OpenAI Kills Sora to Fast-Track 'Spud': AI's Biggest Compute Bet of 2026
OpenAI has sunset Sora, its highly publicized AI video generation model, to free up compute resources for a new flagship model codenamed Spud &emdash; which CEO Sam Altman says will be ready in weeks. The decision signals a sharp strategic pivot at one of the world's most-watched AI labs, and it arrives alongside two other significant developments: Meta's neuroscience-grounded foundation model and a notable inference speed breakthrough from AWS and Cerebras. Together, these stories illustrate how the industry is rapidly reprioritizing around compute efficiency, model capability, and real-world deployment.
OpenAI Restructures Around Spud: What We Know
According to Radical Data Science's AI News Briefs (March 25, 2026), OpenAI has formally wound down Sora to concentrate compute capacity on Spud. While Sora generated considerable attention as a text-to-video model since its launch, its discontinuation suggests that OpenAI's internal calculus now favors broad-capability models over specialized media generation.
The move is notably timed amid intensifying competition with Anthropic, whose Claude model family has gained significant enterprise traction. Altman's confidence in a weeks-away timeline for Spud suggests the model is already in late-stage evaluation. In an industry where compute allocation is strategy, the decision to sacrifice a launched product for a next-generation model demonstrates how aggressively labs are competing on the model capability frontier.
OpenAI also launched a public bug bounty program this week, expanding beyond traditional security vulnerabilities to cover AI misuse and safety risks &emdash; a sign that safety-focused infrastructure is scaling alongside model ambition.
Meta's TRIBE v2: AI Learns to Read the Brain
In one of the more significant neuroscience-AI integrations published recently, Meta introduced TRIBE v2, a foundation model trained on over 500 hours of fMRI recordings from more than 700 participants. The model is designed to predict human brain responses to visual and auditory stimuli.
What makes TRIBE v2 notably advanced is its zero-shot generalization: the model can generate accurate predictions for neuroscience-grounded foundation model and a notable inference speed breakthrough from AWS and Cerebras. Together, these stories illustrate how the industry is rapidly reprioritizing around compute efficiency, model capability, and real-world deployment.
OpenAI also launched a public bug bounty program this week, expanding beyond traditional security vulnerabilities to cover AI misuse and safety risks &emdash; a sign that safety-focused infrastructure is scaling alongside model ambition.
Meta's TRIBE v2: AI Learns to Read the Brain
In one of the more significant neuroscience-AI integrations published recently, Meta introduced TRIBE v2, a foundation model trained on over 500 hours of fMRI recordings from more than 700 participants. The model is designed to predict human brain responses to visual and auditory stimuli.
What make
Leave a Reply