Pete Soderling @petesoder

3 months ago, an anonymous account commented "Y'all are sleeping on the most important talk of this conference". The speaker was @CompleteSkeptic and his talk was on RLHF's deal with the devil, and the argument behind Jev. Diogo argued that training AI to produce answers people prefer comes with a trade-off. You get polished, coherent responses. You lose the unusual alternatives through mode collapse. The problem is that the unusual alternative can be the correct one. Three months later, @typesafeai's Jev, a model designed for decisions, is now generally available to everyone. If Jev has caught your attention over the weekend (the internet is on fire with it after all!), Diogo’s @AICouncilConf talk is worth watching. It lays out the argument behind what his team is building and why they chose this direction. Diogo's talk: youtube.com/watch?v=o-y1HJ…

Very timely that we are organising a RSI hackathon @UseCorgi Cafe London on 19th Sep. Register here: luma.com/abmar7tg 📷 Corgi Cafe, London · 19 Sept 12 hours. 35 hackathon spots. $1,500 cash + $13,500 in @runware credits.

In January, humans still ran most of the queries on @motherduck. Last month, agents ran 29× more than people did. MotherDuck looked at their own query history (UI = human, MCP = agent) and discovered - perhaps not surprisingly - that agents are now the dominant consumers of the data platform. The average org has twice as many agents as human users. The workload shape is different too: ➔ Median gap between agent queries: 4 seconds. Humans: 60. ➔ Median agent query scans ~7,000 rows. Humans: ~200,000. ➔ Agents fire bursty, small, experimental queries then go quiet. Every assumption baked into warehouse design (latency budgets, concurrency, pricing, warm caches) was made for a user who thinks for a minute between queries. The new user thinks for four seconds and then probes with ten more small queries and loops again. It isn't just MotherDuck who is seeing agent activity skyrocket. OpenAI published a report this week showing a huge climb in internal coding agent usage over the past few months. Cursor cloud agents went from ~10% of merged PRs in December to 56% by July, and Cursor now says it’s more than 60%. Internal agents are exploding and the primary user of data infra has flipped. MotherDuck's report is a useful view into what that feels like for a data platform: motherduck.com/blog/agents-do…

keyboard_arrow_left Previous keyboard_arrow_right Next

I’m hiring a Chief of Staff to help run @ZeroPrimeVC & @AICouncilConf. It's a mini-COO role for an unusually capable generalist. Most of your time will go to operating and growing AI Council. The rest goes to increasing my leverage across both AI Council and Zero Prime: turning priorities into action, preparing decisions, protecting focus, and making sure important threads don't drop. The right person is a high-agency generalist - comfortable owning a P&L, fully AI-pilled in how you work, and low ego about the job. You should move comfortably between commercial work, marketing, events, finance, executive communication, and operational detail. The scope is intentionally broad and expected to grow. The right person will earn ownership quickly and can develop toward broader operating leadership over time. Evidence of ownership matters more to me than a matched resume. $150–180K base, bonus, carry eligibility. Full time, in person, San Francisco. To apply, send resume and LinkedIn profile to [email protected], with short answers to two questions: 1. What's the most ambiguous cross-functional problem you've personally owned through execution? 2. Why does building both an AI community business and a day-zero venture platform appeal to you? Specific examples beat a polished cover letter. If a name came to mind while reading this, please tag them. A repost also helps this find them.

.@AgnoHQ just showed you can pay a frontier model to solve a problem, then let a far cheaper model reuse what it learned and nearly match its score. ARC-AGI-3 drops an agent into a game it has never seen, without instructions or goal. The agent has to experiment its way towards understanding the game, and it's scored on how efficiently it learns. GPT-5.6 + Agno's Learning Machines scored 100 on the public set (183 levels across 25 games). As it played, it saved what it learned into a per-game manual. Then Agno handed those manuals to Gemini-3.7-Flash, which had scored 37.33 starting cold. With the manuals it scored 96.42, above the 95.4 human baseline. And because nothing was retrained and no weights changed, the whole gain came from the smaller model reading what the bigger one had learned. The learning transfer made the much cheaper, small model viable for a task you'd normally trust only to the frontier. Significant, because agents burn a lot of tokens rediscovering lessons an earlier run already paid for. Pay frontier prices for discovery once, and the repeat work runs on cheaper models. Full write-up: agno.com/articles/learn…

