In partnership with
In Today’s Issue:
🧠 MiniMax M3 pushes open weights toward frontier agents
🖥️ NVIDIA and Microsoft bring personal agents to Windows PCs
🍎 Apple keeps pushing intelligence toward the edge
🏦 Brazil lets you put your bank account in ChatGPT
✨ And more AI goodness…
⚡ The Signal
Today's issue starts with MiniMax M3, a new open-weight model aimed straight at frontier coding and agent work. MiniMax is bundling three things that usually live behind closed APIs: strong coding benchmarks, 1M-token context through sparse attention, and native multimodality with desktop-operation ability.
The claim still needs independent pressure-testing, but the direction is clear: open-weight models are no longer competing only on chat quality. They are chasing sustained agent work, long context, tools and real developer workflows.
All the best,
Kim Isenberg
🤖 Cosmos 3 Helps Physical AI Think Ahead
NVIDIA's new Cosmos 3 world foundation model combines vision reasoning, multimodal generation and action prediction for robots, autonomous vehicles and vision AI agents. The model can generate action data such as joint angles, gripper positions and trajectory points, and NVIDIA says developers can use it through build.nvidia.com, Hugging Face, GitHub resources and NIM microservices.
👉 tl;dr: Cosmos 3 is NVIDIA's bet that physical AI needs models that understand scenes over time, predict what happens next and produce action data for real systems.
🍎 Apple Bets on Edge Intelligence
Apple's AI strategy still looks unusually restrained next to the hyperscaler buildout. 24/7 Wall St. reports Apple posted $111.184 billion in March-quarter revenue and a record $30.976 billion in Services while Microsoft is reportedly lifting 2026 CapEx to $190 billion, with AI revenue at a $37 billion annual run rate.
👉 tl;dr: Apple's edge-AI bet is about keeping more intelligence on devices and inside its own product loop, while rivals absorb the cost and friction of massive cloud infrastructure.
🖥️ RTX Spark Turns Windows PCs Into Agent Machines
NVIDIA and Microsoft are pitching RTX Spark as a new Windows PC class built for personal agents. NVIDIA says the systems offer up to 1 petaflop of AI performance, up to 128GB of unified memory, Windows security primitives, NVIDIA OpenShell, and local support for 120B-parameter LLMs with up to 1 million tokens of context. That is roughly a GeForce RTX 5070 laptop GPU sitting inside the machine.
👉 tl;dr: The PC agent story is moving from a sidebar app to the machine itself: local models, security controls, creative tools and Windows workflows in one package.
When you design an agent workflow, decide first where the work should run: on the user's device, on a local PC, in the cloud, or behind a human approval step.
Why it helps: Today's issue is full of the same pattern. NVIDIA wants agents on Windows PCs, Cosmos 3 targets physical systems, Apple is protecting the edge, and in Brazil people are piping their bank data straight into ChatGPT and Claude.
Try this: "Map this workflow into four lanes: local device, local workstation, cloud agent, and human approval. For each step, list the privacy risk, latency need, compute need, and what must be logged before execution."
[text here]
🎬 Watch This
NVIDIA's GTC Taipei 2026 keynote replay has Jensen Huang laying out the company's next AI push from Taipei Music Center: personal agents on Windows PCs, physical AI systems, robotics, embedded AI and the infrastructure behind them. The video matches the day's main thread because it shows NVIDIA treating agents as a hardware, software and deployment problem, not just a chatbot feature.
"For forty years, you launched apps. Click. Type. With RTX Spark and Microsoft Windows, you ask - and the PC does the work."
– Jensen Huang, NVIDIA founder and CEO, in NVIDIA's June 1, 2026 RTX Spark announcement. The quote captures the issue's central shift from app-driven computing toward agent-driven local workflows.
The day's dustup is CNN's new lawsuit against Perplexity. CNN filed suit in New York federal court on May 28, alleging the AI search company unlawfully copies and distributes its content, the latest publisher fight over whether AI answer engines can reuse journalism without a licensing deal.
Publishers say AI search can turn reporting into answer-box output while pulling readers away from the original source; Perplexity's side has framed these fights as another round of publishers suing new technology. The court now gets another version of the same hard question: where does search-style indexing end and content substitution begin?
MiniMax M3 Pushes Open Weights Into Frontier Agent Territory
The Takeaway
👉 MiniMax says M3 is the first open-weight model to combine frontier coding and agentic performance, 1M-token context, and native multimodality in one release.
👉 The model uses MSA, MiniMax Sparse Attention, to make long-context scaling more practical than full attention's quadratic cost curve.
👉 On MiniMax's reported coding and agent benchmarks, M3 scores 59.0% on SWE-Bench Pro, 66.0% on Terminal-Bench 2.1, 34.8% on SWE-fficiency, 28.8% on KernelBench Hard, and 74.2% on MCP Atlas.
👉 MiniMax also claims M3 can handle image and video input, operate a desktop computer, and run long autonomous work such as a nearly 12-hour paper reproduction task.
MiniMax M3 is a direct shot at the closed frontier-model bundle. The company says M3 reaches frontier-level performance on coding and agentic tasks, supports context windows up to 1M tokens through MiniMax Sparse Attention, and is natively multimodal with image and video input plus desktop-operation ability. MiniMax's bigger claim is the combination: it says M3 is currently the first open-weight model to bring those three capabilities together.
