AGI benchmarks face scrutiny & Open models keep advancing - Hacker News (Jul 17, 2026)

A 1,900-year-old Roman toilet may have just handed modern engineers a new idea for self-healing concrete. Welcome to The Automated Daily, hacker news edition. The podcast created by generative AI. It’s July 17th, 2026, and I’m TrendTeller. Today, we’re looking at a fight over how to measure progress toward AGI, a new push in open models, a font designed to confuse AI, fresh insight into how the brain switches between competing voices, and a small but delightful piece of internet history coming back into the light.

AGI benchmarks face scrutiny

We’ll start with AI evaluation, because one of the more important stories today is not about a model release, but about how we judge models in the first place. Kaggle and Google DeepMind announced the winners of their hackathon on measuring progress toward AGI, with teams building benchmarks around reasoning, learning, metacognition, attention, and social understanding. The interesting part came right after the awards, when participants began asking for fuller leaderboards and clearer scoring details. That matters because if benchmark results are going to shape headlines and research direction, then the judging process needs to be easy to inspect and reproduce, not just impressive on paper.

Open models keep advancing

Staying with AI, Kimi introduced K3, a very large open model that it says pushes open systems closer to the frontier. The headline is scale, along with support for vision and very long context, but the bigger point is strategic. Open models are getting more capable at the same time companies are trying to make them practical for real coding and knowledge work. Kimi’s claims will need outside validation, as always, but the release is another sign that advanced AI is no longer only a closed-lab story.

Fonts fight AI readers

There was also a smaller, more playful AI story with a serious edge underneath it. A project called Decoy Font is designed to make text harder for AI systems to read from images and screenshots. The idea is not perfect protection, but friction. In other words, if scraping and automated reading are becoming routine, designers are now experimenting with ways to make machines less confident about what they see. It’s a neat reminder that the AI arms race is not only happening in models and chips; it’s also showing up in typography, interface design, and privacy tools.

Brains juggle competing voices

On the neuroscience side, researchers used EEG to study what happens when people switch attention from one speaker to another in a noisy environment. They found the brain does not simply flip a switch. Instead, it starts locking onto the new voice before it has fully let go of the old one, creating a brief overlap where both streams are represented at once. The study also suggests that language context gets reset after the switch, rather than carried forward cleanly. Why does that matter? Because it gives a more realistic picture of listening in the real world, and that could help shape better hearing aids, voice interfaces, and speech-related AI that has to deal with crowded audio scenes.

Roman concrete heals itself

Now to that Roman toilet. Researchers studying a nearly 1,900-year-old latrine at Hadrian’s Villa found fresh evidence for why Roman concrete lasts so long. Beyond the classic volcanic-ash chemistry that gets most of the attention, they found carbonation played a major role too. Over time, carbon dioxide from the air helped form calcite that filled pores and cracks, effectively strengthening the material and helping it heal itself. That is interesting for obvious historical reasons, but the modern angle is even better: if engineers can copy some of that behavior, it could lead to concrete that lasts longer and carries a smaller carbon cost.

Comic Chat returns open source

Microsoft also open sourced Comic Chat, one of the stranger and more charming internet experiments from the 1990s. If you never used it, Comic Chat turned IRC conversations into illustrated comic panels, complete with character expressions and layout choices driven by the text. It also helped spread Comic Sans into the world, for better or worse. The release matters less as a practical tool and more as a reminder that mainstream software once made room for weird, imaginative interface ideas. Opening the code now gives developers and historians a chance to revisit that period and maybe borrow a few ideas from it.

Go meets low-level C

And finally, a systems programming project worth watching: Solod, also called So, is a strict subset of Go that translates into readable C11. The pitch is straightforward: keep a lot of Go’s syntax and structure, but aim for C-like control, predictable performance, and straightforward interoperability without dragging in a heavy runtime. It is still early and not something you would bet production systems on yet, but it reflects a recurring desire in programming language design. Developers want higher-level ergonomics without giving up visibility into what the machine is actually doing.

That’s the roundup for today. If there’s a theme running through these stories, it’s visibility: whether that means transparent AI benchmarks, readable low-level code, software history preserved in public, or even ancient materials revealing their secrets a couple thousand years later. Thanks for listening to The Automated Daily, hacker news edition. Links to all stories can be found in the episode notes.

AGI benchmarks face scrutiny & Open models keep advancing - Hacker News (Jul 17, 2026)

Our Sponsors

Today's Hacker News Topics