Jon Lerner

Thrilled to announce that Together Fund is investing in Sentra, alongside a16z speedrun! You track results in Jira. Decisions in Notion. Conversations in Slack. But the reasoning, the debates, trade-offs, and context behind why you chose A over B, disappears into what we call "Dark Matter." A decision made in March looks insane by July because no one remembers the constraints that made it smart. I lived this firsthand at Twitter scaling from 800 to 8,000 employees, and at Google while launching AI Overviews to billions at planet scale. The problem isn't process. Process is compensation for something deeper: organizational amnesia. An organization’s "Systems of Record" doesn’t solve this, they encode it. They store what happened, never why. That's why we are investing in Sentra. Sentra is the always-on collective memory that eliminates organizational amnesia by maintaining accurate context for all members and agents, functioning as an operational nervous system. It connects to every channel where work happens, meetings, Slack, email, code commits, docs, calendars, and treats them not as artifacts to search, but as living signals to synthesize. The fleeting and the permanent, unified into a memory that understands. The founding team is built for this: - Jae Gwan Park (CEO): Product-first founder, memory systems research at UofT and MIT - Ashwin Gopinath (CSO): Former MIT professor, created "Reflexion" (NeurIPS 2023), agents that learn from mistakes, 2x founder - Andrey Starenky (CTO): Early Vapi engineer, ex-IBM, built to process enterprise-scale data firehose Together is an operator-led fund. We invest in problems we've lived. This is one of them. Many congrats Jae, Ashwin and Andrey, we are so excited to partner with you! Read the full thesis: https://lnkd.in/gixj9cE4 Book a demo: https://www.sentra.app/ #OrganizationalMemory #AI #Sentra #TogetherFund #a16z #ContextGraphs

71

3 Comments

CRV Security: Request for Startups I never know if this actually works for our friends over at YC but figured we'd try. Here's what we want to fund in 2026! 1. Golden Artifacts: Think Chainguard but more broad. Artifact attestation exists for open source. Almost nothing exists for internal software — especially the vibe-coded tooling now running in production. We want the company building cryptographic proof of secure software delivered from secure artifacts: who built it, how, and whether it was reviewed. If more things are being yeeted into the world via Claude Code (myself included), this feels like an issue. 2. MCP & Agentic Security: Agents are getting real credentials and taking real actions. The security posture of most orgs around this is basically zero. That changes fast. You'd never give an employee hardcoded API keys or write access to your email without supervision/trust. Why give it to agents? 3. AI Governance: Boards are asking CISOs to account for AI risk. CISOs have no good answer other than "Palo has a module" 4. Next-Gen Endpoint: CrowdStrike was built for a world of static binaries and human operators. AI workloads, cloud-native infra, and AI-assisted attackers need a new architecture. The category is ready to be reinvented. 5. Networking in the AI Era: Zero trust was designed for humans. What does network security look like when the entity requesting access is an agent? Nobody's really solved this. 6. Email Security + Next-Gen Phishing: LLMs have made spear phishing infinitely scalable. I've never truly understood why Abnormal and KnowBe4 aren't one company. Maybe this time it's different. 7. Frontier Security Lab: We'd back a credible, well-staffed lab focused entirely on red-teaming models and setting the evidentiary standard the industry needs as LLM built apps become the norm. 8. Dependency Security: That Actually Remediates Malicious and vulnerable dependencies are a top attack vector. The tooling is mostly noise — scanners that don't close the loop. The winner here ships fixes, not just alerts. 9. Critical Infrastructure Cyber: Data centers, satellites, power grids, undersea cables. The physical backbone of the internet is increasingly exposed and wildly under-defended. We have data centers in space, for God's sake. Surely we need better cyber for critical infrastructure? 10. PAM for the Modern Era Legacy: PAM was built for static roles, human users, on-prem directories. Cyberark was founded in 1999.....Agents, ephemeral workloads, and cloud-native infra have broken all of those assumptions. Is anyone rebuilding this from scratch? If you're building in any of these areas — or something we haven't thought of — reach out. james@crv.com

445

48 Comments

🚀 New partnership: Mento VC × Modal We teamed up with Modal so founders can spend less time on infra and more time shipping! Serverless compute for AI inference, fine-tuning, and batch - when you need it, at the scale you need. Founders: what’s the toughest part of your AI infra today - GPU availability, cold starts, or cost control? Comment below; we’ll share what’s working across teams we back. Transparency note: Mento portfolio companies can redeem $5,000 in Modal credits.

