New ChatGPT, Claude Whiteboards, Google Carbon: What These AI Launches Mean for Builders

Keeping up with AI is starting to feel like a full-time job. Google ships a stronger model for big tasks. Claude moves into whiteboards, spreadsheets, and data analysis. Cursor brings teammates’ agents into a shared workspace. ChatGPT experiments with local business listings.

It’s a lot. But the interesting question isn’t just, “What launched?”

It’s: How can you use what launched to build a business, make money, or become meaningfully more productive?

That’s the lens I’m using. And underneath all these announcements, a few big shifts are becoming clear: AI is moving from single-player to multiplayer, services can be rebuilt around agents, and distribution matters more as building gets easier.

Table of Contents

• Google Carbon: A Better Fit for Big, Long-Running Tasks
• Claude Whiteboards, Sheets, and Data: The Service Opportunity
• Qwen 4 Max: Another Model Worth Testing, Not Another Tool You Must Adopt
• Multiplayer AI Is Becoming the New Default
• ChatGPT Business Listings Could Create a New Local Agency Category
• Model Routing and Billing Are Becoming Part of the Product Stack
• How I’d Start Becoming a Marketing Engineer
• Agents Need a Control Panel, Not Blind Trust
• Better AI Design Starts with Better References and Skills
• Ask Your Coding Agent to Make Changes Verifiable
• Software Isn’t Dead. Low-Value Software Is Under Pressure.
• Other Opportunities Worth Keeping in View
• What Makes an AI Business Durable?
• Don’t Let AI Overwhelm Turn into Inaction

Google Carbon: A Better Fit for Big, Long-Running Tasks

Full disclosure: I was an early tester of Google’s Gemini 4 Carbon, with about eight days of hands-on use before discussing the launch.

My biggest takeaway is that it’s particularly good at large, involved tasks. Think full codebase refactors, complicated applications, substantial research, or security reviews that would normally require a lot of supervision.

The promise here is not simply a faster answer. It’s a model that can work through something substantial and check its own work along the way.

One launch example involved reverse-engineering the original Theme Park game binaries, adding a DeepMind-themed ride, and bringing the result to the web. That’s a much more demanding assignment than generating a small component or answering an isolated coding question.

There’s an important distinction in my own experience, though:

• Argon has been especially useful for writing. It’s an excellent writer, and I haven’t found Carbon to be clearly better at that.
• Carbon has stood out on big tasks. The improvement is more noticeable when the assignment involves a large application or a lot of connected work.

This is why I don’t think about models as one universal leaderboard. A model can be the better choice for a complicated refactor without being my preferred writing partner.

If you have a substantial task you’ve been postponing because it would require too much babysitting, that’s the kind of assignment I’d use to test Carbon.

Claude Whiteboards, Sheets, and Data: The Service Opportunity

Claude’s whiteboards, sheets, and data tools are interesting because they move AI into the places where actual work happens.

• Whiteboards: A collaborative planning surface that feels closer to FigJam than a traditional chat window.
• Sheets: A spreadsheet environment positioned against familiar Excel-style workflows.
• Data: An analyst-like experience for working with information and turning it into explanations and presentations.

The data integrations are a big part of the story. The demonstration included connections to tools such as Airtable, Databricks, and Snowflake. That moves the conversation beyond asking a chatbot to interpret a pasted spreadsheet. It points toward working with the systems where a business already keeps its information.

I see this as competitive pressure on established productivity software. But as a founder, I’m more interested in the businesses you can build around it.

Start with a productized data service

You could build a service around whiteboards, sheets, or data. Of those, I’d probably explore data first.

Businesses already pay people to analyze information, explain what’s happening, and help them make decisions. If an AI-powered workflow can handle a meaningful portion of that work, there may be room for a productized service at a much lower price than a full-time hire.

Here’s the rough thought experiment I was using: a U.S. data analyst might earn somewhere around $70,000 to $170,000 a year. Could you offer a narrower, useful analytics service for $2,000 a month, with Claude Data powering much of the work?

That is an idea to validate, not a promise of automatic margins. But the potential alignment is compelling: the customer spends less, you deliver a valuable outcome, and AI changes the economics of providing it.

The opportunity is not simply reselling access to Claude. It’s packaging a result that a business wants.

The tools make more work possible. You still have to figure out what customers need, what you can reliably deliver, and whether they’ll pay for it.

Qwen 4 Max: Another Model Worth Testing, Not Another Tool You Must Adopt

Qwen 4 Max was presented as a preview rather than a full release, with early access across Alibaba’s Qwen offerings. The proposed use case that stood out was getting a second opinion on difficult coding tasks.

