Logo
Bono AI
3 min de lecture Bono AI Team

News of the Day — April 9, 2026

Daily AI watch: Meta Muse Spark, Claude Mythos and Project Glasswing, NotebookLM in Gemini, GLM-5.1 open source, Perplexity agents, agentic risk standard.

News of the Day — April 9, 2026

1. Meta Launches Muse Spark, First Model from Meta Superintelligence Labs

Summary — Meta unveiled Muse Spark, the first model built by Meta Superintelligence Labs under the leadership of Alexandr Wang. The model is multimodal (voice, text, image input), supports visual chain-of-thought reasoning, tool use, and multi-agent orchestration. It is being deployed across WhatsApp, Instagram, Facebook, Messenger, and Meta’s smart glasses.

Why it matters — This marks a major strategic shift for Meta. After the disappointing reception of Llama 4, Zuckerberg restructured the AI team and brought in Alexandr Wang (formerly Scale AI) to lead a new division. Unlike Meta’s open-source tradition, Muse Spark is proprietary for now — a first. Meta hints at a possible open-source release down the line.

Suggested angle — Meta’s pivot from open AI (Llama) to proprietary AI (Muse Spark). What does this mean for the open-source ecosystem and users who relied on Llama?

Sources

2. Anthropic Withholds Claude Mythos and Launches Project Glasswing for Cybersecurity

Summary — Anthropic introduced Claude Mythos Preview, which it describes as the most powerful AI model it has ever built. In an unprecedented move, the model will not be made publicly available. It is being distributed exclusively to over 50 tech organizations through Project Glasswing, backed by more than $100 million in usage credits. The model has already identified thousands of critical vulnerabilities, including a 27-year-old flaw in OpenBSD.

Why it matters — This is the first time in nearly seven years that a leading AI lab has publicly withheld a model over safety concerns. It reignites the debate over AI lab responsibility and controlled deployment of frontier models. The cybersecurity angle is particularly striking: an LLM that discovers zero-day vulnerabilities across major operating systems and browsers.

Suggested angle — The dilemma of models that are “too powerful”: when AI finds flaws faster than humans. Implications for both defensive and offensive cybersecurity.

Sources

3. Google Integrates NotebookLM Directly Into Gemini

Summary — Google is merging NotebookLM and Gemini by launching “Notebooks” within the Gemini app. Users can now organize their conversations and files into personal knowledge bases shared between both products. Sources added in one automatically appear in the other, with access to each platform’s unique features (video overviews and infographics in NotebookLM).

Why it matters — Google is moving from a fragmented landscape of AI products to convergence. The integration creates a more cohesive AI-assisted research ecosystem. The rollout starts this week for paid subscribers (AI Ultra, Pro, Plus), with expansion to free users (up to 100 notebooks) in the coming weeks.

Suggested angle — Google’s AI tool consolidation: toward a unified research assistant. Comparison with Microsoft’s (Copilot) and Apple’s (Apple Intelligence) approaches.

Sources

4. Zhipu Releases GLM-5.1 as Open Source: #1 on SWE-Bench Pro

Summary — Zhipu AI (Z.ai) released GLM-5.1, a 744 billion parameter model (Mixture-of-Experts architecture, 40B active) under the MIT license. The model specializes in agentic coding and can work autonomously on a programming task for up to eight hours. It ranks #1 on SWE-Bench Pro with a score of 58.4, ahead of GPT-5.4, Claude Opus 4.6, and Gemini 3.1 Pro.

Why it matters — Zhipu became the first publicly traded foundation model company in January 2026 (Hong Kong IPO, ~$31B valuation). With GLM-5.1 under MIT license, the Chinese company now offers the highest-performing open-source model on code benchmarks. The race for open AI now plays out between China and the US.

Suggested angle — Zhipu’s meteoric rise and China’s growing role in open-source AI. Comparison with Gemma 4 (Google, Apache 2.0) and Qwen 3.6-Plus (Alibaba).

Sources

5. Perplexity Pivots to AI Agents and Launches the “Billion Dollar Build”

Summary — Perplexity announced the “Billion Dollar Build,” an 8-week competition where teams use Perplexity Computer to build a company with a path to $1B. Finalists can secure up to $1M in investment and $1M in credits. Meanwhile, Perplexity reported a 50% revenue increase driven by its strategic shift from conversational search to autonomous AI agents.

Why it matters — This is a strong signal that the AI industry is transitioning from search to action (agents). Perplexity Computer orchestrates up to 19 different models (OpenAI, Anthropic, Google) to execute multi-step workflows — a concrete use case for multi-model orchestration.

Suggested angle — From search to agents: how Perplexity and others are redefining AI interaction. The rise of multi-model orchestration platforms.

Sources

6. A Financial Risk Standard for Autonomous AI Agents (ARS)

Summary — Researchers from Google DeepMind, Microsoft Research, Columbia University, and T54 Labs published the Agentic Risk Standard (ARS), a framework for managing financial risk when AI agents transact autonomously. The framework borrows mechanisms from traditional finance — escrow, underwriting, and collateral. Simulations show a 24% to 61% reduction in user losses.

Why it matters — As AI agents begin managing money and assets, there was no dedicated risk standard. ARS takes a pragmatic approach: instead of trying to make models perfect, it financially hedges their mistakes — just as we do for humans in construction, insurance, and capital markets.

Suggested angle — When AI agents manage your money: the financial safeguards we need. Comparison with existing regulations (EU AI Act, national frameworks).

Sources