AI builds a world you can steer

Good morning! Yesterday’s AI releases offered three very different sizes of ambition: a world you can steer, a model small enough to specialize, and an image editor that follows finer instructions. The common thread is control. What, exactly, can you change?

Three Wednesday highlights:

Plus, a practical exercise for rehearsing business decisions. No headset required; a clear question helps considerably.

Google DeepMind announced Genie 4 on October 27. A world model generates an environment that can respond as you interact with it, rather than simply producing a finished video. Google says this version adds native sound, including ambient noise, music and dialogue.

The research preview is rolling out to trusted testers. Broader access is planned through Project Genie for Google AI Ultra subscribers and early access partners using Google’s developer products. A public announcement is therefore not the same as general availability.

For game designers, creative teams and researchers building environments for agents, the appeal is understandable: describe a scene, change something, and explore what follows. That could make early ideas easier to discuss before committing to production.

Google also lists material limitations. Details can drift, objects may not stay consistent, real locations are imperfect, and continuous interaction lasts minutes rather than extended sessions. These are important boundaries for anyone imagining a reliable training simulator.

The useful business question is narrower than whether the demo looks convincing. What would this environment help someone learn, rehearse or decide? A realistic appearance alone does not establish accurate physics or dependable outcomes.

Announcement · Availability and limitations

October 27: Liquid introduced an open-weight family covering text, vision, audio and other specialized tasks on a shared backbone, starting around the two-billion-parameter range. Models are available through Hugging Face and Liquid’s own services. For builders, this offers another candidate for focused applications. Hardware compatibility, license terms and task quality still need checking for the particular model selected. Official release.

October 27: MAI Image 2.7 launched in preview for editing, including relighting, restoration and changing text within images. Microsoft cites gains on external evaluation leaderboards; those scores do not predict every creative brief. Access starts through MAI Playground and evaluation platforms, with text-to-image generation described as coming later. Marketers should test whether unchanged product details actually remain unchanged. Official announcement.

Published October 27: Anthropic described months of research with six organizations developing medicines. Work spans several stages, beyond proposing molecules, and the company also reports internal discovery efforts. These are early research programs and laboratory findings, not proof of approved treatments or patient benefit. The operational lesson is useful elsewhere: find the bottlenecks across the whole process. Research update.

A simple whiteboard for drawing scenarios, dependencies and handoffs. Founders and operations teams can map a workflow before asking AI to improve it. Availability: a continuing free editor; paid Excalidraw+ adds cloud organization and sharing features. Draw the current process first, then mark the proposed changes. Features and plans.

A browser image editor supporting layers, masks and many file formats. Marketers can crop reference images or annotate a comparison without turning every adjustment into a new generation. Availability: the editor is free to use. Its official learning guide explains the workflow; start there before choosing extra services with similar names. Official guide.

The free whiteboard at tldraw.com gives teams a canvas for diagrams and notes. Useful when a simulation needs a clear starting state and visible branches. Availability: the hosted whiteboard is free; embedding its underlying software in a commercial product has separate licensing requirements. Keep that distinction clear when moving from an internal exercise to building an app. Official FAQ.

You can borrow the idea of simulation without access to a world model. Try this four-step exercise for a customer-support handoff, campaign review or supplier delay.

1. Define the starting state. Write the actual situation, available evidence, people involved and decision deadline. Separate known facts from assumptions. Give the exercise one owner.

2. Create three branches. Ask your assistant to explore an ordinary outcome, a plausible disruption and a case where crucial information is missing. Request a sequence of events, not invented probabilities.

3. Test the actions. For each branch, identify the next action, who can approve it, the information required and what could reverse the decision. Ask a colleague to challenge the weakest assumption.

4. Save the lesson. Choose one reversible improvement, assign responsibility and set a review date. Keep the scenarios beside the real result so the team can compare expectations with events.

For example, rehearse what happens when a supplier misses a delivery. A useful output identifies who contacts the customer and which alternatives exist. A dramatic story about the warehouse is less helpful.

Keep the exercise honest by asking what the assistant cannot know. A scenario explores a possibility; it does not forecast demand, prove a strategy or replace a conversation with the people doing the work. Invite the owner of the current process to edit the branches. Their awkward exceptions are often the most valuable inputs. Save those corrections, because the next rehearsal should begin with a better map of the actual business.

Money raised: $45 million Series A.
Valuation: undisclosed in the announcement.
Announced: October 27.

Nexus Venture Partners led, with B Capital and existing investors participating. The company-issued release puts total funding at $54 million. Hyphenate builds agents for finance operations, including cash reconciliation, invoicing and accounting workflows across existing systems.

The release reports substantial customer and revenue growth, but those are company claims rather than independently audited results supplied here. The round’s amount should also stay separate from lifetime funding and equity valuation.

Our read: investors are backing software that completes specific operational work. Buyers should ask which exceptions remain with staff and how an action’s evidence can be inspected. Company-issued announcement.

Our robot has simulated the perfect office. The virtual coffee is excellent. Unfortunately, the real invoices have not agreed to participate.

A reminder for every impressive demo: ask which part of Monday morning it actually removes.

Would you use AI to rehearse decisions or create visuals? Reply with one, and forward this to a colleague who enjoys testing ideas.