I’ve generally been impressed with Qwen’s speed, pricing, and rate of improvement. It’s a serious model family.

At the same time, model selection isn’t just about capability. Alibaba is a Chinese company, and some people and businesses will have concerns about using a provider based in that jurisdiction. Others will make a different decision.

My view is straightforward: recognize the capability, then decide whether the provider fits your requirements and comfort level. There are plenty of strong alternatives. You don’t need to use every model just because it exists.

Multiplayer AI Is Becoming the New Default

If there’s one product trend connecting several of these launches, it’s the shift from single-player AI to multiplayer AI.

The first wave was largely one person working with one assistant. That’s useful, but it’s not how most teams operate. People collaborate, divide responsibilities, share context, and coordinate around the same project.

Their agents should be able to do that too.

Cursor: Your agents working with your teammates’ agents

Cursor’s rebuilt Projects experience brings agents into a shared team workspace. The practical use case is splitting a feature across a team without everyone, or everyone’s agents, stepping on the same code.

That makes a lot more sense than keeping every agent isolated.

I’ve been spending more time in Cursor, and this kind of coordination makes me more interested in bringing my team into the same environment. When the workspace supports both the people and their agents, choosing an editor becomes a team workflow decision, not just an individual preference.

I hadn’t tested the multiplayer experience deeply enough to give a final verdict. But the direction is exciting.

Grokbot: A shared front door for agents

The Grokbot 1.0 announcement pointed in the same direction. The features being highlighted included:

• Team bots that colleagues or friends can interact with together.
• Bots joining video calls and taking actions based on the discussion.
• Better memory of how individuals and teams work.
• Improved browser use.

The video-call capability is particularly interesting. An agent that participates in a conversation and acts on it starts to feel less like a separate tool and more like part of the workflow.

Memory matters just as much. A big part of trusting an employee is knowing they’ll learn how the team operates. The same expectation applies to an agent. If it repeatedly forgets preferences and working conventions, you keep paying the coordination cost.

I also think it’s useful to separate the orchestrator from the underlying model. You can like an agent interface without believing its default model is best at every task.

In the setup I discussed with Ras Mic, Grokbot connected with tools such as Claude Code as part of a broader workflow. That’s the model I find compelling: one front door, with different systems handling the jobs they’re best suited for.

If you’re building an AI startup, ask yourself where collaboration belongs in your product. What changes when multiple people, and their agents, need to work together?

ChatGPT Business Listings Could Create a New Local Agency Category

The Businesses in ChatGPT beta is one of the announcements that immediately made me think about startup ideas.

The concept is simple: a business claims its listing and adds an agent that can answer customer questions inside ChatGPT. Local stores and service businesses get the same questions all day. This gives them another potential place to answer those questions and connect with customers.

The comparison that comes to mind is Google’s business listings. An entire ecosystem grew around helping businesses claim, improve, and manage their presence there.

Could something similar happen inside ChatGPT? I think it’s worth taking seriously.

There is a caveat: my understanding of the beta was that it focused on physical, local businesses, such as toy stores and coffee shops. I wouldn’t assume that an internet-only business can claim a listing in the same way.

The startup idea: Set it up for local businesses

Many business owners won’t want to learn another platform, work out the setup, and maintain the experience themselves.

That creates a potential service opportunity:

1. Help an eligible local business establish its presence.
2. Set up the information and agent experience around common customer questions.
3. Help the owner improve and maintain that presence over time.

This is the same broad pattern that supported agencies around Google business profiles: a new platform creates a useful capability, and businesses need someone they trust to help them use it.

I’m actively exploring this space with a business partner. I don’t yet know exactly what we’ll build. But I’m happy to share the idea because I don’t see opportunity as a tiny pie that disappears when someone else hears about it.

Model Routing and Billing Are Becoming Part of the Product Stack

The Stripe and OpenRouter announcement caught my attention for a different reason: infrastructure.

The strategic idea is bringing payments and model routing closer together, so businesses can build AI products without treating usage billing as an entirely separate problem.

If you’re unfamiliar with model routing, here’s the simplest explanation: you may not want every task going to the same model.

• A repetitive, straightforward task may work well with a cheaper model.
• A difficult reasoning or coding problem may justify a more capable, expensive model.

A routing layer helps direct work accordingly. It can also reduce dependence on a single model provider. OpenRouter’s documentation provides a practical starting point for understanding its multi-model interface.

This gets more important as token spending grows. A small difference in cost per task can become a meaningful business expense at scale.

I wouldn’t anchor a decision on the acquisition price estimates being passed around during the discussion. The more useful takeaway is the direction: model selection, token economics, and billing are becoming core parts of building an AI product.

