Where to Invest $100,000 Right Now, According to Experts

Investors face a dilemma. When the S&P 500 finished its worst quarter since 2022 last month, diversifiers like bonds and bitcoin fell too.

Even with the turnaround in mid-April, analysts at Goldman Sachs and Vanguard have projected low-single-digit annualized returns from 2024-2034.

Bloomberg asked where experts would personally invest $100,000 for their March monthly edition.

One answer that surfaced for a second time? Art.

It's what billionaires like Bezos and the Rockefellers have privately used to diversify for decades.

Why?

1. Appreciation. The ArtPrice100 Index outpaced the S&P 500 overall from 2000 to 2025

2. Low-correlation. The postwar contemporary segment has moved independently of traditional investments like stocks since ‘95.*

3. Resilience. A scarce, physical, and global asset class with decades of demonstrated demand.

Thanks to the world's premier art investing platform, now anyone can invest in works featuring legends like Banksy, Basquiat, and Picasso, without needing millions.

Shares in new offerings can sell quickly but...

Our subscribers skip their waitlist*

^{*According to Masterworks data. Investing involves risk. Past performance is not indicative of future returns. See important Reg A disclosures at masterworks.com/cd.}

Today:

• Google Launches Gemma 4 12B for the Edge 

• Anthropic Previews AI Self-Improvement 

• NVIDIA Unveils Nemotron 3 Ultra for Agents 

• OpenAI Upgrades GPT-Rosalind for Science 

• Google Debuts Dreambeans AI Journal

Google's Gemma 4 12B: Bringing Multimodal Power to the Edge

Google has launched Gemma 4 12B, a new mid-sized multimodal model built to run complex, agent-driven workflows locally on consumer hardware. By ditching traditional separate encoders for a streamlined architecture, Google has packed advanced reasoning—approaching the performance of its massive 26B model—into a lightweight, edge-friendly footprint.

Key Details:

• Encoder-Free Architecture: Instead of using separate modules to translate images and audio before passing them to the language model, Gemma 4 12B feeds visual and raw audio signals directly into the LLM backbone. This drastically cuts latency and memory usage.

• Native Audio Processing: It is Google's first mid-sized model to natively understand audio, allowing for entirely offline voice transcription, formatting, and translation.

• Laptop Ready: The model requires only 16GB of VRAM or unified memory, making local development highly accessible.

• Open and Optimized: Released under an Apache 2.0 license, it is available across the developer ecosystem (including Hugging Face and Kaggle) and comes equipped with Multi-Token Prediction (MTP) drafters to speed up output generation.

AI Building AI: Anthropic's Path to Recursive Self-Improvement

Anthropic has released internal data highlighting a profound shift in software engineering: AI is accelerating the development of its own successors. While full "recursive self-improvement" (where an AI autonomously designs the next generation of models) isn't here yet, Anthropic's human engineers are increasingly delegating the heavy lifting of coding and experimentation entirely to Claude.

Key Details:

• Massive Productivity Leaps: As of May 2026, Claude authors over 80% of the code merged into Anthropic's codebase. The average engineer now ships 8x as much code per day compared to the 2021–2024 baseline.

• Autonomous Execution: Claude Opus 4.6 can now independently manage complex software tasks that take up to 12 hours. For context, in early 2024, Claude Opus 3 maxed out at tasks that took about four minutes.

• Superhuman Experimentation: When handed a fixed goal (like optimizing training code for speed), Claude achieved a 52x speedup. A skilled human researcher typically needs 4 to 8 hours just to achieve a 4x speedup.

• The Shifting Human Role: Anthropic notes that the human comparative advantage is rapidly narrowing to "research taste and judgment"—deciding which problems actually matter and setting the direction—while AI takes over the execution.

NVIDIA Nemotron 3 Ultra: The Engine for Long-Running AI Agents

As AI moves from single-turn chatbots to long-running autonomous agents, NVIDIA has unveiled Nemotron 3 Ultra to handle the orchestration. This 550-billion-parameter Mixture-of-Experts (MoE) model is specifically engineered to solve the soaring compute costs and context-drift issues that happen when agents plan, use tools, and correct errors over long periods.

