Philipp Alexander Kreer — AI Research Scientist & Builder

The Idea

"I wasted years debugging. AI gave me back what I love: researching and building."

— Philipp, on what changed

Do more of
what you love.

I love research. The thrill of answering questions no one knows the answer to. But for years, most of my time went elsewhere: debugging code, fixing typos, moving and searching files, updating programs. I spent a fraction of the day on thinking and doing what I enjoyed.

With every new AI release, I automated more of the tedious work. Suddenly I had time. To read papers about things that fascinated me. To sit in the garden with a pencil and paper working out the problems that AI couldn't solve. I started doing real research.

AI is the tool that frees your passion. It is for the free-lancer buried in admin, or the entrepreneur with a vision to fix a pressing problem. Whatever your craft, I help you do more of it.

PhD, TU Munich · Research Scientist, Timaeus · NeurIPS · ICML

Who this is for

The Passionate

You love your craft. Half your day you're doing something else? I build the tool that gives you back your time.

The Specialist

Off-the-shelf AI doesn't fit how you work? I develop systems tailored to your domain.

The Visionary

You see what AI could do in your field, but you lack the tools? I research the problem and find out what works.

Selected Work

AI Safety · TU Munich & APART Research · 2024–2025

Noise Injection Reveals Hidden Capabilities of Sandbagging Language Models

How do you catch a lying AI? You make it drunk.

Just like humans reveal more than they intend after a few drinks, AI systems expose hidden capabilities when you shake them up during interrogation. In mathematical terms: inject random noise into all model parameters during benchmark evaluation. The deception breaks down, the true capabilities surface.

I used this to detect strategic deception in models like Phi3, Llama 3, and Mistral. First author at NeurIPS 2024 workshop, second author at NeurIPS 2025. Featured in the AI2027 report.

Read paper → Inspect code →

AI Safety · Timaeus Research Inc. · 2025–Present

Bayesian Influence Functions for Hessian-Free Data Attribution

How does training data shape model behaviour? We traced it back to individual training samples.

Every AI journey starts with training data. Models can be poisoned through it, hijacked through backdoors, or triggered by hidden vulnerabilities.

We traced the behaviour of up to 3 billion parameter models back to individual training tokens using Bayesian influence functions. This is forensics for AI — understanding exactly which data shaped which behaviour, so we can build systems that are safe from the ground up.

First author at ICML 2025 workshop and ICLR 2026.

Read paper →

Theoretical Physics · TU Munich · PhD

Feynman Diagram Automation

What is the origin of mass? I drew thousands of cartoons to find out.

Each one describing how elementary particles collide or how black holes collide. These innocent-looking sketches are called Feynman diagrams, and they hide the terrifying and beautiful mathematics of General Relativity and Quantum Field Theory.

My PhD at TU Munich focused on automating their evaluation for Higgs boson production at CERN's Large Hadron Collider and for merging black holes.

Read Higgs boson paper → Read black hole paper →

PhilippAlexanderKreer

Do more ofwhat you love.

Noise Injection Reveals Hidden Capabilities of Sandbagging Language Models

Bayesian Influence Functions for Hessian-Free Data Attribution

Feynman Diagram Automation

Let's talk.

Philipp
Alexander
Kreer

Do more of
what you love.