The AI Dispatch — March 24, 2026

Strategic Retreat

OpenAI Kills Sora, Disney’s $1 Billion Investment Collapses

Six months after its standalone app launched, OpenAI is shutting down its AI video generation platform entirely — and the fallout has already torpedoed a landmark entertainment partnership.

Sources: Bloomberg • Variety • CNBC

OpenAI announced on March 24 that it is shutting down Sora — the iOS app, the API, and sora.com will all go dark — citing unsustainable compute costs and a strategic pivot toward enterprise products and world-simulation robotics research. The platform launched as a standalone app just six months ago in September 2025, positioning it as a creative tool for filmmakers, advertisers, and social media creators.

The immediate casualty is Disney. Variety reports that the entertainment giant’s partnership with OpenAI, which had been building toward a $1 billion equity investment, has collapsed. Disney had planned to license over 200 characters from Marvel, Pixar, and Star Wars for Sora-generated video content. The deal was never formally finalized, and with Sora’s shutdown, Disney is reportedly evaluating alternatives including Google’s Veo 3 and the open-source LTX-2.3.

The retreat is a rare public acknowledgment that not every frontier AI product can sustain its own weight. Video generation’s extreme GPU demands — industry estimates put Sora’s per-minute generation cost at 10–50x that of text — made it a money pit even by OpenAI’s standards. The company is now concentrating resources on its core ChatGPT ecosystem, the Codex developer platform, and the emerging world-simulation research that CEO Sam Altman has called “the real endgame.”

National Security

Federal Judge Signals Sympathy for Anthropic in Pentagon Blacklist Battle

At a March 24 hearing in San Francisco federal court, Judge Rita Lin pressed the Department of Defense on its decision to blacklist Anthropic’s Claude, saying the move “looks like an attempt to cripple” the company. Anthropic sued the Trump administration after being designated a supply-chain risk in early March, seeking a temporary injunction to pause the blacklist and a separate Trump directive barring federal agencies from using its technology. The conflict stems from CEO Dario Amodei’s refusal to allow Claude to be used for autonomous lethal weapons or mass domestic surveillance of Americans. Without an injunction, Anthropic says it could lose billions in business — defense contractors are currently required to certify they don’t use Claude in Pentagon work.

Sources: CNBC • Euronews • Android Headlines

Product Launch

Anthropic Launches Claude Computer Use on macOS

Anthropic released a research preview of “computer use” for Claude on March 24, available to Pro and Max subscribers on macOS. Users can assign tasks from their iPhone via the Dispatch mobile tool, and Claude will autonomously open apps, browse the web, fill spreadsheets, and interact with software interfaces. When Claude lacks a direct integration, it falls back to controlling the screen as a human would. Anthropic cautioned the feature is early-stage and that Claude will always request permission before accessing new applications.

Sources: CNBC • SiliconANGLE • MacRumors

Models & Research

Open Source

NVIDIA Open-Sources the Training Recipe Behind Nemotron-Cascade 2

NVIDIA released Nemotron-Cascade 2, a 30B-parameter MoE model that activates only 3B parameters at inference — fitting on a single GPU — while achieving gold-medal performance on the 2025 IMO, IOI, and ICPC World Finals. More significantly, NVIDIA simultaneously open-sourced the full Cascade RL post-training pipeline via the NeMo-RL repository: a reproducible sequential reinforcement learning framework covering instruction-following, STEM reasoning, tool-calling, code, and software engineering. This releases the training recipe, not just the weights — a first for a model at this performance tier.

Sources: VentureBeat • Open Source For You • MarkTechPost

Research Frontier

Meta’s HyperAgents: AI That Rewrites Its Own Learning Rules

Meta FAIR published “Hyperagents” (arXiv:2603.19461), introducing metacognitive self-modification — agents that don’t just improve at tasks, but autonomously rewrite the procedures by which they generate future improvements. Built on the Darwin Gödel Machine architecture, HyperAgents integrate a task agent and a meta-agent into a single editable program. The key finding: meta-level improvements transfer across domains. An agent that learns to improve robotics reward design can apply those optimization strategies to Olympiad-level math grading — a form of cross-domain metacognition not previously demonstrated.

