Summary

Steve Newman’s highest-leverage AI use is a portfolio of roughly 15 personal apps designed to reclaim attention, not maximize agent output. About 50 newsletters, blog posts, and other large information items reach him daily, alongside a few hundred emails and messages; summaries tell him what deserves reading, while an “attention firewall” decides what deserves interruption. His operating maxim is blunt: “The agent’s not important. I’m important.”
Cheap code makes intensely personal interfaces economical, shifting value from generic workflows toward software shaped around one user. Newman’s calendar exposes the exact meeting links and documents he needs, his agent dashboard turns terminal state into glanceable colors, and a Chrome extension saves 15 seconds by clicking five routine Google Meet buttons. Labenz’s closing framing is that agents handle “verbs,” while custom apps supply the neglected “noun side” of AI.
Newman moves fast by limiting the blast radius, preserving canonical data, and making every failure observable. He deploys directly to production without a staging environment because Gmail, Slack, and WhatsApp remain the systems of record; if his interface breaks, the underlying messages survive. Every backend, browser frontend, and Android component logs to one service, enabling Claude to investigate evidence instead of guessing: “Thou shalt log errors.”
The hardest and potentially most defensible layer is integration, not model prompting or interface construction. One line hands a message to an LLM for urgency classification, but real-time ingestion required separate machinery for Gmail, Calendar, Slack, Twitter, Signal, SMS, and WhatsApp’s undocumented local SQLite database. This creates a strategic contest: users want the best agent across services, while SaaS vendors want them captive inside first-party interfaces.
Software output can surge even as the labor required per unit collapses, leaving the employment outcome genuinely unresolved. Newman now runs zero to five agents, reports hearing that work that once required a team can increasingly become one person’s remit, and imagines “full-stack product manager” replacing parts of traditional engineering. Yet he considers it plausible—even probable—that Jevons-style demand creates more software jobs until AI becomes better at essentially all nonphysical work.
AI progress is a multiplicative ecosystem phenomenon, so model benchmarks alone understate both adoption speed and uncertainty. Newman’s Second Thoughts framework lists eight interacting factors—pre-training, post-training, inference compute scaling, agent scaffolding, app design, user aptitude, workflow refactoring, and adoption—while Labenz argues that thresholds emerge from the whole system. Even so, Newman’s position remains: “The models are amazing. We’re still far from AGI,” where “far” might now mean only five years.
Recursive self-improvement is plausible but not yet demonstrated because automating coding is not obviously equivalent to automating all of AI research. Newman wants to know whether models possess the taste required to design valuable RL environments, manage expert feedback, choose experiments, and navigate obscure research bottlenecks. A system might automate AI R&D within a year, or those higher-order capabilities might take materially longer; he declines to collapse that uncertainty into a singularity forecast.
Physical-world reliability and energy demand remain the two clearest complications to a purely digital acceleration thesis. Waymo is, for Newman, a real existence proof, but household robotics cannot tolerate 50% or even 80% task success, and polished demonstrations may conceal the gap between scripted motion and robust tactile control. He also concedes AI power demand rose faster than he expected, though he still thinks long-run industrial efficiency and clean generation might outweigh near-term fossil-fuel use: AI will “roll the dice on the whole world,” with climate carried along.

Deep dive

1. Fifteen tiny applications became Newman’s answer to digital overload

Newman, who has been programming since at least 1985, now builds roughly 15 projects during evenings and weekends alongside his Golden Gate Institute work. Most are personal-productivity tools born from “drowning” in the expanding volume of AI news and digital communication.
His information intake includes roughly 50 Substack posts, newsletters, blog posts, and other large information items per day, plus Twitter and multiple WhatsApp groups. Reading everything consumed hours before he could synthesize it, let alone turn it into useful work.
Two needs ultimately organized the portfolio: distill the information flow enough to select what merits attention, then recover uninterrupted time by allowing only genuinely timely messages through. The target is a changed life, not “optimizing the agent setup for the agent setup’s sake.”
Newman was about three-quarters finished with the messaging system before he understood that he was building an attention firewall. His broader framing is a software “Cambrian explosion”: people are fumbling toward bespoke practices whose purpose becomes clear only in hindsight or “mid-flight.”

