AI Builders
Digest

The Takeaway: Real AI adoption at a hedge fund starts with the CEO treating it as non-negotiable culture, not an optional tool.

Will, CEO/CIO/managing partner of Walleye (nearly $10B AUM multi-strategy hedge fund) and Every consulting client, has a deep technical background as a Princeton engineer and Oxford math PhD who built quant strategies for years. The generative AI pivot accelerated after a 2023 demo from an analyst who used early GPT tools to augment research, and a more recent Chris Sacca podcast that pushed him to mandate firm-wide AI training for all 400 employees regardless of technical skill.

He wrote to the entire firm: “Using ChatGPT is not cheating. That’s a non applicable idea from academia. I use ChatGPT to write this email. You should be using it too and be proud of it. 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. Not using these tools is like refusing to use the Internet in 1995 because it wasn’t perfect.”

As owner-operator he can steer the “cruise ship” fast. Internally they run Current for fundamental long-short teams (usage spikes in earnings season), ~75% of the firm uses ChatGPT-like tools weekly, and about a third uses AI coding tools. They record nearly every meeting and call, building an internal “Borg” data lake so LLMs can process risk-call transcripts for memory and predictive insight. His core stance: tools are imperfect but the direction is clear; only the leader can set the tone that falling behind is the real risk.

Claude Blog: Claude now has its own built-in browser in Cowork, and Claude in Chrome is generally available on every paid plan.

When a task needs a website, a separate browser opens in the Cowork side panel on the desktop app. Claude navigates, reads, clicks, types, and finishes the job without any extension and without seeing the user’s own tabs, bookmarks, or passwords. Logins can be imported site-by-site (banking, email, and SSO excluded by default). Existing Claude in Chrome users keep it as default and can switch in Settings.

Claude in Chrome simultaneously goes GA with more autonomous actions: a safety classifier validates every action against the original request before execution. Defenses against prompt injection include continuous training on an expanding attack library, probes that screen tool results, and the new action classifier. Evaluations show 0% success rate against Sonnet 5 and Opus 5 when probes + classifier are active (0.3% on Fable 5, all low-severity). Enterprise admins can restrict domains.

Practical impact: internal dashboards, legacy systems, and vendor portals without connectors are now directly operable by Claude.

Box CEO Aaron Levie summarized Q2 results: $321.1M revenue, up 9% (11% constant currency, highest in 14 quarters), full-year guidance raised to $1.290B. Growth is driven by enterprise demand to unlock unstructured content with AI.

He argues the world’s most advanced superintelligence is only as useful as the enterprise knowledge it can access. Contracts, research, financials, marketing assets, roadmaps—the bulk of corporate knowledge—live in unstructured data. AI agents can now query it, mine intelligence, and automate nearly any workflow involving that content.

Leaders also recognize that as models from OpenAI, Anthropic, Google, NVIDIA and others advance, enterprises need applied platforms that let them swap models or agents on existing workflows without migration. Security is non-negotiable after incidents like the OpenAI Hugging Face event; robust guardrails, audit logs, and real-time alerts are required. Box is focused on becoming the leading intelligent content management platform for unstructured data at scale.

SPC General Partner Aditya Agarwal announced DeepCogito’s $43M Series A, backed by Benchmark, TQ Ventures, Atreides, Nexus VP, Zscaler and others.

DeepCogito is a post-training research lab whose core method is iterated distillation and amplification (IDA). They have already demonstrated the approach publicly on models from 3B to 600B+ parameters. Founders Ishaan Arora and Dhruv Malrana have known each other 14 years and met the SPC team as Founder Fellows in March 2024. Aditya’s thesis: the AI frontier will be determined by post-training, and the team is pushing large-scale RL and recursive self-improvement.

FirstMark VC Matt Turck called it a win-win-win: NVIDIA is now officially the center of open-source AI with Nemotron and Hugging Face; Hugging Face gets a massive boost and the perfect home (no more business-model nagging); open-source AI wins. Congrats to Clément Delangue, Thomas Wolf and team.

Vercel CEO Guillermo Rauch highlighted global compute at agents’ fingertips: multiple regions, failover, up to 10,000 concurrent sandboxes and 5,000 vCPUs/min ramp by default, with more regions coming. They also shipped a security dashboard and the vercel security check CLI so agents can improve security posture with human-in-the-loop or via crons.

Anthropic’s Thariq (Claude Code) announced his change is live: Claude now has a SendFeedback tool so users can simply tell Claude to draft and approve a report instead of manually hitting /feedback. Feedback is highly valued for improvement. He also noted customers are being targeted by fraudulent requests and appreciated Stripe’s work to protect legitimate usage.

Peter Steinberger noted that Codex’s visualization feature has gotten really good. Dan Shipper separately announced that Codex native apps are here.

Google VP Josh Woodward (Google Labs, Gemini App, AI Studio) shared that he is experimenting with multiplayer AI.

OpenAI’s Thibault Sottiaux (Codex & ChatGPT) observed that a few weeks at OpenAI feel like years elsewhere in terms of output, and that one can age a lot in just a few days. In a follow-up he mused that it feels like 20 years since he last pressed the reset button and is intrigued to try dusting it off.

Sam Altman said the 5.5 party was a lot of fun and suggested doing another for the next model release, asking what would make it awesome. He also quoted a report on a negative incident, calling it “a good report about a bad thing.”

Zara Zhang observed a common double standard on AI writing: “it’s fine if I use it; it’s not fine if YOU use it.” She also argued that most PR teams at large companies end up blocking real marketing and branding—the function designed to enhance brand becomes antithetical to it.

Every CEO Dan Shipper amplified the observation that this is a definite golden age for polymaths and philosophers—and a good antidote for epistemic despair. An incredible time.

Matt Turck appeared as guest (instead of host) and covered why the best venture firm in the world might be NVIDIA, the “hyper power law,” why apps are becoming labs faster than labs become apps, RSI and “building God,” AI bubble vs duration-mismatch concerns, and learning to stop underestimating China.