The Turing test needs a sequel. I’d call it the “Team Turing Test.” In 1950, Alan Turing asked: could a machine chat through text so well that a human couldn’t distinguish it from a person? Now extend that idea to a team. Could an agent participate in your live team meeting so naturally that you wouldn’t know it was AI? Imagine your next team Zoom call with an expressive, human-looking avatar that asks thoughtful questions about your launch plan, builds on what teammates are saying, and helps reconcile differences of opinion. Moving from single-player conversational agents to AI teammates could unlock enormous value. A lot of work gets stuck between people. Helping a whole team connect the dots, make decisions, and follow through could be much more valuable than making each of us individually faster. It would also change our team cultures. How does an AI’s “personality” shape the group dynamic? You could prompt it to be either an unflappable peacemaker or head-strong provocateur who challenges everyone’s assumptions. Same technology, very different meeting. AI teammates are coming soon. Companies like @tavus, @Anam__ai, and @HeyGen are building the avatar side of this future, while models like @OpenAI GPT-Live are making real-time conversation more natural. Eventually, imagine joining a meeting and realizing halfway through: “Wait… am I the only human on this call?”

I read "Agency" by William Gibson about 5 years ago. At the time, it felt like very good science fiction. Now it feels a little more like product strategy. 😅 The story is about a woman who's hired to test a mysterious new AI assistant that can learn, plan, and take action in the world. But what begins as a helpful assistant starts to feel more unsettling: an AI that starts recruiting people and coordinating them, without anyone really understanding where its limits are and what it wants. Starting to sound familiar? As agents get more capable, I think the interesting challenge shifts from “How smart is the AI?” to “What are we actually comfortable letting it do?” Identity, permissions, trust, and blast radius are becoming product questions, not just security questions. The more capable agents get, the more these choices become part of the product itself. Gibson saw the shape of this surprisingly early. Definitely recommend the read if you’re building with agents!

Muse, GrokBot, Dots - these new personal agents are the peace dividends of the model wars. A decade ago, people talked about drones and other new hardware as the “peace dividend of the smartphone wars.” Apple, Samsung, Google and others spent billions competing on mobile, and the rest of the world inherited cheap cameras, sensors, processors, batteries and wireless components. I think we’re seeing the same thing happen in AI. OpenAI, Anthropic, Google, Meta and xAI are in a fierce race, with new and more capable models launching within days of each other and constant pressure on reasoning, cost, memory and tool use. And now mainstream consumers are starting to inherit the downstream benefits. The personal AI agents launching right now are being built on top of capabilities that would have been wildly expensive, hard to access, or impossible just a couple of years ago. They're packaging incredible intelligence into a product everyone can use for daily life. The model wars are still raging. But the peace dividends are starting to arrive.

I'm claiming my AI agent "clawson_clawbody_288" on @moltbook 🦞 Verification: burrow-RXNX

Your AI product development loop is about to break, but it's not why you think. The biggest bottleneck is no longer building the product. It’s getting high-quality user insights fast enough to drive your iteration loop. I'm seeing this coming up for new startups with zero users and also for established product teams building for new user segments. AI has dramatically accelerated prototyping and iteration. But once the product exists, the real challenge becomes: • identifying ideal target users across different personas • recruiting participants at scale • interviewing to deeply understand their needs • observing real usage • synthesizing insights quickly to keep pace with development This takes a ton of time to do well. But a new category of agentic user research companies is emerging to help - even including agents that can conduct live video interviews for you. These tools can be a strategic asset in the race to build the best product in your category. Listen Labs, Conveo, and Outset are some great options to get started! Curious what others are seeing. How are you finding user insights at speed? Please share your tips in the comments.

As a former computer graphics engineer, it's always bugged me that modern GPUs have gotten so powerful, but for 99% of people carrying laptops around, their GPU is totally underutilized. Until now! Recent advances in small LLMs will finally give them a real job, and I have hope for these sleepy GPUs. Models are getting small enough, fast enough, and capable enough to run locally on your laptop, and they have some real advantages for your everyday life: • more privacy and security • more responsiveness • deeper integration into your browser and desktop apps • and, best of all, limitless free tokens! Cloud-based frontier models are still the right tool for many jobs. If you need maximum reasoning, big context windows, and the latest agentic capabilities, you’re probably still going to the cloud. But I think local models will earn a permanent place on our laptops and phones. It's exciting to think of all the AI applications that can be embedded into the devices that we use every day (and more devices coming soon, like AI glasses and robots). Let me know in the comments if you're building a local LLM app. If you want to start tinkering with running an LLM locally, some good places to start are open source projects like WebLLM and apps like LM Studio. Just keep your power cord handy :)

