“don’t train your own model” is common ai advice. it's wrong. your token bill's the proof. today, we’re excited to launch castform into open preview. castform is the easiest way for you to train your own model, on your own data. open-weights models are performant and much cheaper. when trained on your task & proprietary data, they beat closed models. the thing standing between you and that was weeks of plumbing & years of ml expertise. with castform, model training is as simple as prompt engineering. @castformai bring your agent traces or raw corpora. castform turns it into training data, picks the right algorithmic recipes, manages gpus, and gives you an ide to watch and chat with your model as it learns. see what you can build with castform👇

We gave an agent one job: email our biggest prospect. A day later I checked what else it could open, and it could pull up our largest live renewal, pricing and all. Nobody set that up on purpose. Until I went looking, nobody could have told me it was true. @forrester has been bracing for this. They expect an agentic AI deployment to cause a publicly disclosed breach in 2026, and to trace back to governance failures rather than sophisticated attackers. This April it got real. Researchers showed that a prompt injection hidden in a customer-facing form could turn both @salesforce Agentforce and @Microsoft Copilot Studio into data-exfiltration tools. The data walked out, sent through a normal action the system saw no reason to block. The agent broke no rules. It used access it had been given, which is what made the leak so hard to spot. You cannot patch your way out of that. The only thing that limits the damage is knowing, ahead of time, exactly what an agent can reach. Most companies cannot produce that list today. Anyone who has run RevOps knows the request. Security or Legal asks who can see a given account and how they got there. You open the Sharing button in Salesforce, and next to half the names it just says "associated record owner or sharing". So you start opening rules, checking territories and guessing. A full afternoon gone and you still answer with "I think so." That is not a chain, it's a shrug. There is a name for the thing that question is really asking for: Access provenance. For any record, the full set of who and what can reach it, and the exact path each one took to get there. At @rox_ai, that question takes one view. Pick any account and you get every identity that can reach it, person or agent, with the path each one took. In the video, I show how this works in Rox. I set up an outbound agent with one job to outbound to just one account. Then I ran it on an account the agent was never pointed at. It came back reachable. The trace showed why in seconds. That same agent was sitting in a broader team too, left from a RevOps change months ago, and GTM West rolls up to it through a user-based hierarchy. A rep on that team owns the account, so access ran straight up the chain to the agent. I caught it before the agent ever ran. Read the same graph backward and it answers the other half: everyone, every person and every agent, who can reach a record, and the route each of them took. Access provenance is the layer nobody bothered to make legible, because until an agent was acting on it at speed, you could get away with not knowing. That era is over. We are building Rox enterprise-ready so this is the first thing you can answer about your data, not the last. Kudos to the platform team that worked on making the boring parts of enterprise software so "cool". We are hiring! jobs.ashbyhq.com/Rox-Data-Co…

Agents now Run Revenue, End-to-End. Rox Autopilot is here 🚀 The world’s largest enterprises now run their critical revenue systems on Autopilot. Coding and support agents had their moment. Now it’s time for Revenue Agents No SaaS. No enrichment. No CRM. Just an Agent. We’re offering 2× free agent actions for the next 30 days. Start Autopilot at rox.com Thank you to @sequoia, @generalcatalyst, @googleventures, @marcbhargava, @vedantsuri, @davemuni, @MongoDB, @cloudsoftware, @togethercompute, @Microsoft, @databricks, and all of our other customers & partners!

Weekly drop 102: we’ve tightened the link between insights, signals, and source documents. Research Insights now refer back to the signal that generated them. & Financial Documents (10Ks, 10Qs, Earnings Reports) now have clickable sources. Shoutout @luhemarora for these features! Release notes: rox.com/docs/about-rox/relea…

@joininteract has opened applications for the 2026 class. Interact was and is meaningful for me; if you resonate with the "young technologist" label, you should consider applying.

🧑‍🍳

No one: GPT 5.2: Nice 👍🏻

The newest Rox Slack emoji will definitely be :holiday-marco:🎄 @marcofleming

Shoutout to the eng team that never stops shipping! Still more we're building before EOY - and even bigger things coming in 2026 💪

Sitting at a table on the Caltrain if you're not using it should be illegal

demo nights are back. come hang out with us "at the edge" @southpkcommons (SF) next Thursday. we will have a *very* wide range of demos, including: - space lasers - emotionally expressive robots - "wise" jewelry - research around scaling multimodal inference

1/ The future of general-purpose robotics will be decided by one major question: which flavor of data scales reasoning? Every major lab represents a different bet. Over the past 3 months, @adam_patni, @vriishin, and I read the core research papers, spoke with staff at the major labs, and mapped the talent pool. This has completely changed how we think about general-purpose robotics. Our paper builds intuition, step-by step, across the 2025 frontier: from architectures → evals → data → industry dynamics. Each layer reveals a different bottleneck, but they all converge on one truth—data decides everything. Our takeaways process below👇 If you want access to our graph (sound on), comment or DM me

Beyond Being There is the Bitter Lesson equivalent of HCI

There’s been a lot of work on unlearning in LLMs, trying to erase memorization without hurting capabilities — but we haven’t seen much success. ❓What if unlearning is actually doomed from the start? 👇This thread explains why and how *memorization sinks* offer a new way forward.

Incredible things happening over in the twitter repo