- Top249
- Posts
- Dion optimizer | Claude safeguards | Lunar dust bricks | Perplexity’s Chrome bid
Dion optimizer | Claude safeguards | Lunar dust bricks | Perplexity’s Chrome bid
Learn more, scroll less. Curated AI/Tech/Business news and articles.

Welcome to the latest edition of Top249 🤩
We're here to keep you updated on AI, tech, and business news and articles, so you can save time and focus on learning and growth.
Let's get started! 😎
Important: If you're reading this in the Promotions tab, please drag it to your Inbox and click ‘Yes’ when your email provider asks to do it for future emails.
Dion is a new optimizer designed to make training large AI models more efficient by orthonormalizing only the top singular vector subspace, reducing communication and compute overhead while maintaining performance.
It leverages amortized power iteration plus QR decomposition to approximate orthonormal updates without full-scale matrix operations and is fully compatible with distributed training techniques like FSDP and tensor parallelism.
At large scales (e.g., LLaMA-3 at 405B parameters), using low-rank Dion (as little as 1/16 rank) offers massive wall-clock time speedups over Muon, and the code is open-source with a PyTorch implementation available.
Anthropic’s Safeguards team develops multi-layered protections across the model lifecycle—including policy setting, adversarial testing, and real-time enforcement—to ensure Claude remains both helpful and safe.
They employ a “Unified Harm Framework” to assess potential misuse across physical, psychological, economic, societal, and autonomy dimensions, and iterate policies using expert-driven stress tests (e.g., during the 2024 U.S. election).
Post-deployment, real-time classifiers detect harmful outputs, enabling tactics like response steering or account enforcement; ongoing monitoring aggregates behavior over time, and efforts include a bug bounty and partnerships for system improvements.
'Building cities on the Moon': Chinese scientists develop machine to turn lunar dust into bricks
(Click 👆️ to read more)
Researchers from China’s Deep Space Exploration Laboratory in Hefei built a solar-powered 3D-printing device that uses concentrated sunlight—via parabolic reflectors and fiber optics—to melt lunar soil at over 1,300 °C and produce dense, durable bricks solely from regolith.
While these bricks alone can’t hold pressure in the Moon’s environment, they are envisioned as protective outer layers for pressurized rigid or inflatable habitat modules, offering radiation and micrometeorite shielding.
The project, spanning two years, tackled challenges like variable soil composition using simulants, and prototypes (including brick samples) were sent via Tianzhou-8 in November 2024 to the Tiangong station for exposure testing in space conditions.
AI search startup Perplexity reportedly extended an unsolicited bid of $34.5 billion to acquire Google Chrome—significantly above Chrome’s own valuation and higher than Perplexity’s estimated worth (~$18 billion).
This follows a prior statement by Perplexity that it would purchase Chrome if regulators compelled Google to divest—though no such divestment has been mandated.
Perplexity confirmed media reports, and plans to invest over $3 billion into Chrome and Chromium over the next two years, backed by large investment funds; Google has not responded or expressed intent to sell.
If you find Top249 useful, consider sharing this with your friends or colleagues.
Till next time.