Building a company in Europe is hard mode. Which is why we spend a lot of time across the pond hunting for startups! By the time a European founder has product in front of customers, they've already cleared obstacles a Delaware C-corp founder never meets. 2 of my early European deals were co-investments alongside @YTR4N_, then at the European seed fund Speedinvest. Yang built his career in Europe, and when I asked him to join me as partner at @ZeroPrimeVC, his European roots were part of the plan. Some of the European companies we've backed: ➔ @cusp_ai (Cambridge): AI-powered materials breakthroughs ➔ @pydantic (London): the end-to-end AI engineering stack ➔ @runware (London): screaming fast inference for video & image gen ➔ DataLinks (Zurich): your data, ready for agents in minutes ➔ @sodadata (Brussels): data contracts at scale ➔ QontextAI (Berlin): a brain for the AI-native company And our conference @AICouncilConf puts European teams on stage every year, like: @duckdb (Amsterdam), @UnikraftCloud (Heidelberg), LogicStar AI (Zurich), @Doubleword_ (London), Datalinks (Zurich, from our own portfolio). Then there's companies like n8n (Berlin), @Spotify (Stockholm), Mistral (Paris), @Lovable (Stockholm), Klarna (Stockholm), and DeepMind (London, now part of Google). We'll keep making the trip.

OpenAI measured 17M messages inside 1,500+ companies. Per active user, analysts out-message every job title class, engineers included. So the most intense AI use in the enterprise is landing on the work @isidoremiller calls "the cursed domain": "Easy questions look hard. Hard questions look easy. Many questions are impossible to answer; to even try is to fail. Bugs are usually silent and subtle. Innocuous assumptions (LLM’s favorite!) make or break analyses. There are no linters, no test suite, no formalization language. There is almost no realistic public data to train on or build environments from, and there is a surplus of unrealistic tutorial-slop jamming up the pretrain. Everyone’s data warehouse is out of distribution. For every right answer, there are ten plausible but subtly incorrect wrong answers, and no way to verify or validate the result." TL;DR: analytics is a hard domain, fraught with booby traps inviting confidently incorrect answers, and public benchmarks mostly test text-to-SQL on clean, self-contained datasets. Izzy's team at @_hex_tech published a new benchmark called DataBench. DataBench runs 100 analytical tasks inside a fake $129M company with 6 years of deliberately broken data, built around the conditions that make analytics hard in practice. A few highlights: > Opus 5 at high effort caught every one of the 10 signature traps. > The traps are what separate the models. Scores there run from 100% down to 30%, against a much tighter 91% to 61% range on straightforward Q&A. > Fable 5 Max leads on Q&A at 91% and is the only model that keeps improving as you spend more. If you operate in analytics, the report is worth reading. Hex: hex.tech/blog/databench… OpenAI: cdn.openai.com/pdf/how-organi…

Running a technical conference is a strange thing for a venture fund to do. Every year, we sell sponsorships, tickets and welcome a mass of attendees and AI/infra thought leaders for a week's worth of events where mistakes are very visible. The ops are pUnIsHiNg. For over a decade I've built the community behind @AICouncilConf, formerly Data Council, and it remains one of our strongest sources of deal flow for @ZeroPrimeVC. Here's the recipe: speakers turn into -> founders -> and then we back them. We invest at the earliest stage in companies like Modal, Runware, Hex, Pydantic and Higgsfield. It's time for the next stage of growth - so I'm hiring a Chief of Staff in San Francisco to help me run it all. → Own the AI Council operating plan and the numbers under it. Sponsorship and ticket sales, marketing, programming, event production, budgets, P&L. → Direct the teammates and partners who already run those functions, our event agency included. You won't be building it solo. → Take some off my plate across both businesses, which mostly means turning priorities into action and keeping threads from dropping. → Occasional special projects on the Zero Prime side, usually portfolio support or fund ops. You'll build a sponsorship forecast in the morning, then spend the afternoon figuring out whether the boat for the evening happy hour community cruise has power. It's full-time and in person in SF. You should have an MBA and want to put it to use. VC, PE, consulting, or fast-growing startup experience helps. Send me an email w/ the most ambiguous cross-functional problem you've personally owned through execution, and why running a community business and a day-zero fund at once appeals to you. Know someone you'd bet on? Send them my way. Details here: zeroprime.notion.site/Chief-of-Staff…

keyboard_arrow_left Previous keyboard_arrow_right Next

Higgsfield: jobs.ashbyhq.com/higgsfieldai/8… jobs.ashbyhq.com/higgsfieldai/8… jobs.ashbyhq.com/higgsfieldai/8… Modal: jobs.ashbyhq.com/modal/e1915603… Cusp AI: jobs.ashbyhq.com/cuspai/916e148… StarTree: job-boards.greenhouse.io/startree/jobs/… job-boards.greenhouse.io/startree/jobs/… Lightning AI: job-boards.greenhouse.io/lightningai/jo… job-boards.greenhouse.io/lightningai/jo… job-boards.greenhouse.io/lightningai/jo… Soda: apply.workable.com/soda-data-nv/j… apply.workable.com/soda-data-nv/j… Runware: apply.workable.com/j/771442E40D MotherDuck: jobs.ashbyhq.com/motherduck/d81… jobs.ashbyhq.com/motherduck/2e2… LanceDB: jobs.ashbyhq.com/lancedb/d86e02… Hightouch: job-boards.greenhouse.io/hightouch/jobs…

