AI Builders
Digest

Box CEO Aaron Levie argues that the simple litmus test for whether you're pushing AI hard enough is this: if the output is something you could already produce before AI, just faster, you're likely not pushing it enough. We are entering a period where AI output can and should dramatically exceed what you could do before, in a fraction of the time. Product prototypes should be fully functional and interactive with meaningfully better features; engineering architectures should deliver performance gains previously out of reach; research should consume every fact across every dimension; business processes should operate entirely differently and learn from all past failures in ways no human could. AI will be smarter than us on most axes (intelligence remains jagged), and combining it with human strengths yields qualitatively different outcomes.

Latent Space host and AI engineer Swyx notes that strong opinions on model effort levels have become the tabs-versus-spaces debate of our time—just tweak to taste. He points out that Andrej Karpathy uses the medium setting on Opus 5.5 for his latest project (Hackertube, launched the same day). Swyx himself prefers low-effort plus a critique loop, especially for throughput-constrained workflows, and cheerfully claims his preference is superior.

The Takeaway: At Arca, an AI-native wealth management firm, agents proactively handle multi-day (often 10-day) financial workflows and prompt human advisors more often than the reverse, delivering a level of service previously impossible while keeping humans firmly in the loop for money movement and client communication.

Aman Singh leads the AI platform team at Arca, which recently raised $64 million. Their goal is to run the business itself with AI rather than sell SaaS tools. Clients range from founders who just sold a company and tech executives planning early retirement to families buying a first home; services span advice, planning, tax, insurance and estate work.

The core principle: AI replaces all the work around the advisor, never the advisor. Clients rarely touch AI directly—they simply experience dramatically more proactive, responsive care. Agents own analysis (e.g., optimal liquidation to free $100k for a tax bill) but cannot execute trades, move money or speak to clients without human approval. They run a fully model-agnostic internal harness, currently favoring Claude, guided solely by proprietary benchmarks. They deliberately avoid dependency on any lab’s managed-agent product because “owning your intelligence” is what lets you build something unique enough to move the business.

Long-horizon state is maintained by persistent memory systems plus a daily reflective “dreaming” process. Safety is entirely deterministic plus fast LLM verifiers that strip decision rights from the agent at capability edges. Every staff edit or cancellation is logged with a “why” and becomes an eval; a nightly meta-agent proposes harness improvements (still human-gated). Because they are the firm rather than a vendor, they can standardize workflows so agents participate in predictable ways.

Memorable lines: “Agents prompt people more than people prompt agents.” And “We’re actually not a SaaS company. We are the firm ourselves.”

Recent posts from Thibault Sottiaux (OpenAI Codex & ChatGPT), Peter Yang, Guillermo Rauch (Vercel CEO), Garry Tan (Y Combinator President & CEO), Matt Turck and Nikunj Kothari were largely seasonal, personal or observational and contained no substantive AI product, technical or industry insights.