Vibe-Coding an Attention Firewall, w/ Steve Newman, creator of The Curve

Episode Summary

Executive Summary: Steve Newman describes building a deeply personalized AI productivity stack: custom apps for newsletter summarization, attention filtering, task management, and agent monitoring, all built with Claude and Cloudflare. The discussion also covers security tradeoffs, why custom UI matters, how AI may reshape software work, and Steve’s cautiously skeptical but increasingly open view on AI progress, robotics, and climate impacts.

Main Topics: Personal AI productivity stack and bespoke tooling (Priority: 5/5): Steve walks through a suite of custom apps that transform how he reads, triages, and acts on digital information, including a summarized feed reader, attention firewall, to-do app, agent dashboard, and logging infrastructure. Attention firewall and focus preservation (Priority: 5/5): A major theme is reducing context-switching by classifying messages as urgent or not, surfacing only what needs immediate action, and protecting attention for deep work and life outside the computer. Security, trust, and controlled autonomy (Priority: 5/5): Steve explains why he is cautious about giving AI direct control over email and other sensitive systems, emphasizing that other people’s data is also at stake and that utility must be balanced against security. How to build effective agentic workflows (Priority: 4/5): The conversation details practical patterns: use simple prompts, keep projects isolated, rely on logs, accept fast iteration, and build custom UIs for recurring tiny actions that agents alone do poorly. AI capability, thresholds, and the possibility of takeoff (Priority: 4/5): Steve is skeptical of immediate singularity claims but recognizes that AI systems are rapidly improving and that threshold effects may emerge from the interaction of model capability, user adoption, scaffolding, and workflow design. Climate, infrastructure, and the broader effects of AI (Priority: 3/5): He revises prior climate views slightly, acknowledging faster-than-expected data center growth and weaker hyperscaler climate commitments, while still believing AI could ultimately improve overall efficiency and emissions. Golden Gate Institute and the Curve conference (Priority: 3/5): Steve describes the nonprofit’s mission to improve collective sense-making across AI-related fields and to convene cross-disciplinary, in-person conversations that bridge isolated communities.

Key Arguments: Custom AI tools deliver the most value when they are tailored to the user’s actual workflow, not when they are optimized for the agent’s efficiency. A simple LLM-based summarizer can be highly useful even without memory or cross-document context if it helps decide what is worth reading. The most valuable productivity gain is an attention firewall that filters urgent from non-urgent messages, preserving uninterrupted focus time. Security concerns are real because AI systems may access not just your data but other people’s trusted communications as well. The integration layer is the hardest part of practical AI tooling, especially for systems like WhatsApp, Twitter, and cross-app synchronization. Building robust logging across all components makes agentic coding far more debuggable and productive. Custom UIs matter because many everyday tasks are too small and frequent to justify prompting an agent each time. Current model progress is impressive, but human-level general intelligence across all tasks remains far away from a solved problem. AI may amplify software creation enough to increase total software demand even if per-task labor decreases, but the nature of jobs will change significantly. AI’s climate impact may be mixed in the short run due to data center energy use, but longer-term effects could be net positive if AI accelerates efficiency and clean technology.

Data Points: Projects in progress: About 15 - Steve says he has around 15 side projects, mostly personal productivity tools. Incoming information volume: About 50 Substack/blog/newsletter items per day - He describes being flooded by reading material across inboxes, Twitter, and WhatsApp groups. Message volume: A few hundred emails/Slack/WhatsApp messages per day - This is the load his attention firewall is designed to filter. Monitor setup: Second monitor - Used to display urgent messages and a rolling view of the calendar. Agent concurrency: 0 to 5 parallel agents - When coding, he may run several Claude agents at once across separate projects. Subscription spend: $200 Claude plan - He says all of his coding fits within the Claude Pro tier. Additional Claude credits: $30 increments - Claude occasionally auto-purchases extra credits when he exceeds the subscription limit. Backup frequency for to-do app: Every 5 minutes if changed - The to-do app is treated as the most important database and is backed up very frequently. Conference attendance: About 350 people - Curve brings together people from multiple AI-related communities. Conference date announced: October 2nd through 4th - Steve says this year’s Curve conference is scheduled for those dates. Historical coding start year: 1985 - The host introduces Steve as programming professionally since 1985. AI model benchmark: Claude was #1 on his personal leaderboard for 99% of the days over the last couple of years - He uses podcast intro drafting as a personal model benchmark.

Pivotal Quotes: "the agent's not important, I'm important" — Steve Newman: His principle for avoiding token-maxing and optimizing his own attention rather than the model’s uptime. "I want to optimize my time, not, you know, I'll give Claude its next prompt when I'm good and ready" — Steve Newman: Explaining his shift away from keeping agents constantly fed toward preserving his own workflow and attention. "the simplest possible thing" — Steve Newman: How he characterizes several of his tools, especially the feed reader and attention-filtering systems, despite their practical impact.

Implications: Listeners are encouraged to build AI around their own workflows, prioritize logging and safeguards, and treat custom UI as a major unlock. The episode suggests AI’s biggest near-term value is personal and organizational augmentation, not full autonomy.