Higgsfield, Modal, Lightning AI, MotherDuck, and 6 other @ZeroPrimeVC portcos are actively hiring for technical roles across AI/infra right now. Here's a list of the freshest technical roles from our portfolio: @higgsfield_ai | AI creative studio for video & image generation ➔ ML Engineer, Data Engine (Kazakhstan) ➔ ML Systems Performance Engineer, MFU (Kazakhstan) ➔ ML Engineer, Ads (SF Bay Area) @modal | serverless compute for AI ➔ Detection and Response Engineer (NYC) @cusp_ai | AI for materials discovery ➔ Applied AI/ML Engineer, Agents (Amsterdam) @startreedata | real-time analytics database ➔ Software Engineer, SRE (India) ➔ Senior Software Engineer, Data Platform (India, remote) @LightningAI | full-stack AI infrastructure platform ➔ Data Scientist (NYC) ➔ Senior Research Engineer, LLM Training & Post-Training (NYC / SF / Seattle / remote) ➔ Senior Software Engineer, Agents (London / NYC / SF) @sodadata | AI-native, fully automated data quality platform ➔ Senior Backend + Infrastructure Engineer, Platform (Europe, remote) ➔ Senior Full-Stack Engineer, Backend + Frontend (Europe, remote) @runware | unified AI inference API ➔ Engineering Manager (London, remote) @motherduck | data warehouse for agents powered by DuckDB ➔ Customer Engineer (NYC) ➔ Software Engineer, Database (Amsterdam / remote) @lancedb | multimodal lakehouse for AI ➔ Senior Product Security Engineer (Americas or APAC) @HightouchAI | customer data and AI platform for marketers ➔ Forward Deployed Analytics Engineer (North America, remote) Direct links to every role in the comments.

The ducks are doing deals. @motherduck acquired Tower.dev, and today Amazon announced it's acquiring Duck Labs, the team behind the @duckdb engine MotherDuck runs on. MotherDuck is a @ZeroPrimeVC portco, so I have ducks in this race. Many of the data infra pitches I read still assume the user is an engineer at a keyboard w/ a connector catalog on one side and a workflow UI on the other. That assumption is aging quickly. A model writes the pipeline on demand now, and the harder problem has become giving that code a sandboxed place to run on a schedule, with enough control and visibility for a human to review what happened. The runtime became the product. Earlier this year @jrdntgn predicted that data engineering becomes an agent supervision problem, agents writing and running the pipelines while engineers review the work. MotherDuck had gone shopping for a connector tool before changing its mind about the category. "We didn't need canned connectors; Claude can write the code to move data from one place to another." Flights, MotherDuck's feature that hosts and schedules those LLM-written pipelines, was built on Tower's runtime and shipped in a matter of weeks. There’s a good backstory here too. Jordan has known Tower founder @datancoffee since their Google days. He was disappointed when Serhii teamed up with a Snowflake colleague, Brad Heller, to start Tower, since it meant losing his shot at hiring him for MotherDuck. Looks like Jordan found a workaround. Then there’s Amazon. Under the deal, the DuckDB Foundation keeps control of the DuckDB IP and the license remains permissively open. Jordan wrote today that Amazon picking up an open source project brings "distribution, awareness, and standardization." If DuckDB becomes more deeply established as a standard engine, there’s less room to differentiate on the engine itself. The interesting product surface moves above the engine, to where workloads run and how humans supervise the agents doing the work. MotherDuck just bought into that layer. Jordan's sign-off: "MotherDuck, true to our name, is both proud of the moment and a little bit nervous about what they'll get up to next." Congrats to Jordan Tigani, @ryguyrg and the MotherDuck team, and to Serhii Sokolenko and Brad Heller. And to my friend Hannes Mühleisen, who created DuckDB with @mraasveldt, congrats on the start of DuckDB's next chapter.

keyboard_arrow_left Previous keyboard_arrow_right Next

True or False? Are we reading the code or not reading the code?

