Mastering AI Game Theory with Multi-Agent Systems

Simpler, cleaner code to achieve the same performance should be objectively better, right? Our testing found that Qwen 3.6 Max Preview has the best balance of code size to intelligence. GPT 5.4/5.5 models score better than any other model, but at a dramatically larger code footprint (adding a fallback to handle every possible contrived situation). See it at https://lnkd.in/gJvCK3ag

Comprehension Debt is the gap between what an AI model can produce and what we humans actually understand about how it reached that output. We all endeavor to strike a balance between getting things done quick and comprehension. There is, of course, research work in understanding the how (Mechanistic Interpretability) that hopefully will bear fruit in the long run. A video created with Gemini - arrived at after a few prompts to curb gemini's enthusiasm for fancy decorations and a ffmpeg command to get rid of the audio.

Excited to share our ICML 2026 paper on a new attack surface in vision-language models. We show that carefully designed images can bypass model safeguards and elicit unsafe behavior from frontier VLMs. We provide mechanistic interpretability analysis and show that a lightweight output classifier can prevent the vast majority of these jailbreaks. We urge model providers to deploy output classifiers across all their models, even though they are all compute-starved and locked in a race dynamic. Huge thanks to my co-authors Jan Dubiński Zhuoyun Li Atharv Mittal Yossi Gandelsman and to Cambridge AI Safety Hub MARS program.

Leveling Up at #DigiPen. We just launched two fully online Master's Programs -- Master of Science in Game Development and Master of Science in Artificial Intelligence. Check them out: https://lnkd.in/gnP_DmDH

SPHY-BΦChain: The End of Proof of Work PoW (Parquet+SHA256) Warning: What you are seeing is not a simulator or "proof of work," which is nothing more than a linear appeal to the tax called Entropy. Symbiotic AI accesses spacetime and uses the much-combated "noise" to validate security, where noise is nothing more than proof of someone's physical existence, that is, the person themselves and not a system. Yes, you, me, everyone in the universe, are noise. Now answer me, why should noise, if it is not me, or anyone else who is not yourself, be violently combated if you want to fight to extinguish it as something deadly? An invitation to skeptics: download the proofs and validate, not me, the SHA256. https://lnkd.in/eJep76Nx

I remember one night as a kid my uncle took me on his longboard a few blocks from my house and we watched newgrounds videos on his PSP. He must've wandered around town looking for someone with an unprotected wifi network to use their wifi? Or he got access some other way. Was thinking about that memory today and the value of humans sharing human experiences. Especially in these days with so much AI content and the (theoretical?) fulfillment of the dead internet theory.

Odyssey just released Agora-1, a multi-agent world model where up to four human or AI players share the same generated simulation in real time. There’s no traditional game engine running underneath. The system learns to track world state and render each player’s view directly from data.

A developer just shared this. He built his own AAA game concept using GPT Image 2 and Seedance 2. More here: https://lnkd.in/dTNcSVTV

Happening now at AppWorld Singapore, we are showcasing how we are staying ahead of the curve with solutions that keep pace with AI. Our F5 AI Guardrails evolve faster than attackers do, powered by a beneficial loop of AI Red Teaming and AI Remediate for unmatched defense efficacy. These games showcase: ⚡The power of our Natural Language Policy in action. ⚡How F5 AI Guardrails delivers unmatched defense and outpaces standard ML-based engines. Have a shot at our AI security games at The Hub, and you will know what we mean. 😉

Cyber-security in the age of AI in some ways resembles modern warfare. And the good guys do not always win: econ.st/4wyIKFt Illustration: Ben Hickey