Key Details:

• Optimized for Agents: Featuring 55B active parameters, it uses a "Hybrid Mamba-Transformer" design. Mamba layers handle massive context windows efficiently, while Transformer layers preserve the model's ability to accurately recall specific facts.

• Speed and Cost Efficiency: It achieves up to 5x higher throughput than comparable open models and lowers the overall cost of multi-step agentic tasks by up to 30%.

• Multi-Teacher Training: It was trained using "Multi-Teacher On-Policy Distillation" (MOPD). During training, the model learned from over ten specialized teacher models simultaneously to gain deep, domain-specific expertise.

• Hardware Agnostic: Thanks to NVFP4 precision quantization, the exact same checkpoint runs efficiently across NVIDIA Hopper, Blackwell, and Ampere GPUs.

• Open Ecosystem: Released under the permissive OpenMDW-1.1 license alongside two new guardrail and voice models (Nemotron 3.5 Content Safety and Nemotron 3.5 ASR), complete with full weights, data transparency, and training recipes.

🧠RESEARCH

Cosmos 3: Omnimodal World Models for Physical AI

NVIDIA introduced Cosmos 3, a unified AI model that processes text, images, video, audio, and physical actions all at once. Instead of using separate tools for each task, it combines everything into one system built for powering robots and smart agents with state-of-the-art, open-source technology.

World Models Meet Language Models: On the Complementarity of Concrete and Abstract Reasoning

Researchers merged text-based reasoning with visual video-prediction AI to solve complex physical problems. They trained the AI to know exactly when to rely on visual simulations versus logical rules. This method prevents the AI from being fooled by incorrect visual guesses, improving problem-solving accuracy by nearly 11%.

Language Models Need Sleep: Learning to Self-Modify and Consolidate Memories

Just like humans, AI models now benefit from "sleep." Researchers developed a process where an AI takes a break to convert temporary information into permanent knowledge. During its "dreaming" phase, the AI practices by generating its own training data, which significantly improves its ability to learn continuously over time.

📲SOCIAL MEDIA

🗞️MORE NEWS

OpenAI GPT-Rosalind Update OpenAI updated GPT-Rosalind, an AI model built specifically for life sciences research. It uses the technology behind GPT-5.5 to help scientists with complex tasks like discovering drugs, analyzing genetics, and troubleshooting lab work. The system outperformed previous models on industry tests while running faster and using less computing power.

Google Dreambeans Google launched Dreambeans, an experimental app that generates a personalized daily feed of stories, events, and ideas. By privately analyzing your connected Google apps like Gmail and Photos, it proactively suggests topics you might find interesting. Each story is uniquely illustrated using custom AI artwork and your own photo memories.

OpenAI Codex and ChatGPT Integration OpenAI is merging its coding tool, Codex, with ChatGPT to create a single platform for all types of office work. The move makes advanced tools accessible to non-technical workers by adding features like role-specific templates and the ability to instantly turn files into shareable web apps. As a result, Codex has seen rapid growth among everyday professionals doing data analysis and creative work.

ChatGPT Memory "Dreaming" OpenAI introduced "Dreaming," a background process that helps ChatGPT better remember and organize your information over time. Instead of relying on manual instructions to save facts, the system automatically reviews past conversations to keep its memory fresh and accurate. This allows the AI to better understand your preferences and seamlessly pick up where you left off on long-term projects.

Manus Shopify Connector Manus released a feature that lets users build and manage a complete Shopify store through a single chat interface. You simply describe the business you want, and the AI will design the storefront, upload products, write descriptions, and analyze sales data to create marketing campaigns. It allows beginners to launch a store from scratch while helping existing owners run their business entirely through conversation.

ElevenLabs Iconic Marketplace ElevenLabs launched a marketplace where creators can legally license the AI-generated voices of famous historical figures, actors, and characters. Users can pay to use the voices of icons like Judy Garland, John Wayne, or Transformers characters for audiobooks, video games, and ads. The platform connects creators directly with official rights holders to ensure fair compensation and proper commercial licensing.