19

2 Comments

SF AI-Native Operator Takeaway #2: In AI-native PLG, the hard part isn’t conversion... it’s discovery. Many AI-native teams are still talking about PLG using a classic SaaS mental model, but based on operator conversations in SF, that model is starting to break down in fairly obvious ways. The biggest bottleneck right now isn’t conversion. It’s discovery. In traditional PLG, users generally understood the category before they ever signed up. The problem was obvious, the product’s value was legible from the homepage, and the “aha” moment tended to show up quickly in first use. In that world, PLG meant optimizing onboarding, reducing friction, and improving free-to-paid conversion because user intent already existed. AI changes that assumption. In AI-native products, users are often curious but unclear. They don’t yet know what’s possible, value depends heavily on workflow, context, data, and role, and the product can feel abstract until it’s applied directly to their job. As a result, many users stall not because the product isn’t valuable, but because they haven’t discovered how it fits into their world and how they can't live without it. This is the real distinction people kept coming back to. PLG conversion answers, “Is this worth paying for?” PLG discovery answers, “What problem does this solve for me, right now?” What’s working best in practice is less about funnel polish and more about clarity up front: role- or workflow-specific entry points, guided examples instead of blank states, and opinionated first actions that show users a concrete outcome before asking them to explore. This also explains a broader pattern across AI-native companies. Forward-deployed teams and services-heavy delivery aren’t just implementation tools; they’re discovery mechanisms. They translate abstract AI capability into concrete workflow value, observe real use cases users wouldn’t self-discover, and feed those learnings back into what eventually becomes productized. PLG isn’t going away, but in AI-native companies it’s being redefined. Self-serve no longer means self-explanatory. Education becomes part of the product, and discovery has to come before optimization. The teams making progress aren’t obsessing over conversion rates yet. They’re focused on whether users see themselves in the product, how quickly they reach a meaningful outcome, and whether the product helps users get to a meaningful outcome for themselves quickly, without too much guesswork. Bottom line: in AI, PLG is less about removing conversion friction early and much more about creating understanding first. Once they understand, they may be hooked. Tomorrow is my last SF AI operator takeaway focusing on everyone's favorite topic du jour: 996 work schedules.

15

1 Comment

Deal of the Week https://lnkd.in/dbwycUFN Augmented Intelligence Inc. — $20M at $750M valuation https://aui.io Investors: eGateway Ventures (lead), New Era Capital investor group, strategic partners AUI's Stealth Play for the Post-Transformer Era AUI continues to operate under deep stealth while charting a bold path beyond the transformer paradigm — a contrarian stance in an AI landscape dominated by compute-heavy scaling wars led by OpenAI, Anthropic, and Google DeepMind. The company's recently closed $20M bridge SAFE at a $750M valuation cap underscores deep investor conviction that the next wave of AI breakthroughs won't come from bigger GPUs, but from smarter architectures. While the broader market remains fixated on model size and synthetic data pipelines, AUI's research focus lies in neuro-symbolic reasoning, blending the statistical power of deep learning with the logical rigor of symbolic systems. This hybrid approach promises greater reasoning reliability, interpretability, and energy efficiency — three attributes that address growing skepticism around current foundation models' opacity and cost of scaling. The raise brings AUI's total capital to nearly $60M, fueling the quiet development of its Apollo-1 reasoning engine, rumored to be a foundation-model architecture that fuses symbolic logic modules with dynamic neural representations. Insiders describe Apollo-1 as a potential step-change in "grounded AI" — one capable of multi-hop reasoning and factual traceability without massive compute budgets. Within New York's fast-emerging AI corridor, AUI is becoming one of the most closely watched under-the-radar companies, blending research sophistication with enterprise-grade product discipline. Its investor syndicate — composed largely of deep-tech and frontier-AI specialists — sees this as an early bet on the post-transformer paradigm: leaner, more explainable AI that could redefine how intelligence is modeled, trained, and deployed. Deal Significance AUI's bridge round reflects a broader inflection point in AI investing — where capital is shifting from pure scale bets to architecture differentiation and reasoning fidelity. The company represents a hedge against the saturation of transformer economics, where model improvements deliver diminishing returns on exponentially higher costs. If Apollo-1 delivers on its promise of interpretable, low-energy reasoning, it could position AUI as a strategic acquisition target for hyperscalers seeking new compute-efficient architectures, or as the foundation for a new class of hybrid AI platforms capable of combining symbolic logic with generative fluency. For investors, this deal underscores rising conviction that the next "OpenAI-moment" may emerge not from scale, but from structure — and AUI's stealth trajectory gives it the mystique and momentum to be one of the few serious contenders for that title.

