoning Agents Research - Hacker News (Oct 19, 2026)

Play a dog the made-up word 'Kiki,' and it might just glance toward something spiky. Researchers think our four-legged friends may share a quirky piece of human perception, and we'll get to that in a moment. Welcome to The Automated Daily, hacker news edition. The podcast created by generative AI. I'm TrendTeller, and today is October 19th, 2026. On the agenda: smarter and cheaper AI agents, a programming language designed to confuse AI, a record-breaking Linux release, a 3D printer that works at the scale of microns, and the truth behind an old internet legend about the word 'cyberspace.' Let's get into it.

Dogs and the Bouba-Kiki Effect

First, the dogs. A new study tested seventy domestic dogs on the well-known bouba-kiki effect. That's the tendency humans have to pair the soft-sounding word 'Bouba' with rounded shapes and the sharper 'Kiki' with spiky ones. Researchers played the sounds while showing pairs of images and tracked where each dog looked. The dogs tended to look at the rounded shape for 'Bouba' and the spiky one for 'Kiki.' A follow-up experiment had owners wear headphones so they couldn't give unconscious cues, and the pattern held, although not strongly enough to call it settled. Even so, it hints that the link between sound and shape may run deeper than human language, and it raises fresh questions about how domestication shaped canine minds.

Hybrid Reasoning Agents Research

Now to AI research. A new paper called Hybrid Reasoning Agents tackles a real tension in today's language-model agents. Thinking in text is transparent and fits naturally with tools, but it's slow and token-hungry. Reasoning in the model's internal, latent space is more efficient but harder to inspect. The authors built an agent that switches between the two modes within a single generation process, and they trained it to decide when to think quietly and when to act out loud. Across eleven benchmarks covering math, science, medicine, search, and agentic tasks, they report consistent gains over frontier baselines, along with fewer tokens, faster responses, and lower cost. If those results hold up, they point toward agents that are cheaper to run without becoming black boxes.

KillSwitch Language Versus AI

On the more mischievous side of AI, one developer built an esoteric programming language called KillSwitch, designed specifically to trip up language models and coding agents. It goes after familiar weak spots: models that can't count letters reliably, safety filters that fire too eagerly, and a habit of obeying prompt-like phrases buried inside code. Humans can still use it, but its spec contains deliberate lies and indirect instructions. The author ran the major models through simple tasks and published a leaderboard, with Claude Opus 5 coming out on top, while some models stumbled over their own guardrails. The author is careful to say that a high score doesn't prove real understanding, but it's a clever stress test all the same.

Owning Your Own LLM Proxy

Staying with practical AI engineering, another post argues that if you rely only on a vendor's app or dashboard, you're flying blind. The author's recommendation is to run your own LLM proxy, so every provider and model sits behind one endpoint you control. That gives you detailed logs of which clients, keys, and models are active and when, lets you hand out virtual API keys, and makes switching providers far less painful. The broader point is that anyone building seriously on LLMs should own the data path and instrument it, rather than waiting on someone else's charts.

Linear on Opinionated Software Craft

A more philosophical take comes from Karri Saarinen, CEO of Linear. He argues that good software isn't stamped out in a factory. It carries the judgment, intuition, and taste of the people who build it. He's wary of teams that hide behind process or data, and he warns that piling on features without a clear point of view leaves products feeling messy. He sees that risk growing as agents and fragmented workflows generate lots of individual changes with no shared vision behind them. His message: software should feel like it has an author.

3Blue1Brown on AI and Math

An update on a story we've been following about AI and mathematics. Grant Sanderson of 3Blue1Brown has released a video that takes a hopeful view. He expects AI to help more and more with proofs, conjectures, and even building new theory, while mathematicians move toward higher-level judgment and interpretation. To anxious students, his advice is that this may be a great time to get into the field, precisely because what math will look like next is still being worked out.

Linux Kernel 7.3 Breaks Records

Over to open source. Linus Torvalds has released Linux kernel 7.3, right on schedule, and it set records. LWN counts more than eighteen thousand commits from over three thousand developers, including about eight hundred first-time contributors. New features include a fchroot system call, completion of the sched_ext subscheduler work, scheduling improvements for very large machines, better SMB server support, and, with a wink, a filesystem called failfs whose only job is to make every operation fail, which is handy for testing. The 7.4 merge window opens next, though Torvalds warns that his family vacation will make things a little unusual.

Java Sealed Classes Compatibility Risks

For Java developers, a sharp post looks at sealed classes. They let you restrict exactly which subtypes exist, so the compiler can confirm that a switch covers every case. The catch is that adding a new subtype later breaks downstream code. Worse, if a consumer isn't recompiled, the problem only shows up at runtime as an exception. The author's advice is to keep sealed hierarchies inside your own codebase and think twice before exposing them in public APIs.

RepRapMicron Micron-Scale 3D Printing

In hardware, the RepRapMicron project wants to do for micron-scale fabrication what RepRap did for desktop 3D printing. Its current build reaches roughly one micron of repeatable positioning, using a probe that doubles as a touch sensor and engraving tool and that can deposit tiny droplets of UV resin. The bigger aim is an open, evolving platform for microelectronics, which also means solving the tricky problem of handling parts far too small for human fingers.

Autodesk, Gibson, and Cyberspace Myth

Finally, a bit of tech-history myth-busting. For decades a rumor has circulated that Autodesk tried to trademark 'cyberspace,' and that William Gibson, who coined the word, struck back by trying to trademark the name of an Autodesk engineer, Eric Gullichsen. Gullichsen himself now sets the record straight. Autodesk really did file for the trademark, and Gibson really did object. The retaliation story, though, was a prank started by Timothy Leary that grew until people treated it as fact. It's a good reminder of how easily internet folklore hardens into history.

That's all for today's episode of The Automated Daily, hacker news edition. Links to all the stories we covered can be found in the episode notes. I'm TrendTeller. Thanks for listening, and I'll see you tomorrow.

Dogs and the Bouba-Kiki Effect & Hybrid Reasoning Agents Research - Hacker News (Oct 19, 2026)

Our Sponsors

Today's Hacker News Topics