Sources: arXiv • MarkTechPost

Benchmark Watch

Gemini 3.1 Pro Retakes the Crown: 77.1% on ARC-AGI-2

Google’s Gemini 3.1 Pro has posted 77.1% on ARC-AGI-2 — a benchmark specifically designed to resist memorization by testing novel logic pattern generalization. The score more than doubled its predecessor Gemini 3 Pro (31.1%) and surpassed Claude Opus 4.6 (68.8%) and GPT-5.2 (52.9%). It also leads 13 of 16 major benchmarks overall, marking Google’s return to the top of frontier model rankings for the first time in several cycles.

Meanwhile, OpenAI open-sourced GDPval, a benchmark spanning 1,320 real-world tasks across 44 occupations — from lawyers to nurses to mechanical engineers — representing $3 trillion in annual U.S. wages. Expert graders with an average of 14 years’ experience blindly ranked AI outputs against human work. The result: current frontier models now approach expert-level quality while completing tasks approximately 100x faster and 100x cheaper. The automated grading service is publicly accessible at evals.openai.com.

Sources: Google • HumAI • OpenAI GDPval • arXiv

Quick Dispatches

DeepSeek Quietly Drops V3-0324 With Improved Reasoning

DeepSeek released V3-0324 as a silent update with an improved post-training pipeline drawing from R1’s RL techniques. The model reportedly surpasses GPT-4.5 on some coding evaluations, preserves the 164K context window, and is priced at $0.20/$0.77 per million tokens. Weights are fully open on Hugging Face. Hugging Face

Luma AI Launches Uni-1: Reasoning-First Image Generation

Uni-1 is an autoregressive decoder-only transformer that interleaves text and image tokens — the same architecture as LLMs — allowing it to reason about composition before rendering pixels. It tops Google’s Nano Banana 2 and OpenAI’s GPT Image 1.5 on logic-based benchmarks at $0.09/image at 2K resolution. VentureBeat

OpenAI Foundation Pledges $1 Billion in Grants for 2026

OpenAI’s nonprofit arm formally named its leadership team and announced at least $1 billion this year for AI safety and societal-benefit programs — part of a $25 billion long-term commitment. Grant focus areas include biosecurity, life sciences, medical research, and workforce transition programs. Bloomberg

LTX-2.3: Open-Source 4K Video With Synchronized Audio

Lightricks open-sourced LTX-2.3, a 22B-parameter Diffusion Transformer that generates video and audio in a single forward pass at up to 4K/50fps. Weights are on Hugging Face under Apache 2.0, including FP8 quantized variants. A companion desktop editor runs the full model locally on consumer hardware. VP-Land

AI Drug Discovery Hits Cell: MSU’s GPS Platform Finds Cancer Candidates

Published in Cell, MSU’s GPS platform predicts how novel chemical structures alter gene expression without synthesizing compounds first. In mouse trials, GPS-identified compounds reduced tumors in hepatocellular carcinoma; a separate track found three candidates for idiopathic pulmonary fibrosis. Code and a public portal are open-sourced. Cell • EurekAlert

Google DeepMind Partners With Agile Robots on Industrial AI

DeepMind and Agile Robots SE announced a strategic research partnership to integrate Gemini Robotics foundation models with Agile’s industrial platform of 20,000+ deployed systems. The partnership targets manufacturing automation with a data flywheel: real-world operational data improves models, and better models expand capabilities. TechCrunch

GitHub Trending

Repository	Language	Stars	Description
bytedance/deer-flow	Python	43.2k	Open-source SuperAgent harness from ByteDance — researches, codes, and creates using sandboxes, memories, tools, skills, and subagents
obra/superpowers	Shell	108.4k	Composable agentic skills framework enforcing spec-driven development, TDD, and autonomous subagent execution for coding agents
NousResearch/hermes-agent	Python	12.5k	Self-improving AI agent with built-in learning loop — creates skills from experience, supports 200+ model providers and multi-platform messaging
harry0703/MoneyPrinterTurbo	Python	52.6k	Generate high-definition short videos with a single click using AI — the more capable follow-up to MoneyPrinter
jingyaogong/minimind	Python	42.5k	Train a fully functional 26M-parameter GPT-style language model from scratch on a single GPU in approximately 2 hours
ruvnet/ruflo	TypeScript	25.1k	Enterprise-grade agent orchestration platform for Claude enabling intelligent multi-agent swarms and autonomous workflows
pascalorg/editor	TypeScript	5.2k	3D building editor built with React Three Fiber and WebGPU — uses Zustand + IndexedDB for state management