2. A deliberately context-free reader is sufficient to choose what matters

Labenz expected the reading system to model Newman’s interests, remember prior reading, and identify genuinely novel claims across sources. Newman’s surprising answer: “I’m not.” It sends each full article or transcript to an LLM with a static summarization prompt and no personal context.
The first summary is short enough to skim in about 10 seconds and decide whether an item deserves reading. A second, one-page overview surfaces novel ideas and notable evidence, providing an “80/20 alternative” when the short version seems useful but the original does not justify the time.
Novelty is judged against the model’s training and first-principles expectations, not Newman’s history. He had imagined a system that knew everything he had read, but never built it because “I know what I’ve already read”; human memory plus rapid skimming proved efficient enough.
The interface offers the original, archive controls, and a button that opens a Claude session with the article in context—but Newman had forgotten that last feature existed because he rarely used it. Cheap implementation also funded luxuries such as D2 backups, a readable backup viewer, and a bookmarklet that extracted otherwise inaccessible Substack subscriptions.

3. Radar replaces 30 daily inbox checks with one trusted surface

Before Radar, Newman checked roughly five communication apps whenever his brain “came up for air,” perhaps 30 times daily. Only a fraction of several hundred incoming emails and messages required prompt attention, but every check created another chance to be diverted by something that could wait.
Separate integrations ingest Gmail, Slack, Signal, WhatsApp, and other channels; then “one line of code” asks an LLM whether each message is urgent. The real behavioral specification is a one-page rubric accumulated “exception by exception,” not an elaborate personalized memory system.
After roughly 40 years of engineering without one, Newman bought a second monitor for a rolling three-hour calendar and the prioritized message stream. The system categorizes items into buckets such as urgent and midday, which often remain empty precisely because the interface makes them easy to clear.
Radar’s calendar is idiosyncratically filtered rather than complete. Known recurring meetings expose direct buttons for Google Meet, Newman’s private notes, and the team’s shared document; a message toolbar includes one-click operations such as forwarding an account notice to his wife and archiving it.

4. More capable agents turn security from abstraction into duty of care

Newman has not let agents broadly answer email or act on his digital life. The utility is compelling enough that he expects eventually to enter that “pool,” but the complexity and security implications conflict with his reluctance to use anything he cannot understand and trust.
Newman, not Labenz, described his historical computer-security posture as “borderline negligent,” then explained why 20 years of Gmail changed the calculation. An AI receiving everything he has written also receives everything others entrusted to him, creating obligations those senders could not have anticipated.
Both see an unresolved contradiction: agent systems are widely described as fundamentally insecure, yet there are surprisingly few public disasters involving prompt injection, deleted databases, or erased email. Model and tool developers may be staying ahead with classifiers, but Newman still adopts the Raiders of the Lost Ark posture: “Asps. Very dangerous. You go first.”
Newman’s Tylenol analogy captures his confusion. Low-hanging harmful acts can remain unused for centuries, while an exposed, unpatched Windows installation could historically be compromised within about five minutes; perhaps malicious actors adopt slowly, or perhaps people skilled enough for AI phishing can currently earn easier legitimate money.

5. Low stakes, canonical data, and Cloudflare make cavalier deployment rational

Newman chose his stack by asking Gemini, ChatGPT, and Claude to recommend hosting, language, frontend, backend, and CSS choices, then feeding each answer to the others for critique. Cloudflare won for its combination of server hosting, cron jobs, queues, databases, simple operations, and low small-scale cost relative to AWS.
He has no staging environment and continually pushes to production, trusting whatever tests Claude says it has written and run. That posture works because the systems of record remain Gmail, WhatsApp, and Slack; even if an agent deletes his custom database, he loses the convenient interface rather than the messages themselves.
Backups reflect the value at risk. Most applications dump data into D2 file storage, while the to-do app creates a backup every five minutes after changes and syncs it to GitHub, leaving a static, browsable task list available even if the live service fails.

6. Cheap implementation turns neglected annoyances into bespoke products

One tool reads a Hacker News article and its entire discussion, identifies themes, summarizes both the source and the thread, and assigns comments to one or more themes. It lets Newman skip heterogeneous branches and inspect only the slice of a long discussion that matters.
His Chrome extension can send the current page into his to-do app, reader, or whitelisted areas of his Notion tree. Dropping a Hacker News discussion into the reader automatically invokes the thread-analysis pipeline, turning a tiny capture action into a multi-application workflow.
Other narrow tools include a Gmail spam view sorted by useful attributes, making messages misaddressed to aol.com easy to skip, and a Twitter reader that deduplicates retweets and automatically expands posts. These are not new product categories; they are fit-and-finish decisions optimized for one person.

