The robot messed up the fold. Then folded it correctly. No one taught it that. Check out the new research from Sereact.

Robots that learn by copying people hit a ceiling. Cortex 2.6 now learns from practice: after 50 attempts per task, its success on three returns-handling tasks rose from 60–70% to 85–100%.

Tried a bunch of neoclouds / neocloud adjacent services (@FireworksAI_HQ , @togethercompute , @baseten , @nebiusai) the other day and me and my agent made the following observations. I was running benchmarks and deploying a bunch of OSS models via Claude Code on these services, and doing some internal benchmarking. Maybe I messed something up, but has anyone does some comparisons across these providers?

Reinforcement learning on real robots New research developed in the field by @TuscherMarc @SereactAI

I've built this little website to view my @goodreads bookshelf in a much nicer way: bookshelf-aea8f1.netlify.app If you'd like to get a version for your books, LMK and I'll open-source the code on Github ... or I'll just give you the prompt ... or you just point Claude/Codex to the URL and tell it to copy/rebuild it. :)

POV: you visit a customer site running @SensmoreAI

We are still living in interesting times. The world is changing very fast, and after 10 years of full-time investing, it is time for me to open a new chapter of my career. I have therefore decided to step down from my role as a General Partner and transition to a new role. My journey with Point Nine has been beyond anything I could have imagined. I joined as a fresh MIT graduate almost exactly 10 years ago after meeting Pawel and Christoph on Skype, became a General Partner alongside Ricardo 7 years ago, and have loved every single bit of my day-to-day ever since. To my P9 partners, to all P9 team members (current and past), to all P9 LPs, and to all P9 entrepreneurs (P9 family members) I’ve worked with over the past 10 years, thank you. I am proud of the work we’ve done to build a venture firm that is different, more human, more approachable, and entirely dedicated to serving entrepreneurs building some of the most ambitious companies in Europe and beyond. Over the past few months, I've spent most of my time speaking with AI researchers at various universities across Europe about how to bring their research to market. This led to our investment in a legal-focused AI lab called grubel and made it clear to me that I should get involved with companies as early as possible, including taking an increasingly operational role. Starting now, I’ll continue to be part of Point Nine on a part-time basis as a Venture Partner, working with my portfolio companies, investing with the team in a few AI research projects at their inception, and figuring out what the next chapter of my career will be. To more adventures like this one, Louis

Most teams running at scale want to move their agent off OpenAI or Anthropic onto a cheaper model. Few do, because it's kinda hard to guarantee quality stays the same and there are no regressions (harder then it looks). We've been working on this problem the last few weeks. How to give a better guarantee of swapping to OSS looking at historical traces? In one example, we swapped the model for a fintech and a retail support agent across thousands of conversations and here are some findings: 1/ The usual check compares the cheap model’s next reply to the old one’s. It approved swaps that resolved 6–10 points fewer requests correctly. Replaying full conversations caught every one. 2/ A cheap model can look fine reply by reply and still fail the conversation: skip an identity check, offer a refund it shouldn’t, hand off too early. 3/ Serving matters as much as the model. FP8 cost ~10 points. One wrong config setting broke 95% of conversations. 4/ And recheck your judge. One tuned on GPT’s answers won’t automatically grade Qwen fairly. We’re building Kitaru to capture, replay, and evaluate agent runs. If you run an agent on a closed model, send me a sample of your traces and I’ll tell you what’s safe to swap and what you’d save. Can sign an NDA if needed. DMs open.

The real benchmark for physical AI is production. 300+ live systems across Europe and the US. One model learning from all of them. Proud to be #5 in @Siftedeu’s AI 100. Thank you to everyone building this with us. 🦾 #SiftedAI100

In the age of agents, systems of record have never mattered more. They’ve also never been less prepared for what’s coming. Today, we’re announcing @attio's hybrid architecture, with the controls of application software at the scale of infrastructure. To agents, Attio looks like a database, queryable in SQL - a language they're RL’d on. Unlike a database, the business logic, policies, and compliance sit behind the query. This makes all of Attio’s data accessible by agents in the most powerful and flexible syntax. It gives GTM teams an order of magnitude more value from their time, with the full power of agents working across the entire customer lifecycle. A new shape of software is here. attio.ai/architecture

🚀 Your customer data just got a lot more actionable! We just shipped: 📋 Contact Lists with powerful filters 👤 Contact Profiles ✉️ Batch email & sequence enrollment 🤖 Automated outreach with workflows All connected to your subscription and revenue data. chartmogul.com/blog/contact-…

Every EHR has its own data model, its own quirks and its own six-month integration project. Or you skip all of that.

Clio named to Fortune’s Future 120 - a list of companies with exceptional long term growth potential. @goclio

Another saying that I recently heard and completely agree with is that during a raise the founders should first solve for the people on the other side of the deal, then for the raise amount, and then the valuation.

Odyssey is pushing the frontier of world models day in day out

For many B2B software startups going B2C is a good idea now (back to consulting)

As everyone is doing it, here's mine 👇 asked Opus 5.5 to turn our website into a short explainer video of what CharMogul is all about. Incredible output for a 8 minute 1-shot attempt

With the acquisition of PTC by Schneider, there is no decent b-rep kernel owned by a US company anymore.

Most Nigerian enterprise CFOs are logging into five, sometimes ten, sometimes twelve different banking portals just to answer one question: Where is our cash right now? That question should take thirty seconds. For most African enterprise businesses managing multiple entities, subsidiaries, and banking relationships, it takes hours. And by the time the answer is assembled, it is already out of date.  We are changing that. Duplo Treasury Management Solution is live. It is the first treasury management platform built specifically for African enterprise businesses, giving finance teams near real-time visibility across all their bank accounts and entities, automated cash sweeping rules that move idle funds to where they are needed, AI-powered liquidity forecasting, bulk vendor payment execution, ERP integration with Microsoft Dynamics 365, and complete audit trail governance, all from a single dashboard. African enterprises shouldn't have to manage cash with tools not built for their operating environment. Global TMS platforms like Kyriba and HighRadius are powerful, but they were built for Western corporate treasury functions. Duplo Treasury Management is built for NIBSS connectivity, multi-bank Nigerian operations, CBN compliance, and the multi-entity complexity that comes with growing an African enterprise. Treasury that tracks cash is a cost centre. Treasury that optimises it is a competitive advantage. Read the full announcement at the link below.👇 hubs.li/Q04zg83j0  #DuploTreasury #TreasuryManagement #AfricanEnterprise #NigerianBusiness #Duplo #EnterpriseFinance