Jason Demant

10 years ago, I backed @andrewdfeldman before there was a company. Now, he is defining the future of AI. I first met Andrew in 2009 when he was building SeaMicro. Even back then, he stood out. He was a force of nature. Extremely intense. Deeply technical. A builder in the purest sense. You could feel the sparks flying off him. After Andrew joined AMD, we stayed in touch. Over the next year or so, we met 6-7 times. Sometimes in our office. Sometimes at a coffee shop in Portola Valley. Sometimes at our local tennis and swim club. We were talking, riffing on ideas and exploring what might be interesting or broken. Conversation after conversation, we kept coming back to one thing: deep learning workloads were growing exponentially and traditional compute architectures and data center models were buckling under that pressure. By the time Andrew and his all-star co-founders Sean Lie, Gary Lauterbach, Jean-Philippe Fricker and Michael James were ready to go all-in, we already had a deep belief in them. In fact, I asked Andrew if we could be his first term sheet. Shortly after our investment, @ericvishria from Benchmark, our co-lead in Cerebras’ first round joked with me “you did 2 years of work on Cerebras, I did 4 days.” A testament to Eric’s conviction in the opportunity. Fast forward 10 years and it has been an unbelievable journey. I’m both incredibly proud to have been the first partner in Cerebras and to call Andrew my friend. Watching the team overcome adversity after adversity to shape the most impactful technology of our lifetime has been a defining part of my last decade. And I still can’t wait for everything that is still to come. Keep at it my friend.

Introducing Daydream — a desktop AI video editor that you can chat with. Create videos like this one all with prompts. No prior editing skill required. No need to upload GBs of video to the cloud. Check it out at daydreamvideo.com and follow @daydream_video for more.

In 2023, I learned one important lesson at #burningman: if a diligence meeting matters enough, you will find a way out of a flooded playa on a Sunday. Covered in dust, mud, and questionable decisions, still made the call. That call became our seed investment in Fulcrum. Conviction > conditions 😅🚀 Congrats team Fulcrum. @JayaGup10 and I and the entire team at @FoundationCap are so proud of you. foundationcapital.com/building-the-c…

@akoratana @VirtusaCorp Huge milestone, Animesh! Great to see you tackling the real operational pain around running and supporting software in production with a partner at Virtusa’s scale.

Very excited to announce our partnership with @VirtusaCorp! So much of the conversation is about writing code, but it’s easy to forget the people who have to operate and support that software once it leaves the IDE. There’s an asymmetric pain here that the market still has to work through before the future we’re imagining really shows up. In most enterprises, the knowledge that matters is scattered across tickets, repos, dashboards, and people’s heads. Building agents (and the context graphs they run on) that work in those environments is hard, and it’s the work that'll make the difference. Virtusa is an incredibly thoughtful partner with a rare ability to see where things are going and bring the future into the messy reality of today for companies around the world. Super excited to partner and bring AI production engineering to a lot more teams!

Today we’re launching Automations - automatic onramps that turn ACH & wires into stablecoins on 20+ blockchains. Unique account & routing per customer. Fiat in, stablecoins out. Starting at $0.25/month and 0 bps. 🧵👇

AI has given venture capital a new way to repeat an old mistake: kingmaking. The pattern from 2021 is back: a category becomes "obvious," a top-tier firm anoints its winner, and everyone else acts like the decision is final. Sierra for support. Harvey for legal. Applied Compute for RL-as-a-service.

So excited to launch Autograph today. We raised $2.1M from elite operators, founders, and investors to build a platform that feels like God Mode for Operators. It's a single platform that integrates with existing systems — the alphabet soup of enterprise software, like your HRIS, ERP, FP&A, ATS, CRM, along with data warehouses and spreadsheets — to combine, clean, and make sense of your data. This data feeds a series of pre-built apps and agents that allow you to level up as an operator: run your complex, multi-functional workflows 10x better. We're solving the processes that takes forever and is really frustrating, but is essential to operating and scaling effectively: our first few apps are headcount planning, compensation reviews, and sales capacity. We're moving insanely fast. We're shipping a new app every couple of weeks. Soon, we'll have hiring pipelines, vendor analytics, resource allocation, unit economics, CX capacity, people analytics, and more. The best part? It's incredibly fast to set up thanks to our powerful agentic data engine, that processes any data source like magic. That means no long implementations, no multi-year contracts. We think it's the biggest leap forward for operators since the spreadsheet. Book an 20-minute onboarding to get started!

It is intuitively easy to understand why self play *can* work for LLMs, if we are able to provide a value function at intermediate steps (although not as clearly guaranteed as in two-player zero-sum games). In chess / go / poker, we have a reward associated with every next move, but as Noam points out, natural language is messy. It is hard to define a value function at intermediate steps like tokens. As a result, in usual reinforcement learning (like RLVR), LLMs get a reward at the end. They end up learning to 'meander' more for hard problems. In a way, we reward brute forcing with more tokens to end up at the right answer as the right approach. However, at @DeepCogito, we provide a signal for the thinking process itself. Conceptually, you can imagine this as post-hoc assigning a reward to better search trajectories. This teaches the model to develop a stronger intuition for 'how to search' while reasoning. In practice, the model ends up with significantly shorter reasoning chains for harder problems in a reasoning mode. Somewhat surprisingly, it also ends up being better in a non-thinking mode. One way to think about it is that since the model knows how to search better, it 'picks' the most likely trajectory better in the non-thinking mode.

Congrats to @JasonDemant and the @FoundationCap team for launching their first batch of Fintech AI companies in SF today — great to be here and see so many friends

🚀 We’ve raised a $9M seed to power Maximor – the AI-powered finance team that fixes the back-end grind. 💡 Finance is broken. And the solution isn’t another ERP. A quick 🧵

🚀 Excited for tomorrow’s Fintech AI Build Demo Day: foundationcapital.com/fintechaibuild/ Eleven fintech + AI companies presenting. A couple of last-minute investor spots are open—if someone should join, tag a fintech investor or DM for details. #Fintech #AI #DemoDay

@awwstn 🚀🚀🚀

Meet Howie, the people's secretary. Out now. (see thread for signup link + fundraise announcement)

Interested in the intersection of RL and coding? (not RL environments) Ping @akoratana or I for an exclusive invite to an event Wednesday!