How I’d Start Becoming a Marketing Engineer

A marketing engineer is someone who does marketing with AI agents.

That sounds like a small change in wording. It’s a big change in the rate of experimentation.

Marketing has moved through print, radio, television, and the internet. Each shift changed how campaigns were produced and distributed. The internet made it possible to take more shots. AI pushes that further, toward workflows capable of dozens or potentially hundreds of experiments a day.

But more output doesn’t automatically mean better marketing.

1. Learn the fundamentals before automating them

I’d start with classic marketing, including David Ogilvy and other strong work from before the AI era.

Human psychology, positioning, copywriting, and creative judgment still matter. A faster production system doesn’t replace understanding why someone should care.

I’d also put those fundamentals into my second brain, meaning the reference material my AI can draw on. One mistake I see is filling a knowledge system entirely with recent AI-era advice while ignoring decades of useful thinking.

Use books, notes, and other material you can legitimately access. The goal is to give the system a stronger foundation than whatever happened to trend this week.

2. Learn the creative platforms one at a time

Next, I’d get hands-on with the tools for images, video, audio, writing, and other creative work.

I would not try to master everything simultaneously. Go platform by platform. Learn what each can do, where it struggles, and what kinds of direction improve the result.

A small example: shortly before the live session, we needed a thumbnail. Working with Ashley, our producer, I used Claude to come up with one in roughly six minutes or less.

I’ve spent hundreds of dollars on individual thumbnails before. This was dramatically faster and cheaper. Whether it would outperform the alternatives was still an open question, but that’s exactly the point: the cost of trying a creative direction has changed.

3. Get a real business result

Once you’ve learned enough to be useful, work on something real.

Maybe a friend works at an ice cream shop. Offer to help with a small project for free. Use it to develop your ability and build a case study. Then move toward paid work.

That’s the direction I’d take for a 30-day start:

1. Study durable marketing principles.
2. Build an AI-accessible reference library.
3. Practice with creative tools sequentially.
4. Apply the skills to a real business.
5. Turn demonstrated results into paid work.

The objective isn’t to collect tool subscriptions. It’s to become someone who can produce outcomes.

Why “learn marketing” is the new “learn to code”

I studied computer science, and I’ve built and coded things. I still find marketing extraordinarily hard.

Coding often feels more scientific to me. Marketing involves art, psychology, culture, writing, and visual communication. Marketing engineering is where that creative judgment meets a more systematic process.

As building gets easier, distribution becomes even more valuable. If you’re a developer, designer, or operations person, learning marketing can make the rest of your skill set much more powerful.

Agents Need a Control Panel, Not Blind Trust

I love experimenting with agents. I also understand the anxiety that comes from connecting them to more and more of your life.

That’s why Shane Mack’s Agent Picnic caught my attention. The screenshots suggested a central place to understand agent connections, shared memory, data access, and controls, including removing access for individual agent platforms.

I hadn’t used it yet, so this was interest rather than a tested recommendation. But the need feels obvious.

Using agents and trusting agents are not the same thing. As they gain more access and autonomy, understanding what they can reach becomes part of using them responsibly.

This is also why I won’t casually recommend how someone should connect third-party models to a database-backed second brain without understanding the setup. The details matter, especially when the question involves access and security.

Better AI Design Starts with Better References and Skills

AI-generated products do not all have to look the same. But the agent needs context about what good looks and feels like.

Stage Show: More polished visual demonstrations

Benji Taylor’s Stage Show skill stood out for helping agents create better UI stills, animations, bento grids, and related visual work.

The broader lesson is that specialized skills can give a general-purpose coding agent more specific design direction.

One example used Stage Show to help create a demonstration for a hand-drawn UI library. The workflow scripted a macOS cursor through real components, flagged visual bugs, and recorded the result frame by frame.

I loved how different it felt. There was personality in the components and interactions.

That’s an opportunity. Instead of accepting the first generic design an agent produces, give it a reference with a clear point of view.

Gestures: The small interactions that build trust

Emil Kowalski’s gestures skill focused on details such as rubber-banding, drag velocity, and momentum.

Those can sound like tiny implementation choices. In a mobile app, they affect how the whole product feels.

A drag that responds naturally, carries momentum, and settles properly can make an interface feel more considered. Those details contribute to trust and brand, even when someone can’t explain exactly why the app feels better.

Skills are also a distribution opportunity

There’s a business idea here beyond using the skills yourself.

You can create a skill, open-source it, sell it, or make it free and offer something else around it. If agent platforms become a primary interface for work, useful skills become a way to reach people inside that interface.

Think about the sequence: build something useful, earn trust, establish a brand, then offer a deeper product or service.

