Menlo Ventures

The Chime Scholars Foundation® is celebrating a big milestone: we have awarded close to $10 million in scholarships to nearly 1,500 students across the country since 2022. Today, we are excited to announce the new 2026-2027 Chime Scholars Foundation cohort. Among the 800 scholars in this year’s class, 74% are first-generation college students and 100% of them are Pell-eligible, pursuing paths from four-year universities to trade schools and community colleges. Financial progress starts with access, and access means having the resources and support to reach their goals. Now in its fifth year, the scholarship program is already showing meaningful impact: our scholars graduate at nearly 2x the national average rate*. Congratulations to this year's class. 💚 Read more about the 2026–2027 class in our first comment. *87% of Chime Scholars graduate college — nearly 2x the national average for students eligible for Pell Grants.

Writing code with AI is becoming the easy part. Reviewing it, building it, and shipping it safely is the new bottleneck. Today, we're launching agent-ready Harness Code Repository with built-in AI Code Review. Most source control systems - GitHub included - were designed around human pacing. A developer writes code, opens a PR, and a few people review it over the next few hours or days. That model starts to break when agents are writing code. PRs pile up faster than humans can review them. Indexing and search have to keep up with a much higher volume. And permissions designed around developers don't have a clean answer for an agent that may merge or deploy code on its own. We built Harness Code Repository for that world: - Agent-scale throughput. Scale-tested with thousands of PRs and commits at once in a single repository, with branches in the tens of thousands. - Scoped permissions for agents. Agents inherit permissions from the developer who triggers them, with RBAC and OPA policies that can scope them further to a specific repository, branch, or environment. - AI review focused on risk. AI Code Review groups changes by risk instead of just by file, so behavior-changing code doesn't get buried under mass renames, dependency bumps, and generated boilerplate. Teams can make AI Checks mandatory, and failed required checks block the merge. AI Code Review works with GitHub today, so you don't have to move anything. Harness Code Repository is free to start, with a simple migration if you do want to move. AI Code Review is going beyond the diff. Using the Harness SDLC Knowledge Graph — deployment architecture, policies, and past production incidents — it can flag risks that the code alone can't reveal. We've been using Code Repository and AI Code Review on our own pipelines for months. Across hundreds of developers, we estimate AI Code Review saved more than 10,000 hours in the last month alone. Huge credit to the Harness engineering teams who rethought source control and code review for an agent-first world. Link in comments.

Introducing S1, our general robot model that learns from just one example. S1 can carry out 10 minute long tasks it was never trained on, simply by prompting it with a single human demo without fine-tuning. How does it work? Much like a language model, S1 does not update its weights to learn a new task. Instead, it receives a demonstration of the task in its context window, and uses that to decide what to do next. It can be taught extremely long-horizon tasks. By composing manipulation primitives learned in pre-training, it can make a new style of coffee, pot a plant, fry pancakes, and more. The first time S1 flipped a pancake, we assumed pancake flipping must have been in its pre-training data. We searched millions of hours of pre-training data — and found nothing. S1 inferred the out-of-distribution task from one demonstration. Compared to conventional VLAs, S1 is a step change improvement. On unseen tasks after 100K hours of pre-training, language-prompted VLAs reach 9% accuracy. Our in-context learner reaches 66%. S1 does not rigidly replay the demonstration. A model of its size can withstand perturbations, recover from its own errors, and at times execute with more precision than the demonstrator. We're excited to deploy S1 with our limited industrial partners today, rolling out to all customers over the next few months. S1 is the first light on a new path towards general intelligence — grounded in the physical world. Technical blog: https://skild.ai/blogs/s1

What happens when intelligence gets 100X cheaper? Shashank Goyal, our Head of Provider Ecosystem, sat down with Deedy Das, Partner at Menlo Ventures, to unpack a few truths about AI economics: - Token prices are down 40%, while usage has grown 4X. - Optimizing costs too early can be the wrong move. - The best AI products will increasingly be built across multiple models. This conversation is part of New Defaults, our event series exploring the shifts reshaping how AI gets built and used. Watch the full episode in the comments.