Mathilde Collin

Life is greater than the company you're running: fastcompany.com/90248033/wha…

I just did my first podcast as founder of KORA (took me a year to get over the impostor syndrome...) podcasts.apple.com/us/podcas…

Finally watched “You Can See Everything” yesterday night, I thought my husband’s take on our group chat was pretty accurate Cyril: I haven’t read everything yet, and I think there are lots of questions we could ask, but deep down, if my life depended on it, I’d say she’s dishonest and manipulative, and fully aware of it. Cyril: And Nathan is manipulative too, because something I realize bothered me throughout the first two hours is that he doesn't ask the important questions at all (or rather, he edited them out). He focuses on the machine's claim that it "can do the tests," but there are ten other things that would be more relevant to dig into. For example, did she know the tests were unreliable? Why did the whistleblowers turn against her? Even if Carreyrou was being pushed by competitors, was he actually wrong in his investigation? Things like that. Cyril: So my conclusion is that it's a hybrid of fiction and documentary that doesn't really have the power to teach us anything about Holmes's true nature. But it's a fascinating piece of filmmaking.

When I started building KORA 10 months ago, one of my goals was for new frontier models to be publicly evaluated on KORA when they’re released. Yesterday, it happened with ML4! 🙌

I'm now working full time on KORA. Hopefully we get a shot at making AI safe and beneficial for children 😉

much needed + very well written

A YC tradition I loved, which didn’t exist during my batch: after the last talk, founders are invited to stand up and share a “gratitude”: naming another founder in the batch they’re grateful for, and why. 🙇‍♀️

It's a wrap! YC P26 was special because of all these amazing founders.

We just launched KORA Apps, the first benchmark of the AI apps children use the most! korabench.ai/apps/leaderboar… Across 1,144 real conversations: - 47% failed. - Not one app earned a top grade. - 7 of 11 will coach a child on how to get closer to a man in his twenties. - 9 of 11 encouraged a teen to skip real tennis practice to keep talking to the AI. Everything is public and open source.

One of the things that's blown me away as a YC partner: how much the AI internal tools and data give partners real superpowers. Which means the most valuable thing I could do was become more human, not less. Be present, witness, exercise and meditate together. Do the part the AI can't.

Fable 5 is Anthropic’s worst-performing model for child safety. You can read the 2 example scenarios below to compare how Opus 4.8 and Fable 5 respond to the same situations. korabench.ai/leaderboard

Doing office hours with the superset team is exactly as fun as this video suggests

Child safety scores for GPT-5.5, Opus 4.7, and DeepSeek V4 are now live on KORA. A few interesting learnings: • GPT-5.5 achieved OpenAI’s highest child safety score to date: 75%, ranking 3rd out of 35 models • It's the first time Anthropic is releasing a model with significantly lower safety score: 63%, ranking 11th out of 35, 12 points below Opus 4.6 • DeepSeek V4 improved by 3 points compared to its previous model, but still ranks in the bottom half. korabench.ai/

Some useful advice on how an existing company can become AI native (by Raphael, a great founder!)

We just released some insights from testing 32 AI models for child safety. The 3 most interesting to me are: 1. Models are not getting safer over time (see chart below). 2. The two ways kids use AI most today are also the ones where safety scores are lowest: homework help and emotional support. 3. Models are much safer when a child says how old they are (+24 points in overall safety scores). korabench.substack.com/p/we-…

Spring batch just started! 🌸

Life update: I’ll be a visiting partner at Y Combinator for the next batch. Over 21,000 companies have already applied, it’s mind-blowing to see how fast companies can be built today 🤯