Ask Your Coding Agent to Make Changes Verifiable

One of the most useful prompt ideas I came across was to ask an agent how a proposed change could be made fully verifiable.

It’s a simple shift, but an important one. Instead of treating “I implemented it” as the finish line, you ask what evidence would demonstrate that the change actually works.

That can push the workflow toward real front-end interactions and checks against actual services, rather than stopping at a confident explanation.

Make verification part of the assignment, not an afterthought.

This fits the larger direction of AI development: agents can do more work, but the value of that work depends on proving that the result is correct.

Software Isn’t Dead. Low-Value Software Is Under Pressure.

I don’t buy the idea that there’s no software left to build.

Software is evolving. Some products will struggle if they don’t evolve with it. That’s very different from saying the category disappears.

There should be pressure on subscriptions that don’t add value and survive mainly because customers forgot about them. But people still pay for convenience, useful outcomes, and someone else taking responsibility.

They want someone to:

• Keep the product secure.
• Maintain it.
• Ship updates.
• Be accountable when something breaks.

Why would all of that suddenly stop mattering?

At the same time, personal software is becoming more practical. The custom TV interface example I shared made this concrete: someone used AI coding to create a cleaner experience than the bloated interface their device shipped with.

I have a Samsung TV, and that idea immediately resonated. There are frustrating software experiences everywhere. Making them better is still valuable.

Other Opportunities Worth Keeping in View

Local AI implementation

I’m bullish on helping businesses set up local AI.

The analogy is the move from on-premises servers to cloud infrastructure. Businesses of many sizes needed help making that transition, and implementation companies captured real value by guiding them through it.

AI creates another transition that businesses may not want to navigate alone.

Even cloud implementation hasn’t stopped being useful. Companies still want a trusted person behind the wheel rather than simply handing an agent access to infrastructure and hoping for the best. There’s room to position that work around AI workflows too.

SEO fundamentals still matter for AI search

My hot take on AEO and GEO, the efforts to improve visibility in AI-generated answers, is that good SEO remains a strong foundation.

I mean the legitimate fundamentals: build a good brand, publish something useful, and earn attention from credible publications. Not tricks built around gaming search systems.

Google’s SEO Starter Guide is a useful reference for those basics. The terminology may change, but I wouldn’t throw away durable principles just because there’s a new acronym.

Voice AI and upcoming releases

I also think voice AI is underrated. The quality has reached a point where it deserves a serious look, especially for people exploring marketing and service workflows.

Other items on the announcement radar included a planned Grok 4.8 rollout and OpenAI’s upcoming cybersecurity keynote, framed around turning lessons from defending its own systems into practical customer tools.

Those were things to monitor, not finished products I’d already evaluated. Keeping that distinction clear helps separate useful signal from launch excitement.

What Makes an AI Business Durable?

A comment from Mercor’s CEO argued that brand and proprietary data would be the two primary moats for application-layer companies.

I agree that both matter. I also think the picture is broader.

• Brand: Give people a reason to recognize and trust you.
• Proprietary data: Develop something valuable that isn’t equally available to every competitor.
• Network effects: Consider whether participation makes the product more useful.
• Content systems: Build a repeatable way to earn attention and develop distribution.

The last one connects directly to marketing engineering. If you can build a system that consistently produces useful content, you have an advantage beyond shipping a product once.

For someone starting in software engineering, though, I’d keep the first milestone much simpler: build an app, get a customer, and make a dollar on the internet.

You can spend forever debating moats before anyone wants what you’ve built.

Don’t Let AI Overwhelm Turn into Inaction

The pace is overwhelming. I feel it too.

But I don’t want that feeling to turn into quicksand. You can look at these launches as an impossible pile of things to learn, or as a growing set of ways to solve problems.

Questions about industry scrutiny, infrastructure costs, and changing platforms are real. They don’t remove the opportunity to build something useful. On hosting costs, for example, the question I care about is whether the return justifies the expense, not whether the expense exists.

I choose to approach this era with an abundance mindset. Not because every idea will work, but because there are so many useful things still waiting to be built.

If you want a starting point for exploring opportunities, I built Idea Browser to help surface startup ideas and trends. There’s a free option, so you can begin exploring before deciding whether a paid plan is useful.

You don’t need to master every model. Pick a customer problem, learn the tools that help solve it, and build something real. Then verify it, put it in someone’s hands, and learn what they value.

The opportunity isn’t keeping up with every announcement. It’s turning one of them into something that matters.

Disclosure: Brex sponsored the live session behind these observations. I use its banking, card, and finance tools across the businesses in my holding company, Late Checkout.