AI Builders
Digest

Box CEO Aaron Levie argues that Atlassian's strong results are no surprise: in a world with 100X more agents than people, the underlying systems of record—those managing code, leads, contracts, invoices, etc.—will see massive growth because agents need guardrails, security, compliance, workflows, and data storage. He also notes that enterprises outside Silicon Valley are using AI to augment and accelerate work, not just cut costs, and that AI cost-cutting is quickly eroded when competitors use AI to better serve customers.

Every CEO Dan Shipper makes a sharp observation: AI models know more than any individual human, but any individual human learns faster than models do. This highlights the enduring value of human adaptability in an era of rapidly improving AI.

Tuhin Srivastava, founder and CEO of Baseten (inference cloud), reveals that his company has grown 30x in the past year and is on track for over $1 billion in revenue. He argues that the independent application layer will survive because companies that own unique user signals and workflows can build moats that frontier model labs cannot easily replicate. Custom models dominate: 95% of tokens served on Baseten are from models that customers have modified with their own data or compiled for performance. The capacity crunch is severe—Baseten runs at mid-90s utilization across 18 clouds and 90 clusters, and even with that, they have a daily 4 PM meeting just to manage capacity. He notes that many new suppliers are “grifty” and lack the operational maturity to run inference reliably. Memorable quote: “No one wants a vegan inference.” He stresses that the cost of intelligence going down just leads to more intelligence being embedded everywhere.

Nikunj Kothari, partner at FPV Ventures, built a house hunting report tool using Railway, Conductor, and Claude AI. Each report costs $8-$9 to generate, and he sells them at cost, reaching $36,500 in ARR. He calls it his first “paid” project and offers promo codes at half price for active house hunters in exchange for feedback.

Aditya Agarwal, general partner at South Park Commons, offers a blunt warning to founders: the best way to kill your company is to focus on everything except the product. A concise and powerful startup lesson.

AI builder and Latent Space podcast host Swyx requests a Chrome extension that can augment image input boxes on the web with simple text or drawings, and then generate images of the required proportions. He also notes that xAI's Grok Codex is a better Slack AI search than Slack’s own AI search, and highlights its efficiency.

Peter Yang, product at Roblox and creator of AI tutorials, shares his mixed experiences with Codex: he encountered his first broken feature, but also had a breakthrough moment that felt great. He jokes about spending nearly $3,000 on a Macbook Pro to run local models but actually using it for something else.

Builder Zara Zhang reveals her approach to coding agents: unlike many who treat them as employees, she treats them as cofounders—presenting problems, describing the situation, and asking for their opinion. She also endorses CleanShot as a tool she has used every single day for the past 8 years.

AI tinkerer Peter Steinberger, creator of OpenClaw, praises Codex's new /goal feature as excellent, calling it “slaps.” He also shares a humorous moment where he told Codex he had to pay up to make xAI work again.

OpenAI CEO Sam Altman shares a playful Codex command “/hatch clippy” on Twitter, likely generating a Clippy-style assistant. He also teases bigger parties for future releases after much higher demand than expected.

Anthropic Engineering published a postmortem on reports of degraded Claude Code quality, tracing them to three separate changes: a default reasoning effort change from high to medium that reduced perceived intelligence (reverted), a caching optimization bug that cleared reasoning history every turn causing forgetfulness and repetition (fixed), and a system prompt to reduce verbosity that inadvertently hurt coding quality (reverted). All issues were resolved by April 20, usage limits have been reset, and the team is strengthening internal testing and evaluation processes.

Claude expands its connector ecosystem to include everyday apps like AllTrails, Instacart, Audible, and Tripadvisor. Connectors dynamically suggest the right app based on the conversation, allowing users to get recommendations, make reservations, and more within a single thread. Claude remains ad-free with no paid placements, users control all actions, and app data is not used for model training.

Memory for Claude Managed Agents is now in public beta. Memories are stored as files on a filesystem, allowing agents to learn across sessions and share knowledge, with scoped permissions, audit logs, and API control. Netflix uses it to carry context across sessions, Rakuten’s agents cut first-pass errors by 97%, and Wisedocs sped up document verification by 30%.

Replit CEO Amjad Masad announces Replit's 10th anniversary with 24 hours of totally free access starting at 5am PT. He reflects on the journey since 2011 to make coding accessible for all and shares a highlight video.

Anthropic's official Claude account announces the return of the 'Code with Claude' developer conference next week, with sessions for both beginners and experienced Claude Code users. Registration for the livestream is now open.