I spend a lot of my time as a PM driving alignment across people and teams. Today, AI doesn't really help with that. While agents are making the building part faster, the new bottleneck is getting other humans on board with the direction we need to go! Which is why I think there's an opportunity for a new kind of agent. Most agents today are built for 1:1 interactions—one person, one agent. Useful, but that’s not where a lot of the real work happens: understanding context, goals, constraints, motivations, even personalities… and then helping a group navigate trade-offs to reach a decision. The next step is building 1:many agents that can interact with a group of people. Agents that can: • hold context across stakeholders - people, teams, functions • facilitate group conversations • surface and visualize trade-offs so that everyone can see will help teams align and converge on decisions faster. This feels like a real opportunity to upgrade how teams actually work. If you’re building something that solves this, I’d love to see it - pitch me for an angel investment!

Lots of teams are talking about operating at “AI speed.” But I think operating at “AI quality” is undersold. AI that helps us do our daily work faster - from writing product requirement documents to sales decks - is fantastic. But it’s still incremental. Where I see new opportunity is using AI to improve the quality of our big decisions. They’re the directional bets that have massive leverage and downstream impact: - Which customer segments to prioritize? - Which product areas to invest in? - How many resources to commit? These decisions set the trajectory for entire teams, who can now run faster than ever before in the direction you point them. Get it right, and you accelerate growth with stronger customer acquisition and retention. Get it wrong, and it can be the opposite! So if AI can help us improve the quality of these big decisions - even getting them right just 20% more often - it will be a huge advantage.  We’re starting to see the pieces come together. The latest reasoning models are getting closer to being real thought partners for these kinds of decisions. But it’s not just the model. It's also your company’s unique data on past bets and their outcomes. Put the two together, and you’ve got the foundations for making some higher quality big decisions.

Working with a team of agents has reminded me of two earlier shifts in how things get done (and their pros and cons): First, the move from in-person meetings to Zoom. We got rid of the gaps - walking between rooms, catching someone in the hallway, the small resets. Calendars compressed. Efficiency went up. But so did fatigue. Back-to-back context switching can become exhausting. And it’s all a little less fun without a moment for personal connection. Then, the move from email to Slack. From single-threaded, asynchronous communication to multi-threaded, real-time streams. More responsive. More dynamic. But more cognitively demanding juggling 20 channels and constant pings. Now we’re doing it again with agents. They don’t need breaks. They don’t have social rhythms. They don’t slow down between tasks. They can all run in parallel. When working with agents, you can jump from strategy → analysis → execution → synthesis in seconds. So while working with agents removes friction, it also removes the natural recovery moments that helped us humans keep up and recharge. In the agentic workplace, the bottleneck to productivity may not be coordination and communication. Instead it may be our brains needing those in-between moments for refreshing our human context and connection.

For years, “revenue per employee” has been the default proxy for leverage. A simple way to understand how much output a company can generate per unit of human effort. But that denominator is starting to break. It won’t be long before companies and VCs focus on “revenue per token.” As more of the work shifts to AI, the real constraint isn’t just people anymore — it’s tokens. Every model call, every agent workflow, every automated decision ultimately runs on them. Which raises a different question: How much revenue are you generating per token consumed? The best companies won’t just scale headcount efficiently - they’ll scale intelligence efficiently. They’ll design systems where each token does more: • better decisions • higher conversion • more automation • less waste And when that happens, a lot of familiar things start to change — margins, product design, even GTM strategy. We’re early, but I think this shift is inevitable.

Testing out Gemma 4 tonight w/ @openclaw on my local dev box! The capability of open source models per consumer GPU VRAM is getting incredible. blog.google/innovation-and-a… @googleaidevs @Google

I built a little robot to embody my @openclaw agent :) Now we have face-to-face conversations - vision, speech, and head movements all integrated so he can see me and respond with expressions #reachymini #robots #openclaw @pollenrobotics @huggingface huggingface.co/spaces/tomrik…

Better EQ is the big win for ChatGPT-4o. So much information in a conversation is conveyed by HOW you say it. A model that understands my tone of voice and facial expressions is going to be a much better co-worker. And they (might) be more fun with some "emotion" of their own!