4

Terminal Use (YC W26): Vercel for background agents. Terminal Use is hosting infrastructure for background agents that makes harness engineering easy. https://lnkd.in/gv99wN2t

51

2 Comments

Why does this matter? Take a common “stablecoin sandwich 🥪 ” scenario: funds are on-ramped via a provider that supports USDT on Tron, and off-ramped through another that supports USDC on Polygon. Until now, teams had to rely on external DEXs or exchanges to bridge that gap — adding operational complexity, delay, and risk. With this new embedded capability in Utila, that friction is gone. Teams can now seamlessly swap and manage assets across chains 𝐰𝐢𝐭𝐡𝐢𝐧 their core ops stack.

19

With RL techniques like GRPO, it’s become easy to fine-tune small models to outperform foundation models on vertical tasks. We’re open sourcing Osmosis-Apply-1.7B: a small model that merges code (similar to Cursor’s instant apply) better than foundation models. Links in comments! Using foundation models for high specificity, low complexity tasks like code merging can be overkill - instead, it’s better to use a specialized model instead for higher performance and latency. We fine-tuned the model on 100K real-world code commits where the model needs to merge a code edit snippet into an original code snippet. On the test dataset (subset of CommitPackFT), foundation model performance ranged between 0.77-0.93 reward score. In comparison, Osmosis-Apply-1.7B achieved a 0.98 reward score - while also being 3X-5X more affordable than the next cheaper model AND ~10X faster! Why this matters: the best AI agents use multiple models to ensure task specialization and reliability at scale - on top of cost and latency benefits as well. Links to the model and full write-up (with reward function, hyperparameters, etc.) are in the comments - reach out if you’re interested in using reinforcement learning!

240

18 Comments

Supabase's Ant Wilson and Paul Copplestone raise $200m seven months after last raise. TC talks about the growth of vibe coding. https://lnkd.in/gaNYvyRC

53

How Replit Agent 3 Pushes Engineers from “Coding” to “Supervising Systems” Replit’s Agent 3 is a real step toward autonomous software delivery - not just “AI-assisted coding.” Here’s what actually changed and why it matters. What it does (beyond code suggestions) * Self-testing in a real browser: Periodically spins up a browser, drives the UI (clicks, forms, auth), summarizes findings, and patches regressions in-loop. You can watch the cursor in the Agent pane as it validates buttons, forms, APIs, data sources, and Replit Auth flows. * Autonomy with guardrails: Replit reports 10× more autonomy than v2 and a testing system 3× faster and 10× more cost-effective than generic “computer use” models. Tests trigger when enough has changed - not after every message * Max Autonomy (beta): Runs for ~200 minutes with self-supervision and progress tracking - useful for multi-step refactors or feature builds that need uninterrupted execution * Builds other agents & automations: From Slack/Telegram bots to scheduled workflows (e.g., daily Linear/Notion summaries) via a guided credentials flow (Notion, Linear, Dropbox, SharePoint). Turns “glue work” into natural-language recipes the agent implements Why it matters for teams * Shorter feedback loops: A tight build → run → test → fix loop reduces context switching and catches UI/API integration issues earlier * Reduced integration overhead (credentials, SDKs, OAuth) frees engineers to focus on specs and review * Supervisory posture: Engineers shift to specifying intent, constraints, and acceptance criteria - less line-editing Operational considerations * Production safeguards: Keep backups/versioned deploys; gate production mutations until evaluations pass * Stop conditions & scope: Define time, scope, and blast-radius limits for Max Autonomy sessions. Require approvals for schema/migration steps; prefer data migrations behind feature flags * Traceability: Treat the agent like a junior dev with root - mandate PRs, CI, signed commits and human review on sensitive repos Bottom line Agent 3’s browser-level self-testing, multi-hour autonomy, and built-in automations make it a credible candidate for supervised autonomy across parts of the SDLC. Teams that write crisp specs and enforce guardrails can compress iteration cycles - while keeping humans accountable for outcomes. https://lnkd.in/gravaSWA #ai #artificialintelligence #aiagents #coding

