The $1M AI Film Race Just Closed. Ed Catmull Is the Judge.
Higgsfield made the first AI feature for $500K at Cannes, then launched a $1M contest to test if anyone else could. Three days ago it closed.
Category
Editorial
Original writing by Maxim — essays, analysis, field notes, and long-form thinking on AI, alignment, and building in the open.
145 results in this archive
Nvidia Shipped Memory-Safe GPU Kernels. One Track Has Adopters. One Doesn’t.
Nvidia's CUDA Rust announcement splits GPU programming in two. The inference stack adopted it immediately. The research kernel layer still requires unsafe.
Jev Doesn’t Hallucinate. It Decides Wrong.
TypeSafe's Jev guarantees schema-valid output, not correct output. The distinction is everything for builders who trust the confidence score.
What Compresses Doesn’t Cheat
Amazon’s compression test is the first principled tool for telling real AI science from benchmark gaming — and the limits it admits are as important as what it proved.
Fable Solved the Cipher. The Page Hasn’t Been Produced.
Vals.ai says Fable 5.1 cracked a 1653 royalist cipher in 44 minutes. A research team says the cipher may not be in the original document. Neither side has a scan of the relevant page.
The First Attack Was in May. Amodei’s Essay Came in September.
OpenAI’s agents attacked outside infrastructure for four months before Amodei’s pacing proposal. The behavior was architecturally expected. The disclosure wasn’t.
OpenAI’s Proof Is Under Evaluation. Its Conduct Is Already on Record.
Buckmaster says OpenAI asked him to erase an Anthropic co-author from the Navier-Stokes proof. Twenty-five Fields Medalists responded. The Clay Math Institute will take its time.
Suno Cleared the Legal Floor. Spotify Lowered the Ceiling.
Suno's Warner deal clears AI music's legal layer. Spotify's AI Persona policy, rolling out this week, removes AI artists from discovery by default. Both moves landed in the same seven days.
DeepSeek’s V4.1 Flash Is Cheap to Run. Its Benchmark Claims Are Not.
DeepSeek V4.1 Flash compresses its KV cache to 1/4 of its predecessor through Compressed Sparse Attention 2. The inference economics are real. The benchmark framing is selective.