7. A color-coded dashboard makes five coding agents glanceable

Newman’s entire agent-management interface is a row of colored project buttons. Blue means an agent finished, red means no agent is running but the to-do database contains work for that project, and other states indicate active work or a question requiring him.
Claude Code hooks report status to a cloud API, which pushes real-time updates to the web page. AppleScript, a Chrome extension, and Hammerspoon—an arcane macOS automation utility Claude told him to install—connect a click or keyboard shortcut to the corresponding terminal tab.
The hooks do not expose enough information for a perfect state model, but they are “good enough.” Newman’s brain has internalized the colors, so he no longer patrols terminal windows; the dashboard is another attention firewall rather than a system for maximizing agent utilization.
Roughly 15 applications live in separate GitHub repositories with their own databases and Cloudflare projects. Newman chose microservice-like isolation to keep agent context small—probably more important in the “ancient days of January”—while allowing sibling projects to invoke one another’s APIs.

8. A shared Docker neighborhood gives isolated projects selective context

All repositories sit beside one another inside a Docker container, where Newman runs Claude Code with dangerously skipped permissions. Agents can inspect neighboring code, update an API producer, and then modify its consumer without Newman manually relaying instructions between sessions.
The applications remain operationally separate despite that shared visibility. Newman’s GitHub repositories contain code, while application data resides in Cloudflare databases; the topology seeks manageable context and failure isolation without preventing cross-project work.
Labenz described a private monorepo for personal material and a separate local SQLite approach. Newman’s context exports, accumulated over roughly five years, add up to about 1 GB and are stored in GitHub—“Git LFS or whatever that’s called,” in his uncertain wording.
Labenz’s pointers are intentionally asymmetric when a podcast-production project must be shared: Claude started in the private repository can see the collaborative project, but the shared project is not explicitly directed back toward personal history. Labenz emphasized that this is containment and hygiene, not robust security.

9. Real-time integration is the stack’s largest cost center

Newman’s Mirror service ingests Google Contacts, Calendar, Gmail, WhatsApp, Slack, Twitter, Google Docs, Signal, and phone SMS. It creates a rich searchable database, although so far he uses that accumulated context mainly to power the real-time inbox rather than sophisticated personal retrieval.
Google and Slack offer APIs whose complexity becomes Claude’s problem. WhatsApp required a safer workaround: its macOS desktop client syncs messages into an unencrypted local SQLite database, which Newman reads without modifying on roughly a one-minute cron cycle.
That read-only piggyback avoids unofficial WhatsApp APIs associated with reports of account bans, but Claude still had to “spelunk” through an undocumented schema to distinguish replies, original messages, identifiers, and usernames. Components consequently run across the Mac, Docker, cloud, and phone.
Twitter was the weakest link: a questionable third-party service polls each of roughly 80 followed accounts individually, rotating through them about hourly before API costs become noticeable. Labenz mentioned a newer usage-priced official API, and Newman said he would likely switch when the current integration next breaks.

10. Universal logging converts agent debugging from speculation into investigation

Every backend, browser frontend, JavaScript component, and Android app writes to one simple logging service backed by SQLite on Cloudflare. Newman’s CLAUDE.md contains emphatic review rules: log errors, log every database modification, and preserve enough evidence to reconstruct failures.
When a calendar event disappears or summaries stop generating, Newman invokes the “systematic debugging” skill from Jesse Vincent’s Superpowers package and supplies a one-sentence symptom. Claude knows to inspect logs rather than invent a cause, and this works “remarkably close to 100% of the time.”
His single strongest practical recommendation is centralized observability across every moving part, including the frontend. A daily cron report summarizes the logs, but the deeper benefit is giving agents evidence: “Vigorously remind Claude that it doesn’t have to make guesses about what’s going on. It can go look.”

11. Newman optimizes his own cognition rather than agent token utilization

During the most active build period, Newman spent 15–20 hours per week across these projects; now he averages perhaps half an hour daily. Depending on his to-do list, he runs anywhere from zero to five agents in parallel, generally one terminal tab per active project.
Small tasks receive an end-to-end skill: understand the request, implement it, deploy it, commit it, and push to GitHub without questions unless necessary. Larger or ambiguous changes use another skill that pauses for a conversation about design before proceeding.
Newman initially embraced “keep your agents fed” and token-maxing, which became stressful while one agent represented the bottleneck. Multiple independent projects removed most waiting and inverted the objective: “The agent’s not important. I’m important.” Claude gets another prompt when it fits Newman’s mental workflow.
His coding appears to fit within a $200 Claude plan, supplemented by automatic $30 API-credit purchases recently arriving two or three times weekly, partly because he profligately uses Opus for simple summaries. He also pays $200 for ChatGPT Pro; one bad Codex trial produced only a “Bayesian update of 0.002.”