8

AI is reshaping every industry — and VC is no exception. great catching up with Nathaniel Barling, operating partner at Andreessen Horowitz, who leads a 10-person engineering team building internal AI tools to supercharge their VC workflows. 🚀 at Creandum, we’re right there on the same journey. Over the past year, we’ve rolled out multiple AI-enabled apps (some built with Lovable ofc 😎) that now power everything from sourcing + dealflow analysis to portfolio management. the productivity impact is real — and we’re only just getting started. in VC (like everywhere else): adapt and embrace…or get run over. ⚡️

204

11 Comments

The big labs are expansive in their product ambition, especially since foundation models have largely improved in lockstep - in order to compete with them you have to do things they won’t which are: - building a very rich software ecosystem around a primitive - orchestration across multiple models - going insanely deep on product and growth for a narrow vertical domain

58

3 Comments

We've built the best Computer Use Agent (CUA) for QA testing! In our latest benchmark, our agent outperformed OpenAI's Operator, Claude 3.7 Thinking, and Gemini 2.0 Flash in accuracy & robustness, based on real customer E2E testing in staging environments. See the screenshot for the results. Let's dive into our methodology: <Evaluation Criteria> To begin, we started by defining clear criteria based on what QA teams care about. 1. Robustness: Does the agent follow the user’s instructions and test all specified scenarios? For example, if instructed to log in, search for an item and buy it - the agent must verify all these steps. 2. Accuracy: Does the agent accurately detect bugs in user flows without any false positives or negatives? This goes beyond just completing the test. The agent must also correctly identify each test scenario as FAIL/PASS according to the known issues on staging. < Creating Realistic Test Prompts > To create a meaningful test set, we used prompts used by 10 customers on our platform, ranging from startups to Fortune 500 companies, to either recreate existing end-to-end tests from their QA suites or write tests for new scenarios they want to have tests for. The scope of the tests includes ~30 user flows and ~250 steps that cover critical functionalities like account creation, authentication, checkout, content discovery, chat, form submissions, access control, and search. For each of the tests, we created an answer key detailing expected outcomes, clearly identifying steps with or without known bugs. < Evaluation Methodology > To build the evaluation dataset, we executed each test across real customer staging environments through each of the providers twice, and took the best scoring run across the metrics. For the evaluations of Claude 3.7 Thinking, we conducted the runs through both a vision-only computer use framework and a DOM-based computer use framework. For Gemini 2.0 Flash, we conducted both runs on a DOM-based computer use framework because there’s no available packaged product on the market to use it in a vision-only framework. At the end of each test run, we asked the agents to output the test cases it validated and the steps it took to validate them. Then, we evaluated the result by comparing the run results against our answer key to find: - actual cases tested - tested cases with false failures - tested cases with false passed Using this, we calculated the robustness of each test by calculating actual cases tested/total cases in the answer key, and accuracy by calculating actual cases tested with no errors/the total cases in the answer key. Finally, we built the benchmark by taking the best of the two run results for each agent and averaged  the robustness and accuracy across the board. <Getting Access> Interested in early access to our QA agent? Comment below or DM me—we're making rapid improvements each week and would love to work with you!

