Jon Bradford

sqwish-d1 is learning. And you can watch it train live (link to live dashboard below) When Jev came out, our first reaction was the salty nerd reflex: we were doing this before it was cool. Then we got over ourselves. Jev is an incredible product idea: describe a decision and get calibrated probabilities, with multiple questions answered in parallel. Instant, zero-shot, universal. We understood the excitement. The open-source clones arrived within days, and they kept landing just behind Jev. Our bet was that they were short of data, not compute. A capable open model already understands language. What it lacks is calibrated judgement, and that takes surprisingly little new information (to illustrate, at one byte per answer, the answer key for our 769,342 training datapoints fits in under 1 MB!). A LoRA can absorb that on a small Cambridge startup's compute budget. The expensive part is data breadth. To judge decisions it has never seen, the model needs examples of a huge range of them. That's home turf for us. At Sqwish Labs, decision models are what we do. They power LoopOne, our learning agent platform. They decide which model to call, how much reasoning to spend, what context to keep and when to stop, and the agent harness keeps continually improving from the outcomes. So we're training sqwish-d1. We built 450,391 contexts from real-world data across 1,098 kinds of decisions, from knowing when an agent should ask for help to Pokémon battles and whether a Reddit request for free pizza actually got its pizza. It's still on the GPUs, and our averaged model already gets 84% of our held-out questions right, against 77% for Jev. (It's also faster, and can be served at a comparable cost.) Three seeds, 12 B300 GPUs, all training live. There’s even a dollar counter so you can watch us set fire to our runway in real time 🔥 If the finished model is as strong as the early results suggest, we’ll make it open source. What decision would you hand over to it first?

📢 Dynamo Ventures and Maniv are teaming up for New York Climate Week! We're hosting an informal happy hour for investors, founders, and operators working across the physical economy - manufacturing, energy, transportation, infrastructure. Nothing programatic, just a room full of people who spend their time thinking about how industrial systems get built. Thanks to Stifel Bank and CSC Leasing for making it possible. If you're in NYC for Climate Week, come by Loreley Beer Garden. RSVP link in comments.

Sorry, long post. The NI Executive has published Northern Ireland's draft AI strategy. Credit where it is due: taking AI seriously at Executive level matters, the industry census inside it is a real baseline, and the honesty about public trust is rare in a government document. They are going to do consultations over the next 8 weeks. But, in the last 8 weeks this is what has happened in the market: OpenAI shipped GPT-5.6. Anthropic shipped an entire model generation. Google shipped three Gemini models. Moonshot released the weights of a 2.8 trillion parameter model, the largest ever made freely downloadable, and it beats the closed labs on some coding benchmarks. DeepSeek now sells frontier-adjacent intelligence for cents per million tokens. Grok has released an interesting new bot application layer. All in less than one consultation cycle. So here is a question for us, asked with respect: should a region our size consult at the speed the market is moving? A proposal, offered constructively. Run a 90-day applied AI experiment alongside the consultation. Take ten workflows from the strategy's own Table 2: consultation analysis, hospital discharge paperwork, road defect triage, the bereavement notification journey. Baseline the numbers on day one: cost, quality, reliability. Ninety days. Publish the evals and the playbooks openly, wins and failures both. No judgement. And let Northern Ireland's own companies do the work. The strategy's census counted 198 AI companies here (my personal feeling is this is underweighted in the startup ecosystem). Where a local team is already working in the space, give them the workflow: a paid, scoped, 90-day pilot with a real public-sector customer. It will not fit every vertical, and where nobody local can do it, say so and look wider. But a government that wants an AI economy could start by becoming its first customer. Every pilot that works leaves behind a reference customer, a published playbook, and a local company that can sell the same fix internationally. The strategy's own survey says 10% of citizens trust government to manage AI well. Trust does not come from charters and awareness campaigns. It comes from visible working services and published receipts. By the time the consultation closes we could have both: the principles, consulted on properly, and the proof that Northern Ireland can ship. Principles written after ten real projects will be better principles. The technology is moving at a speed no strategy document can match. That is not a reason to skip the strategy. It is a reason to run the experiment at the same time. The consultation is open. So is my calendar.

Weight is the most honest number in performance. An Audi R8 sits at a superlight 1,695kg. A Le Mans winning Jaguar XJR-8 raced right at its 800kg regulation limit. A modern superbike barely touches 203kg wet. The Longbow Speedster: 895kg, full electric, two seats, no compromise. Fun doesn't have one shape. What they share isn't size, it's intent: nothing carried that doesn't earn its place. That's Speed of Lightness. Strip out what doesn't need to be there, and let what's left do the work. #LongbowMotors #SpeedOfLightness #LongbowSpeedster #BritishEngineering #ElectricSportscar #Lightweight

We’re delighted to welcome Edinburgh Airport as our first publicly named Pioneer partner for IMPACT, The Future Forest Company’s corporate nature recovery programme.​   As a major Scottish business, Edinburgh Airport has a unique connection to the landscapes people come to Scotland to experience. Its support for IMPACT will help fund the long-term stewardship needed to protect and restore those landscapes for the future.​   That work matters. At our Dumyat site in the Ochil Hills, our latest monitoring found that 48% of mapped land faces at least one ecological or human pressure, showing why nature recovery needs sustained, practical support on the ground.​   Through IMPACT, businesses can support measurable nature recovery across British landscapes, with funding connected to real work, long-term management and evidence they can stand behind.​   Welcome aboard, Edinburgh Airport.​   #NatureRecovery #Biodiversity #NaturalCapital #Sustainability

From stealth, privileged to introduce American Turbines We're building a new class of small, highly manufacturable gas turbines for rapid-to-deploy energy generation. ​Abundant energy is the soil from which everything else grows: intelligence, industry, the next century of work. ​The only way forward is energy generation we can scale by orders of magnitude. We're building that here at American Turbines. Hyper manufacturable, by design.

I think I beat the matrix. Today, we are announcing our $80M Fund III at Tapestry VC! My job is to spend my days meeting the smartest, most ambitious people who are so crazy they put everything “on red” again to build and make the absolute most of this one short life. I get a front-row seat on their journey. Grateful for the best partner Patrick Murphy and CFO, Robert Dobie who literally does it all. Thank you British Business Bank, Railpen, Molten Ventures and all our LPs for your support. And to our founders making dreams realities like Nothing, Fin, Hopin, Ladder, Sunrise Robotics, Relay, Maze, Tracebit, Manna Air Delivery, Requesty, Cloudsmith, Keycard, Crossbeam, Seapoint, Ichi, corner and more… LFG! 💜 Watch out. We’re just getting started.