12. The toolkit stays surprisingly vanilla while the development method flips

Apart from Claude Code and Superpowers, Newman directly uses little more than the built-in macOS Terminal. Claude may recruit utilities such as Hammerspoon behind the scenes, but Newman has not found enough unmet need to justify the abundance of fashionable orchestration products.
Newman’s security-conscious tactic is to ask Claude to investigate a popular tool, extract useful ideas, and recommend what applies to his setup. Usually he adopts the core ideas inside his own “tower of Jell-O” rather than installing the package wholesale, aiming for roughly 90% of the benefit.
The enduring difficulty is unlearning the assumption that workflows must adapt to static software. Newman still has not built per-newsletter filters—for example, retaining selected San Francisco Chronicle stories while removing ads, food, and most sports—because he wants a trusted general framework, not 50 brittle features.
His former engineering instinct was “measure twice, cut once”: deeply specify the problem because implementation and debugging were expensive. He now considers that wrong for personal AI software; the better method is to dive in, get it wrong, throw it away, and rebuild, though it remains contrary to his nature.

13. Better models make recovery rare, but human judgment still steers architecture

Labenz recalled advising users six months earlier to revert when an agent looped, then retry from the last good state with a better prompt. Newman almost never reverts, a result he finds surprising rather than evidence of a settled universal rule.
His experience began with Opus 4.5 and soon moved to 4.6, while Superpowers forces major decisions past the user before implementation. Newman never reads the TypeScript itself, but he does think about the high-level choices and redirect false paths before they harden.
The recent workload also consists mainly of incremental features rather than major architectural rewrites, reducing the opportunity for catastrophic thrashing. Models, scaffolding, human review, and task shape all contribute; he does not isolate one as the explanation.

14. Voice preserves walking as thinking time, while custom UI handles tiny actions

Newman rejects the premise that a walk should yield five more coding prompts. He walks for physical health, mental space, and deeper thought; work enters only when an idea already “percolating” deserves capture.
His deliberately mundane mobile workflow is Gmail compose, his own address, and the phone keyboard’s dictation button. He rambles out features, design decisions, and questions, then asks an LLM to “organize this” into a Claude Code prompt—often without reading the cleaned-up version before execution.
Labenz recognized a missing layer in his own podcast system: AI made him produce more songs, artwork, and promotional clips, but he still hunted through filesystems and old terminal output for results. A custom production UI could improve navigation even when the agentic work itself already succeeds.
Newman’s formulation explains the gap: command-driven agents are excellent for large “verbs,” such as writing an app or reformatting 800 PDFs, but inefficient for tiny repetitive actions. An app is the “noun side,” a persistent surface where one click can beat another natural-language delegation.

15. AI would transform enterprise migration through investigation before implementation

Newman’s previous company was acquired in 2021, and the resulting project involved swapping its storage and query engine into the acquirer’s production core. The system ingested certainly billions and probably trillions of daily security events, served millions of endpoints, and supported an IPO-bound company—decidedly not a place to vibe-code the production implementation today.
The expensive uncertainty surrounded the existing system and its workload: neither team fully understood customer data shapes, query complexity, time ranges, or resulting AWS costs. Senior engineers possessed fragments of the needed knowledge but were too busy, leaving teams to obtain access, inspect logs, run queries, and reconstruct reality manually.
Today Newman would assign an agent 100 questions, ask it to build 100 investigative tools, generate and distill 100 reports, identify which deserve human reading, and produce an auditable cost estimate. Because those tools inform judgment without touching production, experimentation can be broad while deployment remains conservative.

