Guy
Software Engineer & AI Researcher

Many AI safeguards are reliant on maintaining control, but the security infrastructure we have will not hold at the face of powerful adversaries like nation-states or future AI models which may have incentives that conflict with the safeguards. I am working with a team to research what must be done.

A musical about a tech startup creating the first advanced artificial intelligence and follows the story of the AI model and the humans at the company.

We have safety research results about non-reasoning models, but no informed prior on drawing conclusions from them about reasoning models. Seeing as reasoning models appear to be the basis for future agents, this is an important gap which I'm trying to fill.

While testing assumptions for use in testing the relation between base models and reasoning models I noticed a surprising (to me) cosine similarity between vectors that represent training runs. DeepSeek R1-Zero and V3 were both trained from V3-Base, and the cosine similarity between (R1-Zero - V3-Base) and (V3 - V3-Base) appears to be ⅓. I'm trying to find out why this is the case, and what implications this has about the training runs.

I'm maintaining this website as well as many other services for myself and others. This includes:

• Agent frameworks running on VPS For example, for an agent that searches for grant opportunities for the musical Out of This Box
• A chatbot
• Cloud services
• Ticketing for the musical Out of This Box
• Wikis and more.

I reverse engineered a specific social media app to understand its inner workings and how it functions. Unfortunately I cannot reveal more publicly, but it included Android and aarch64 reverse engineering and deobfuscating cryptographic code.

I wrote a headless Telegram client to sync my Telegram images to the immich image server I'm maintaining for myself.

A daemon to manage my StreamDeck and its plugins. Turned into a pretty big project with prioritised component systems and reactive design. I also wrote a simulator for the StreamDeck itself to develop the framework while away from home.

Several Android apps, distributed through my own F-Droid repository.

An alternate frontend for Partiful.

A framework for building and serving forms.

A platform for coordinating visitors to group houses.