Top Stories
OpenAI Kills Sora, Scraps $1B Disney Deal Ahead of IPO
OpenAI is pulling the plug on Sora, its video generation product, and walking away from a reported $1 billion partnership with Disney. The reason? Pivoting all compute toward AGI, robotics, and a consolidated super app that merges ChatGPT, Codex, and Atlas into a single product. Sora's downloads cratered from 3.3 million to 1.1 million in three months — the market had already voted. Codex, meanwhile, quietly crossed $1 billion ARR, proving where OpenAI's real revenue engine lives.
[Source: VentureBeat](https://venturebeat.com)
Anthropic Launches Claude Computer Use on Mac
Anthropic just shipped "Computer Use" for macOS — Claude can now click, type, scroll, and navigate any application on your desktop. Paired with "Dispatch" for remote task execution, it's the most ambitious agent deployment yet. The catch? Complex workflows succeed about 50% of the time. It's powerful, imperfect, and a clear signal: the desktop is the next agent battleground.
[Source: Anthropic Blog](https://anthropic.com)
Arm Enters Chip Manufacturing with "AGI CPU"
Arm is no longer just licensing designs — it's manufacturing its own data center chip on TSMC's 3nm process. Meta, OpenAI, and Cerebras are among the first customers. This puts Arm in direct competition with AMD and Intel for the AI inference market. The silicon wars just got a third serious player.
[Source: TechCrunch](https://techcrunch.com)
Google's TurboQuant: 6-8x Memory Compression with "Zero Accuracy Loss"
Google released TurboQuant, a compression algorithm that shrinks model memory requirements by 6-8x while claiming zero measurable accuracy degradation. The open-source community ported it to MLX and llama.cpp within 24 hours. If the benchmarks hold, running 70B+ models on consumer hardware just became practical.
[Source: Google Research Blog](https://research.google)
Microsoft VibeVoice Hits 30K Stars — Open-Source Voice AI Is Here
Microsoft's VibeVoice is dominating the voice AI space with 30,000+ GitHub stars and 2,492 added just today. It handles 60-minute single-pass audio transcription with speaker diarization and is already integrated into Transformers v5.3. Mistral also dropped Voxtral-4B-TTS, claiming ElevenLabs parity. The TTS and ASR wars are officially open-source.
[Source: GitHub / HuggingFace](https://github.com)
Papers That Matter
D2Skill: Dynamic Dual-Granularity Skill Bank for Agentic RL
This paper introduces a "skill bank" approach for reinforcement learning in AI agents — letting them store, retrieve, and reuse successful action patterns across tasks. On benchmarks like ALFWorld and WebShop, it improved agent success rates by 10-20 percentage points. If you're building agents that need to learn from experience rather than just follow prompts, this is the pattern to study.
[Paper: ArXiv](https://arxiv.org)
PRCO: Perception-Reasoning Coevolution for Multimodal Models
PRCO trains multimodal models to improve perception and reasoning simultaneously rather than sequentially, boosting average accuracy by 7 points across 8 benchmarks. The key insight is that vision and language reasoning should evolve together, not in isolation. For anyone building multimodal AI products, this addresses a fundamental training bottleneck.
[Paper: ArXiv](https://arxiv.org)
How Atobotz Can Help
**That 50% success rate on Claude's Computer Use?** We've been building production agents that need to hit 95%+ reliability. The gap between a demo and a deployable agent is where we live.
**VibeVoice has 30K stars but zero enterprise integration.** We take open-source breakthroughs and turn them into voice-powered customer experiences that actually work at scale.
**TurboQuant means your models just got 6x cheaper to run.** If you're still paying peak inference costs, you're overpaying. We optimize the full stack — from model compression to production deployment.
*AI Pulse is a daily digest by [Atobotz](https://atobotz.com) — cutting through the noise so you don't have to. Subscribe for sharp takes on what actually matters in AI.*