Can computer-use agents use complicated graphical software? You’re watching GPT-5.6 Sol operate Autodesk Fusion for 250 turns as it attempts a real mechanical-design task. The run is part of a benchmark aimed at the question "how good are agents at CAD?" @seldon_tech's CADBench contains 105 mechanical-design tasks tested across 10 frontier models, and the results show how early reliable, autonomous CAD work remains: More than two-thirds of the tasks are still unsolved. That gap matters as agents make their way into the professional tools used to design the physical world. Getting around Autodesk Fusion is one thing, but producing a correct, editable model an engineer can pick up and use is much harder. CADBench gives us a way to see and measure that frontier. Great work from the Seldon team! Explore CADBench: seldon.global/blog/cadbench

.@higgsfield_ai's @alexmashrabov on @CNBC. "The opportunity goes beyond making content faster. It’s about helping businesses create hyper-personalized content, test more ideas, and ultimately drive better business outcomes." Test more ideas = create a faster learning loop Maximizing the learning loop is baked into the Higgsfield's product, and a core part of their company DNA. $20M to $700M annualized revenue ramp speaks for itself. Useful lesson in there! linkedin.com/posts/amashrab…

.@higgsfield_ai's first product turned a selfie into a video. Today Higgsfield raised a $400M Series B at a $5.4B valuation, led by DST Global. Annualized revenue is $700M, up from about $20M a year ago. In January, business customers were under a quarter of revenue. Today enterprise dominates with 390 of the Fortune 500 running visual production through Higgsfield. @ZeroPrimeVC wrote a check into the $8M seed in 2024 and, looking back on my earlier conversations with @alexmashrabov, what’s remarkable is how closely the company’s trajectory has followed the plan Alex described back then: “We started with consumer use cases to be first in the market and learn what users want, because the largest risk for any startup is building something nobody needs.” Alex was already thinking about what came next: “We’re going to start partnering with enterprise teams and make the model part of their workflows.” Underneath that strategy was another advantage: Higgsfield’s speed of experimentation. The team built its own training framework so researchers could run experiments without getting slowed down by MLOps and GPU orchestration. Alex said they had achieved a speed of experimentation he had never seen at Snap or Yandex, even while managing clusters of up to 1,000 GPUs. The product evolved, but the thesis was there from the start: learn faster than anyone else, then bring that advantage to the enterprise. Congratulations to Alex, Yerzat, and the entire Higgsfield team!

ByteDance's ArkClaw agents fork their memory like code and merge the useful branches back in. They call it 'GitforMemory'. It's built on the branching API in Lance, the open storage format underneath @lancedb. An agent snapshots its memory state and runs down one direction in isolation. The branch either merges into main or gets thrown away. Their support teams can also use the same mechanism to keep sensitive operational memory in its own branch, away from the public knowledge base. The 'git for memory' idea is not new, but instead of versioning a markdown file, ByteDance's 'GitforMemory' versions the entire Lance table the agent searches, indexes and all. The result is versioned, enterprise-grade agent memory that can handle 100,000+ QPS, built on the open lakehouse format Lance. @ZeroPrimeVC is an investor in LanceDB, so weight the enthusiasm accordingly, but I thought this was cool!

.@ZeroPrimeVC is one of 101 GPs invited to the 11th Annual @raiseconference Summit alongside 330+ LPs. We're honored to be included in this year's cohort, and I'm excited to present our perspective on where AI and data infrastructure are headed. Looking forward to connecting with everyone in October. events.raiseglobal.co/event/2026Glob…

Between the OpenAI sandbox escape, rogue Anthropic models, and the return of Shai-Hulud, cyber security seems stuck in an infinite loop of bad headlines. Some optimism from @feross, CEO of @SocketSecurity: "There is not an infinite number of software vulnerabilities." "Every vulnerability that is discovered does get patched, and then it can never be exploited again. That is permanent progress." "Yes, new code is written and there will be new bugs introduced, but I think there's this compounding effect where the models that write exploits can also review pull requests, flag unsafe patterns, catch bugs." "I think as they integrate into more and more development workflows, we're we're not just going to find these old bugs, we're going to write fewer new ones. And so I think it's going to converge into a much better place." "And I think for the first time defenders have this infinitely scalable army of AI agents that can do continuous security analysis." "And they can do work that would have been too expensive or too impractical for humans to do before." "So the trajectory is toward a fundamentally more secure software world than than we've ever had before." From his talk at @AICouncilConf '26: "The Agent Attack Surface: Why AI Is Breaking Software Security As We Know It" youtube.com/watch?v=XnP3td…

.@andy_pavlo, one of the research rockstars of modern databases, joined @ClickHouseDB this week. Seeing his name on the front page of Hacker News reminded me of a story: A decade ago when I was running DataEngConf in NY, Andy gave a talk called "Peloton: The Self-Driving Database Management System." It previewed a paper his CMU research group would publish a few months later, in January 2017. The pitch sounded crazy at the time. A database that could configure, tune, and optimize itself the way a human DBA would, with people only making the calls a machine can't. The system Andy sketched, a neural net forecasting what a live database will need and a planner deciding what to do about it, sounds a lot like what we now call an AI agent today. He was building one to manage a database before the transformer paper existed. One of the research questions Andy lists for ClickHouse Labs is "determining how agents can improve and automate the development of DBMSs themselves." Same question he was asking on our stage a decade ago. Congrats Andy. Lucky ClickHouse! Andy's 2016 talk: youtube.com/watch?v=mzMnyY…