AI Builders Digest

Vercel opens its automated security review to any repository

Vercel Security Review brings the company's internal security agent to customer repositories, including those hosted elsewhere. Guillermo Rauch says it audits every line of code automatically and argues that security checks cannot depend on each engineer remembering to prompt an agent.

• #products
• #security
• #agents

Nebius says execution is the real bottleneck in AI infrastructure

Nebius bets that its software can serve both enormous GPU clusters and small, short jobs, while hyperscalers juggle AI capacity against their own needs. Its CTO says customers now draw up to 700 gigabytes per second from shared storage, a scale the company initially underestimated. “There is one single bottleneck, and that's your ability to execute.”

• #infrastructure
• #hardware
• #agents

Claude now connects teammates to company plugins on first use

Team and Enterprise admins can set up company plugins so teammates connect automatically when they first use them. Boris Cherny says this removes the need for each person to sign in to every plugin separately.

• #products
• #enterprise

Altman says human control must constrain AI progress

Sam Altman says people should remain in control of the future, even as AI systems become more capable. He rejects a future where people are limited to work that AI systems choose to reward, and says preserving human control should constrain AI progress.

• #policy
• #agents

SurrealDB 4 adds a rebuilt engine and Postgres and MySQL support

Matt Turck calls SurrealDB 4 the payoff from years of work after its initial open-source traction. He says the release combines vector, graph, and document data with ACID compliance and native Postgres and MySQL support, while Cloudflare handles peak scale above 10 billion rows per second.

• #databases
• #open-source
• #products

Cat Wu reflects on Claude Code's shift toward autonomy

Announcing her departure from Anthropic, Cat Wu recalls using an early Claude Code to build a working reinforcement learning environment without writing code herself. At launch, users approved most permission prompts and leaned on it for questions and tests; now, she says, it handles large migrations and new products with more trust to act on its own.

• #agents
• #coding

Levie says AI productivity gains now reach 5 to 10 times

Aaron Levie argues that AI adoption is broadening even as the gap between casual users and those working at the frontier grows. He estimates that the upper end of productivity gains has risen from roughly 30–50% two years ago to 5–10 times in some fields, helped by longer agent tasks and less human intervention.

• #agents
• #productivity

Grok's phone agent can handle calls to US numbers

Peter Yang highlights Grok's new phone call feature for tasks such as negotiating bills, canceling subscriptions, waiting on hold, and collecting contractor quotes. He says it currently works only with US numbers.

• #agents
• #products

Nikunj Kothari uses Claude Code skills to improve his workflow

Nikunj Kothari runs Claude Code's /insights every two weeks to identify ways to improve his setup. He uses /playground to visualize unfamiliar architectures and workflows, sometimes pairing Codex with Claude Code to look for dead code, busy paths, and latency improvements.

• #coding
• #agents