AI Builders Digest

NVIDIA buying Hugging Face makes it the official center of open source AI

NVIDIA now owns the default home of open weights, pairing Hugging Face with Nemotron. The read is that Hugging Face finally escapes the business model question that has nagged it for years, and open source AI gets an owner with unlimited capital. Turck also laid out a set of theses from a recent interview: NVIDIA may be the best venture firm in the world, apps are becoming labs faster than labs are becoming apps, and the AI bubble worry is really a duration mismatch worry.

• #open-source
• #funding

Claude now ships with its own browser inside Cowork

Claude has a built-in browser in the Cowork desktop app that opens in a side panel and navigates, clicks, types, and fills forms while you keep working. It is Claude's browser, not yours: no tabs, bookmarks, or passwords are shared, and you port logins over site by site, with banking, email, and SSO excluded unless you opt them in. Rolling out this week to Pro, Max, and Team; already available for Enterprise admins to enable. Claude in Chrome stays the right tool for a page you already have open, and remains the default if you already use it.

• #agents
• #products

Claude in Chrome goes GA and can now act without approving every step

Claude in Chrome is generally available on all paid plans, and Claude now auto-approves actions it judges safe rather than asking each time. Three layers make that defensible: training against a growing library of real prompt injection attacks, probes that scan tool results before Claude acts on them, and a classifier that checks each pending action against what you actually asked for. The numbers are concrete. On the harder red-team evaluation, attacks that reached the model succeeded 17.6% of the time against Opus 4.5 and 3.8% against Opus 5 with no safeguards; with probes plus the safety classifier, nothing succeeded against Sonnet 5 or Opus 5, and Fable 5 saw a 0.3% success rate in low-severity scenarios.

• #agents
• #security
• #evals

Box hits $321M quarter, argues superintelligence is capped by what it can read

Box reported Q2 revenue of $321.1 million, up 9% or 11% in constant currency, its best constant-currency growth in 14 quarters, and raised the full-year target to $1.290 billion. The pitch to enterprise buyers is that the world's most advanced superintelligence is only as useful as the enterprise knowledge it can access, and the bulk of corporate knowledge lives in unstructured content like contracts, financials, and roadmaps rather than in databases. Two demands keep coming up from technology leaders: the ability to swap models and agents on existing workflows as token usage grows, and hard guardrails, audit logs, and real-time alerts so neither humans nor agents reach data they should not.

• #agents
• #products

Claude's browser needs no install and stays walled off from your own

When a Cowork task involves a website, a browser opens in the side panel and Claude navigates, fills forms, and finishes the job. Nothing to install: it is built into the desktop app and stays separate from your own browser and logins, rolling out over the next week to all paid plans. Claude in Chrome is also now generally available on every paid plan, and stays your default if you already use it.

• #agents
• #products

Vercel opens global sandbox compute for agents, plus a security check CLI

Agents get multi-region compute with failover, up to 10,000 concurrent sandboxes and 5,000 vCPUs per minute of ramp by default, with more regions coming. Vercel is also shipping a security dashboard alongside a `vercel security check` CLI, built so agents can improve your security posture with a human in the loop or run on a cron. Same shape as `is-agentic`: the CLI is the interface agents actually use.

• #agents
• #products
• #security

DeepCogito raises $43M Series A betting the frontier is post-training

The thesis is blunt: the AI frontier will be determined by post-training, not pretraining. DeepCogito is a post-training research lab working on large-scale reinforcement learning and recursive self-improvement, with iterated distillation and amplification as its core method, already demonstrated publicly on models from 3B to 600B+ parameters. Benchmark, TQ Ventures, Atreides, Nexus VP, and Zscaler joined the round. Founders Drishan Arora and Dhruv Malrana have known each other 14 years and met SPC as Founder Fellows in March 2024.

• #funding
• #research

A $10B hedge fund made AI use mandatory and told staff to stop hiding it

Walleye's owner-operator sent the whole firm a memo opening with "I use ChatGPT to write this email, you should be using it too and be proud of it," and added that "as a hedge fund, we should be ashamed to leave money on the table by ignoring tools that make us faster, smarter and more effective." The specific problem he was solving was shame: people were generating drafts and then roughing them up so they would not look machine-written. Adoption now runs around 75% of the 400-person firm using a chat assistant weekly and roughly a third using AI coding tools; an internal product called Current turns filings, broker notes, and earnings transcripts into analysis for long-short teams, and over 50 outside firms have asked to be beta users. He is careful about the limit: these tools do not remove the need to think, they buy back the mechanical typing time so you can think more, and an internal report that reads like it was generated without thought still fails. He also rejects the idea that benchmarking productivity gains inside a real company is meaningful, and argues that a leader's job during this shift is to set a tone where a broken demo in front of a third of the company costs nothing.

• #enterprise
• #products

Claude Code adds a SendFeedback tool so you can just tell Claude what broke

Instead of running /feedback and writing up a report yourself, you can now ask Claude to draft one and approve it. Separately, several Anthropic customers are being hit with fraudulent requests, the kind of abuse that degrades what everyone can offer legitimate users, with Stripe doing work to help and plenty left to do.

• #products
• #security

Codex native apps ship

OpenAI's Codex now has native apps. The framing from Shipper is that this is a golden age for polymaths and philosophers, and a decent antidote to epistemic despair.

• #products

Most PR teams at large companies exist to block real marketing

The function built to strengthen a company's brand ends up working against it, which is about as ironic as corporate structure gets. Related double standard worth naming: on AI writing, plenty of people land on "it's fine if I use it, it's not fine if you use it."

• #marketing