39

5 Comments

We Invested in Feltsense, Agentic Founders That Build and Scale Startups At Blitzscaling Ventures, we invest in founders building systems that can reshape how markets form and scale. Feltsense is exploring a new frontier led by Marik Hazan. They are building agentic founders that can create, launch, and scale startups autonomously, (yes, you read that right, AI agents building companies) we call this trend Autonomous Organizations. Historically, startups scaled through human coordination to scale. Teams hired people, built products, and expanded step by step. Feltsense introduces a model where AI agents can ideate, build, test, and iterate continuously. This has the potential to dramatically accelerate company formation and innovation cycles. We believe this represents an early signal of a broader shift. Software is evolving from a tool used by entrepreneurs into an active participant in entrepreneurship itself. If agentic startups can proliferate across markets through shared infrastructure, this may create compounding advantages in data, workflows, and distribution. This is a bold bet on the future of entrepreneurship. We are proud to support the Feltsense team as they explore what autonomous company creation can look like at scale.

44

10 Comments

While there has never been a more exciting time to be a founder building dev tooling or next-gen infra, it has also never been less investable at seed/pre-seed. I'm either really missing something or a lot of my peers are lost. As someone who has not just written, but also SHIPPED, about 75k lines of code in the past 6 months I can tell you that the evolution of how to build products has changed as much in the past year as it did in the entirety of 2007-2017. The complete rise and fail of frameworks, platforms, methodologies, etc... paved over and forgotten... that is of course except for the 1 company that gets a 1000x return from a wildly overvalued hyper-scaler or drunken growth stage investor obsessed with compounding at scale. Imagine a world where any seed investor in trends like Openstack, Hadoop, PaaS, etc all took a full loss on their investment. That's what we're looking at right now. I personally know of over a dozen well-funded seed-stage companies building in these spaces, with years of runway, scrambling to get acquired for a return of capital + several million personally while they're still relevant. If you're not seeing this unfold in front of you, you either aren't paying attention or you're satisfied playing the lottery instead of investing.

75

11 Comments

Very compelling from Cloud Capital - they’ve committed $10m in cashback to startups, giving companies 5% back on their AWS spend.

6

I sent some internal notes to the LocalHost team, and the takeaways apply more widely. Here's the short version. -------- Portable Legitimacy and Coordination Infrastructure LocalHost is building infrastructure and launchpad for companies going from -1 to 0 and human potential. The entire venture apparatus is optimized for companies that already exist – cap tables, traction metrics, unit economics. The harder problem is earlier: conviction + capability, but no proof. Traditional gates are closed not because ideas are bad, but because builders lack the right zip code or credential. We're building a parallel institutional stack. Traditional geography gives you three things: density of ambitious people, access to resources, and legitimacy by association. These are increasingly unbundled. You don't need to be in one place to get all three. Run the experiment: take builders from Benglauru, Cluj, Tokyo. Put them in a house with capital, remove logistical friction, surround them with 15 others shipping products. They build faster. The constraint wasn't access to Sand Hill Road – it was focused time and peer conviction. What one year revealed: Iteration speed beats pedigree. BLR House: 30 days, Three major products launched – Maya Research, Shiro, Zenith. Combined 950K+ views. The common thread was shipping velocity in a high-conviction environment. Legitimacy is memetic, not geographic. went from unknown to backed by LocalHost to term sheets. Public vouching from credible peers matters more than institutional proximity. Infrastructure scales, subsidies don't. $300K+ deployed, but the real unlock was systematic friction removal – free flights, housing with meals, workspace access. All with compounding returns. Each founder strengthens the network. Distribution compounds organically. Zero paid marketing. 2K to 29.5K followers. Caleb's video: 30M views. Genuine stories travel further than manufactured ones. We're building an infrastructure for globally distributed builders who share values but not locations. The playbook: create high-trust physical nodes in multiple geographies, ensure each has local context + global network access, make movement between nodes frictionless, let reputation become portable across borders. This is closer to what university was supposed to be - density of ambition, removal of survival constraints, legitimacy that compounds over time. Except it's opt-in, performance-based, and not geographically gated. Companies will increasingly be built by people who optimize for conviction density over proximity to capital.

64

Everyone knows what YC is about. Build fast. Talk to users. Focus. Ship. And that’s exactly what it is. Had AxionOrbital Space (YC W26)'s CEO and CTO on Vesting talking about their YC experience - and what stood out wasn’t some secret formula. It was the discipline. Sometimes the most powerful thing isn’t new advice. It’s an environment that forces you to execute on what you already know you should be doing. Clip below 👇

59

3 Comments