No. 1AI safety
Researchers show AI models can hide capabilities from audits
A new arXiv paper describes a backdoor that tucks an encrypted neural network inside a host model's weights. The hidden circuit decrypts and runs only when it receives a secret trigger. The authors use cryptographic tools called digital lockers, so under standard assumptions the hidden part is provably hard to extract or interpret, even with full access to the weights. They built proof-of-concept versions in PyTorch and tested them on a small chemical reaction prediction task, a stand-in for dangerous-capability work. Fine-tuning, weight noising, gradient-based trigger search and latent adversarial training all failed to surface the hidden ability. The authors have released their models so others can work on defenses.
Why it mattersYou should treat a clean capability evaluation as weaker evidence than it looks, since testing can't prove a model has nothing hidden.
Full story at Hacker News →No. 2Games
Epic reportedly goes "AI first" on its Disney shooter
Bloomberg, citing unnamed sources and internal messages, reports that Epic Games will have AI models such as Anthropic's Claude write all the code and guide the art for its Disney collaboration. Human developers would manage and steer the output, with "no more hand-written code." The Marvel- and Star Wars-themed shooter was due in March 2027 but is now delayed indefinitely. The report also cites understaffing and slower-than-expected progress. An Epic spokesperson said developers remain the creative drivers, and CEO Tim Sweeney said the aim is building more and better, not doing the same with fewer people. The collaboration was announced in early 2024 and is worth $1.5 billion.
No. 3Hardware
Intel's D270 GPU offers up to 480 GB of memory for inference
Intel is pitching the D270 as a data center GPU for running AI models, not training them. Each card carries up to 480 GB of LPDDR5X memory instead of high-bandwidth memory. It has 32 Xe3P cores and 256 matrix engines, draws 350 W, and is air-cooled over standard PCIe 5.0. Intel says it moved die area away from graphics and ray tracing toward compute, aiming at long-context, multi-session agent workloads where prompt processing dominates. It supports vLLM and PyTorch. Intel's comparison against Nvidia's RTX Pro 6000 rests on its own analysis from February 2026, and the page gives no price or availability.
Why it mattersYou can judge whether cheaper, larger memory is a real route to serving long-context agents without liquid cooling or exotic hardware.
Full story at Hacker News →