One safety story a week, pressure-tested for what's actually happening underneath the surface.

By Lindsay Gross

On October 28, the AI security firm Irregular published its tests of GLM-5.3, an open model from the Chinese lab Zhipu AI. In simulated networks it completed 3 of 13 full cyberattack scenarios from start to finish and refused none of the offensive tasks it was handed. NIST has called it the most capable openly available model for cybersecurity work released so far, about four months behind the top US systems.

It is easy to read that as a verdict on open weights. It isn't, or not a simple one. The same openness that let GLM-5.3 go out with no working refusals is also what let Irregular test it this thoroughly, without asking anyone's permission. So let's stress-test the claim that a result like this settles the open-versus-closed debate.

Start with the case for open weights, because it is strong, and a lot of it is a safety case. Open models are the ones outside researchers can actually inspect, evaluate, and red-team. Most serious interpretability and safety work runs on them. They keep a handful of companies from being the only ones who decide what a model refuses or when it switches off. Even the US Commerce Department's 2024 report credits widely available weights for letting "neutral third-party entities ... assess systems, perform audits, and validate internal developer safety checks," and it did not recommend restricting them.

The cost sits on the other side of the ledger, and it is real. A closed model that turns out to be dangerous can be patched or switched off behind its API. An open one cannot. The same Commerce report says plainly that "the release of model weights also cannot feasibly be reversed," and that "direct model weight access can allow individuals to strip these safety features." In 2023, Palisade Research showed how cheap that stripping is: under $200 on a single GPU took Llama 2-Chat from refusing almost everything harmful to refusing about 1%, with its general scores intact. Shipping safety training with an open model helps, but it is closer to a default than a lock.

The honest way to weigh this is marginal risk: whether an open model lets someone do something they could not already do with a closed model or a search. For years that favored openness, because open models lagged. Stanford's 2025 AI Index found the gap between the top open and closed models on the Chatbot Arena leaderboard fell from 8.0% in January 2024 to 1.7% a year later. The argument for openness is not wrong. It is just doing more work than it used to, because the models it covers now sit close to the frontier.

So the fight was never really open versus closed. It is a per-release question. What can this model do that was not already out there, and how much of that are you making permanent. GLM-5.3 is worth paying attention to because it moves that answer, not because it ends the argument.

The longer version, including the full case for and against open weights and the four questions worth asking before any model ships its weights, is on Hidden Layer this week:

Join the Outliers Summit on November 5-6 in New York City, hosted by Andrew Yeung and Ankur Nagpal with AngelList. Day 1 is hands-on sessions with AI tools. Day 2 moves to candid talks and firesides for founders, builders, and creators.

Speakers include Sam Parr, Ryan Hoover, Sahil Lavingia, Andy Dunn, Nadya Okamoto, and Nat Eliason.

On Thursday, October 22, our Salt Lake City chapter co-hosted Women Shaping AI with Nirnaya Lohani of AI, Her Way. More than 50 women gathered at Adobe in Lehi to hear panelists Sunny Washington, Bree Jones, Florencia Silveira, and Heather Nemelka share their experiences and advice. Leslie Koch, Cameo Doran, and Robin Huling led breakout groups on the problems women in AI keep running into: driving implementation at work, getting visibility and credit, access, and having a say in how AI gets built.

Heather Nemelka put it plainly: “There’s a difference between having access to tools and access to opportunity.” Her concern is that women get more productive with AI without getting more visible, so companies need to give them real problems to solve and projects to lead. Florencia Silveira said of hiring decisions, “We’re looking for demonstrations of critical thinking.” Sunny Washington’s advice was to pick something you’ve always wanted to do and start experimenting on it with AI.

Here are a few moments from week two of The AI Collective Podcast. On Monday, Rissa Cao of Fish Audio walked through Drama 3, the company’s newest text-to-speech model. Say “this fried rice is a masterpiece,” change “masterpiece” to “disaster,” and the model regenerates just that word. As she put it, everything else stays the same. On Tuesday, Dan Green, who leads our Kansas City chapter, shared a hackathon tool he built that scores local ballot measures by reading level. He said the average Kansas City ballot measure reads at a “38th grade” level, so he uses AI to translate them into plain English.

On Wednesday, Nicholas Charriere, who works on AI products at Notion, talked about the AI tutor he built to brush up on math and physics. A camera watches him solve problems with pen and paper, and the tutor makes the next exercise easier or harder based on how he did. He named it after Aristotle and says he now does about an hour of math and physics before work every day. Thursday’s lineup included Yujian Tang of OSS4AI, Andrew Gamino-Cheong of Trustible, and Northwestern PhD candidate Ruiyao Xu.

The show is live every weekday at 9 AM PT. Follow The AI Collective on LinkedIn to catch the next episode live, or watch replays on LinkedIn and YouTube.

Our partners and sponsors get exclusive placements across the newsletter and access to AIC's in-person network — demo nights, dinners, hackathons, and forums across 180+ chapters.

For all inquiries, send us a note at [email protected].

The AI Collective is built by volunteers across 180+ chapters in 40 countries.

Thank you to the thousands of volunteers around the world who make this work possible. We truly could not do this without you.