bryce

Happy to share we signed our first paid pilot to build an organizational world model for a manufacturing business. It gives their team and agents a live view of what’s true, uncertain, changing, and what to do next, a belief state. Today’s agents have memory. But once they start doing real work, they need to know what world they’re in and the current state of things. Mental model: memory is the rearview mirror. belief infrastructure is the windshield. world models need both. PS: new website hero. wdyt?

@liamjmcgregor Credit: @Skymography

Who is behind Anthropic's latest design work? It's an extraordinary mix of biological illustration, ASCII art, halftones… Physical materials + data scale. Call it "Humanist collage"? Especially fable announcement + J space article. It looks different than the OG Geist work.

@Skymography Great storytelling and cinematography.

Change is a cycle, you just haven’t noticed yet

@NVoitenkov @GenwayAI @speedrun Enterprises can’t delegate real work to AI. Agents act on what they think is true, not what is true. We fix that by giving agents shared operating state to plan and act from, a world model. Just landed first six figure paid customer. thinkn.ai/manifesto

its official - we are shutting down @GenwayAI and I am giving away all my @speedrun materials. i hope it will help others raise capital and build something epic Pre-Seed Investor deck and live demo day pitch recording. Comment with a blurb of what you are building, and I will DM you the materials. Alternatively text your email to +1 (567) 485-2649 and my agent will send those to your inbox right away PS - I still have a few a16z/SR scout checks in me, make sure to include some traction/validation points in your blurb to make my life easier.

@capital_3x Enterprises want operating leverage from AI. We built the infrastructure.

you're pitching your startup in exactly 10 words. no more. drop your pitch below. we'll read every one.

@16vchq Enterprises want operating leverage from AI. We built the infrastructure.

you're pitching your startup in exactly 10 words. no more. drop your pitch below. we'll read every one.

@g_k_swamy cool to see that goal-directed agents already encode a world model, even if it is only implicit in the policy. for agents in a business, can it stay hidden or does it need to become explicit and inspectable? this is a safety question as much as a performance question.

Indeed an underrated paper :)

The day you stop checking your agents is the day AI changes your business.

@joannezchen This multi-agent world requires shared operating state to plan and act from. That's what I'm building. Operating state for working agents. More formally, a world model. Just signed a six figure contract with a critical infra manufacturer and its agents.

We're just beginning to see the next big cohort of startups: Companies built for an agentic customer, not a human one. Take email. Today's software is built around a visual interface - for you and me to read and type in. But if an agent is reading the emails, sorting them, and writing the replies, it's not looking at a screen or clicking buttons. As agents take over more of the repeated, operational work, human engagement changes form: we manage the agents rather than do the tasks. That opens the door for startups to redesign software for an agentic world. Each of us will have agents working on our behalf. Soon there will be billions, probably trillions of them. And no one is going to stare at a spreadsheet asking what Agent 482 did yesterday. Tools that orchestrate fleets of agents are becoming a credible investment opportunity. At @FoundationCap, we've already invested in @arizeai and @sixtyfourai. I think we'll see many more players emerge here in the coming years.

@Nickprince Accurate.

claude: seasoned eng, strategic and flirting with moving to product codex: cracked staff eng, gets shit done grok: waterloo intern, absolutely relentless and on the path to greatness but still has a lot to learn

fundraising deserves a repo, not a spreadsheet. so I built Round. fully open source. runs locally. native to Claude Code. manage your investors, pipeline, tasks, cap table, data room, call notes, and multi-round scenarios from files Claude can actually work with. I’ve been using it for my own raise. now it’s open source. github.com/round-sh/round

all software is becoming headless and open source. even frontier labs are subject to disruption. a contrarian belief is that services and the experience you deliver around the software are where the moat builds in the future, not the software. services are no longer heavy.

World models can sound complex or abstract, but humans operate from them every day. Simple world model example: You look outside to see if its going to rain. Clouds are forming. Not enough evidence to know. Next best action: You check the weather on your phone. Rain is coming. Your picture of reality updates. So you grab a jacket. World model = a live belief about the current state, updated by new signals, so you can choose the next best action. Agents need the same thing.

@arcprize @gen_intuition @PimDeWitte @EloiAlonso1 @AdamJelley2 @micheli_vincent Leet.

ARC Prize Foundation + General Intuition We’re grateful to announce that @gen_intuition has donated $1,001,337 to support the development of ARC-AGI-4 Thank you @PimDeWitte, @EloiAlonso1, @AdamJelley2, and @micheli_vincent for helping push open AGI benchmarking forward

Claude coding? Make sure you create adversarial /workflows that regularly challenge load-bearing assumptions in your code. "I misdiagnosed four times from indirect evidence; the adversarial multi-agent workflow found the truth my solo guessing couldn't."

@naval @frankdegods rediscovered** Timing matters.

Great products are not so much invented as discovered.

@vivianmshen I signed a $120k ARR contract. Should I apply to the next batch?

just had a YC S26 founder sign a $60K ARR contract before the batch kickoff this week surprisingly, setting ambitious goals works

@FoundationCap good episode! reminds me of @annimaniac’s L0–L5 framing. most companies are using AI to improve personal productivity, then wondering why the company doesn’t move 2x faster. real delegation needs a shared operating state agents can plan and act from. x.com/annimaniac/sta…

As AI takes on more work inside organizations, it reveals where they are too slow, too rigid, or too dependent on processes built around human constraints. AI is already helping individuals work faster, but many organizations are struggling to translate that into gains at the team or company level. @azeem calls this problem “congestion.” When one part of the company speeds up, the next downstream process becomes the bottleneck. AI-native startups are able to bypass many of these blocks because they’re designed around agent workflows from the ground up. In the latest episode of AI in the Real World with Foundation Capital, @joannezchen and @azeem explore what it will take to build AI-native companies. The org chart is being redrawn, and the companies that thrive will be the ones that learn how to collaborate with agents instead of simply bolting AI to old workflows. Listen to the full episode here, now available on Spotify and other podcast platforms: open.spotify.com/episode/56kYxG…

@shadcn Models don’t have an explicit belief state. When you give them one, they can easily answer this question without drifting.

No model survives the "are you sure?" They all fold instantly.

Models don’t maintain an external belief state. Give the agent one and it can easily answer this question without drifting.

@xuezhao agents making agents only compounds if each new agent inherits a better model of the world. one it can test, challenge, and improve. otherwise the next agent just inherits the last agent’s assumptions and runs with them.

the investment thesis going forward is simple: invest in agents making agents, robots making robots, and energy generating energy. that’s it.