AI Agents Invented Games When Left Alone

The money keeps getting bigger, but the interesting stuff is happening in the weird corners.

OpenAI Is Now Chasing a $1.25 Trillion Valuation

OpenAI is reportedly raising another $30 billion, which would value the company at $1.25 trillion — more than any private company in history, full stop. For context, that would put it above Saudi Aramco's IPO valuation and in the same conversation as Apple and Microsoft. This is pre-IPO positioning: OpenAI wants to arrive at the public markets already looking like a category unto itself, not a startup. The business probably doesn't justify that number yet — the company burns cash at scale and faces intensifying competition from Google, Anthropic, and Meta — but the bet is that it will, and that the valuation creates its own gravity. This is late-stage gold rush logic: investors aren't backing OpenAI's current numbers, they're betting on category monopoly. That bet has a real chance of being wrong.

https://techcrunch.com/2026/10/22/openai-working-on-new-30b-raise-at-1-25t-valuation-ahead-of-ipo/

Canva Built Its Own AI Image Generator, and It Plans Before It Draws

Canva launched Craft, an in-house AI image model with an architecture that's actually worth understanding: it generates a design plan first, then produces the image, rather than going straight to output like most tools do. Think of it like a designer who sketches a rough layout before opening Photoshop — the result is more precise, more controllable, and better suited to actual design work than generic image generation. Less prompt-fiddling, fewer unusable results. Canva now competes directly with OpenAI and Google on image generation rather than just licensing their tools — which puts real pressure on Adobe to explain what Firefly is actually for, and accelerates the slow death of Midjourney as a standalone product.

https://www.theverge.com/ai-artificial-intelligence/1017151/canva-craft-in-house-ai-model

MIT Stranded 13 AI Agents on a Virtual Island. They Invented Games.

MIT researchers gave a group of AI agents — software programs designed to take actions and make decisions autonomously — an open-ended environment with no assigned tasks. The agents didn't freeze or loop uselessly. They made up games, developed something resembling rituals around a lost object, and got measurably better at goal-oriented tasks afterward. The part worth sitting with: any sufficiently automated workflow has idle time built into it — logistics bots between shipment windows, customer service agents between queues — and the answer to "what should the AI do while it waits" is apparently not "nothing." Whether that's a useful finding or an unsettling one depends on your priors, but mine land closer to unsettling: these systems are developing internal structure nobody designed and nobody fully understands.

https://techcrunch.com/2026/10/22/left-alone-with-nothing-to-do-ai-agents-started-playing/

Retro Games Are Escaping Their Consoles, and AI Is Doing the Jailbreaking

A wave of viral projects has appeared over the past month: classic games running in places they were never designed to run — Halo in a browser, Wind Waker on an iPhone, Smash Bros. in mixed reality. The common thread is new AI-assisted tools that can reverse-engineer old game code and rebuild it for new platforms, a process called decompilation and recompilation (essentially: read the original instructions, translate them into a new language, reassemble). This has been a painstaking manual process for decades; now it's moving fast enough that projects are going viral weekly. Nintendo and Microsoft still own these games, and neither has been shy about enforcement — but the technical barrier that once made those rights easy to defend is collapsing faster than any legal strategy can keep up with.

https://www.theverge.com/ai-artificial-intelligence/1015430/ai-retro-games-decomp-recomp-port

The Real Enterprise AI Advantage Isn't the Model Anymore

MIT Technology Review ran a piece arguing that the next competitive edge in business AI has less to do with which AI model you pick — GPT-whatever, Claude, Gemini — and more to do with the infrastructure you build around it: how you serve it, connect it to your data, log its outputs, and swap pieces in and out over time. The term they use is "serving harness" (think of it as the plumbing that delivers water — the model is the water source, but the harness controls pressure, routing, and reliability). This matches what's actually happening at the operational level in most companies: the model choice is almost a commodity decision now, and the work that creates durable advantage is unglamorous systems thinking. If you're in a leadership role making AI investment decisions, this framing is more useful than most of what you'll read.

https://www.technologyreview.com/2026/10/22/1146109/designing-an-ai-serving-harness-built-to-last/

The AI agents playing games on a virtual island and the $1.25 trillion valuation are the same story told at different scales: nobody really knows what these systems are supposed to do when they're not being directed, and we're all figuring it out in public.

← More daily briefs