The benchmark story is aggressive. MiniMax reports 59.0% on SWE-Bench Pro, 66.0% on Terminal-Bench 2.1, 34.8% on SWE-fficiency, 28.8% on KernelBench Hard, and 74.2% on MCP Atlas. The company also says M3 surpasses GPT-5.5 and Gemini 3.1 Pro on SWE-Bench Pro while approaching Opus 4.7, beats Opus 4.7 on SVG-Bench, scores above Gemini 3.1 Pro on OmniDocBench, and posts the highest score on Claw-Eval for autonomous agents.
The more interesting part is how MiniMax frames real-world coding. The blogpost argues that single-turn coding benchmarks miss how developers actually work: requirements change, users clarify, tasks move across contexts, and sessions keep iterating. MiniMax says it built an interactive user-simulator framework to expose models to those multi-round collaboration patterns during training and evaluation.
The showcase task is deliberately long-horizon: MiniMax gave M3 the ICLR 2025 Outstanding Paper Award-winning paper Learning Dynamics of LLM Finetuning and asked it to reproduce the work. According to the company, M3 ran autonomously for nearly 12 hours, produced 18 commits and 23 experimental figures, and completed the core experiments. That is the claim to watch: open-weight competition is moving from static benchmark scores toward sustained agent work.
Why it matters: MiniMax M3 puts pressure on the assumption that frontier coding agents have to stay closed. The caveat is obvious: the strongest claims come from MiniMax's own launch post, so independent testing will decide how much of this survives contact with real developer workflows.
Sources:
🔗 https://www.minimax.io/blog/minimax-m3
Investors see ANOTHER return from Masterworks (!!!!)
That’s 6 sales in 7 months. 29 all time. And the performance?
16.5%, 17.6%, and 17.8%, net annualized returns on sold works held longer than one year (See all 29 at Masterworks.com)
It’s not from stocks, private equity, or real estate… it’s from contemporary and post war art. Crazy, right?
With Masterworks, you don’t need to be a BILLIONAIRE to invest in multi-million dollar art anymore.
Historically, the segment overall has had attractive appreciation and low correlation to stocks.*
Masterworks targets works featuring legends like Banksy, Basquiat, and Picasso, identifying what they believe to have significant long-term appreciation potential, not just at the artist level but at the level of individual artworks.
As one of the largest players in the art market, with $1.3 billion invested over 500 artworks, they pass critical advantages through to their 70,000+ members to add art to their portfolios strategically.
Looking to diversify your investments in 2026?
*According to Masterworks data. Investing involves risk. Past performance is not indicative of future returns. See important Reg A disclosures at masterworks.com/cd.
The chart: NVIDIA's Cosmos 3 blog shows Artificial Analysis leaderboards where Cosmos3-Super-Text2Image and Cosmos3-Super-Image2Video rank first among open-weights text-to-image and image-to-video models, each shown with an ELO of 1,246.
The lesson: Physical AI is pulling image, video and action generation into the same competitive lane as language models: models are being judged by whether they can make useful world data, not only fluent text.
The caveat: This is NVIDIA-selected leaderboard imagery from an NVIDIA launch post. Treat it as a useful benchmark signal, not an independent verdict on the whole physical-AI stack.
Your Bank Account Just Joined the Chat
⚡ Bottom line: A Brazilian payment institution, Cumbuca, now lets you plug your real bank statements into ChatGPT or Claude through the country's regulated Open Finance, using an MCP server.
💡 Why it matters: It answers plain questions like where your money went last month from your actual transactions, not from a guess or a generic budgeting tip.
🔎 What it means: The interface to your money is sliding from the bank app to the chat window, and whoever owns that window owns the relationship.
Here is the AI and finance story worth your time this week, and it is not another bank dashboard. A Brazilian company called Cumbuca, a payment institution licensed by the Central Bank, shipped an MCP server that connects your bank account straight to ChatGPT or Claude. You authenticate through Brazil's Open Finance with your CPF and your own bank login, then just ask: where did my money go last month, what are my five biggest merchants, build me a spreadsheet of October by category. Cumbuca says the data is read in real time and never stored on its servers, and you can revoke access whenever you want.
The reason this is more than a neat hack is who else is moving the same way. In mid-May, OpenAI put personal finance inside ChatGPT for Pro users in the US, wiring in Plaid so people can connect more than 12,000 institutions like Chase, Fidelity and Robinhood and see spending, subscriptions and upcoming payments in one view. OpenAI says more than 200 million people already ask it money questions every month. Payment Expert just asked whether Brazil, with its Pix and Open Finance head start, could become the first AI-native banking economy.
What gets me is how quietly the interface is changing. For years the bank app was the thing you opened to check your money. Now the honest answer to how am I doing might come from a chat window that read your statement two seconds ago. That is useful and a little unsettling at the same time. The open question is trust: who sees the data, where it lives, and whether revoke anytime really holds up once millions of people wire their accounts into a model.
“Who is this person again?”
You’ve had that moment. Walking into a call, scrolling through old emails, trying to remember what you promised. Lindy texts you a brief 15 minutes before: attendee context, past discussions, open items, talking points. All pulled automatically. Try Lindy free.
