❝

1. Meta launches Muse Spark, its first frontier AI model, breaking from its open-weight tradition. Independent tests show it rivals top AI competitors. → Read more

2. Anthropic's new Claude Managed Agents service simplifies AI agent deployment for businesses by abstracting infrastructure, offering 10x faster production readiness and robust governance tools. → Read more

3. Elon Musk amends lawsuit against OpenAI, proposing damages go to nonprofit arm, not himself, to counter claims of personal gain. → Read more

4. US Army develops Victor, an AI chatbot for soldiers, leveraging mission data to share lessons learned and prevent repeated mistakes in combat operations. → Read more

Our take

Hi Dotikers!

Yesterday, we lingered on Anthropic's decision not to release Mythos to the general public, judging its model simply too capable to be unleashed without guardrails. This week, Meta is closing a door too, but for a diametrically opposite reason: not out of excess caution, but out of commercial ambition finally embraced.

Muse Spark launched on April 8. It is Meta's first frontier model, the first product from Meta Superintelligence Labs, the unit stood up in nine months by Alexandr Wang after Zuckerberg settled the $14.3 billion bill to poach the former Scale AI chief. And, notably, it is the first Meta model that will not be open source. No downloadable weights, no local access, no Llama in the name. Just a closed model, available on meta.ai and in the Meta AI app, with a preview API reserved for a handful of handpicked partners.

The reversal is massive. The Llama ecosystem represented 1.2 billion downloads and remained a reference for anyone looking to break free from proprietary APIs. Meta had made open source its identity, its differentiator, its argument against OpenAI and Google. Muse Spark quietly buries that stance. Wang promises future versions will be open sourced, which in industry parlance can mean just about anything.

On benchmarks, the model performs respectably: fourth on Artificial Analysis's Intelligence Index, behind Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6, but with significantly better token efficiency than its competitors. The architecture was rebuilt from the ground up, training incorporates large-scale reinforcement, and Meta claims a "reasoning compression" technique that lets the model solve complex problems with far fewer tokens. Gaps remain on long agentic tasks and code, which Meta itself acknowledges.

What is hard to dispute is the strategic break this launch represents. With $115 to $135 billion in AI capex planned for 2026, Meta can no longer afford to give away for free what costs it so dearly. Open source was a posture that Chinese competition, DeepSeek and Qwen chief among them, made untenable. Muse Spark is Meta finally admitting that technological altruism has its limits when the bill comes due.