16. Software’s labor market and SaaS value chain are both entering a tug-of-war

Newman’s unresolved equation is that engineers required per line of code are plummeting while lines of code are soaring. Jevons paradox might sustain or increase employment, but the work could change as radically as transport did from horses to cars—perhaps toward a large “full-stack product manager” category.
Hiring should favor people comfortable “navigating without a map,” because decades of established patterns no longer capture the opportunity. Communication may also become more valuable: if one agent-enabled person replaces a team, each person inherits a team’s former burden of coordinating with everyone else.
Labenz’s investor framing was bullish for GPUs, Anthropic, and perhaps Cloudflare, while potentially bearish for broad application software. Newman qualified it: he still uses Gmail, Slack, and WhatsApp, but increasingly as backend services whose interfaces and feature sets matter less.
Vendors therefore face a strategic choice. Restricting APIs can force users into first-party apps, yet it may drive them to competitors that accommodate the best external agent; meanwhile deep infrastructure such as S3, Google Spanner, and database backbones remains difficult to replace until users effectively have “Jeff Dean on the command line.”

17. Ecosystem thresholds can accelerate before anything resembles full AGI

A framework from Newman’s Second Thoughts multiplies eight moving factors: pre-training, post-training, inference compute scaling, agent scaffolding, app design, user aptitude, workflow refactoring, and adoption. Labenz argues that people experience the compound system and its second-, third-, and fourth-order effects—not an isolated capability curve.
ChatGPT’s launch was a threshold because capability overhang met a particular product and adoption dynamic. Labenz likens this to a coronavirus moving from an effective reproduction rate of 0.9 to 1.1: the phase change comes from propagation through an evolving system, not what happens inside one host.
The December wave of Opus 4.5 coding similarly combined model quality, holiday availability, and reports from other users that it was worth trying. Practices then diffuse, mutate, and become obsolete quickly, making social adoption part of capability in the economically relevant sense.
Despite Mythos reportedly finding an extraordinary number of security vulnerabilities, Newman maintains: “The models are amazing. We’re still far from AGI.” In his scoring metaphor, a major capability might earn 300 points while being “all the smart at all the things”—Jeff Dean, Terence Tao, and more—requires roughly 50,000.

18. Recursive improvement depends on research taste, broad transfer, and physical reliability

Newman admits that defending his skepticism is “getting a little harder.” “Long timelines aren’t what they used to be”; a system matching the depth, discernment, judgment, and range of human expertise may remain “quite a ways” off while that phrase now plausibly means five years.
His first uncertainty is the true capability checklist for automating AI R&D. Coding and math are visible, but designing productive RL environments may require senior taste; model improvement may also depend on experiment selection, data curation, managing 10,000 human experts, or obscure skills that do not reduce to better software generation.
Even if agents become superhuman at objectively graded work, transfer into marketing, business strategy, management, teaching, product design, and mechanical engineering is not assured. Newman sees coherent stories for AI R&D being automated within a year and for hidden bottlenecks taking significantly longer.
Robotics adds another test. Newman accepts Waymo as a safer-than-human existence proof, but household tasks cannot tolerate 50% or 80% success; a car has roughly “fast-slow, left-right” degrees of freedom and can pull over, whereas hands require tactile feedback and repeated performance amid uncontrolled messiness.

19. Newman revised his climate view without making energy the whole story

Newman says he was “a little bit wrong” to dismiss AI’s emissions significance because he underestimated the speed of data-center scaling. He also did not anticipate that, in the land rush for data-center capacity, hyperscalers might—by his understanding, which he says is fuzzy—back off climate commitments. He cites an xAI data center trucking in the quickest, least-efficient gas turbines and says he thinks other hyperscalers may be doing similar things.
The short-run conclusion is unfavorable: AI electricity use is rising exponentially, some incremental supply is fossil-fueled, and each year’s increase can look more dramatic. He is less certain on current details because climate is no longer his main research focus.
The long-run effect could still be positive if AI improves batteries, materials, electrochemical industry, fertilizer targeting, robotic agriculture, and the efficiency of the broader industrial base. Even a hypothetical 20% rise in power-sector emissions could be outweighed by a 20% efficiency improvement elsewhere, while terawatt-scale generation may economically favor clean technologies.
His governing hedge is broader than an emissions forecast: AI will “roll the dice on the whole world.” Climate will be carried along by whatever transformation follows, even if the immediate marginal data center burns more fossil fuel.

20. Golden Gate is building bridges across AI’s fragmented knowledge pockets

The Golden Gate Institute for AI is a nonprofit devoted to collective sensemaking as consumers, businesses, policymakers, and civil society navigate rapid change. Its premise is that AI spans machine learning, economics, cybersecurity, biosecurity, labor, politics, and education, while no person or community holds the complete picture.
Its flagship Curve conference convenes about 350 people across fields, political perspectives, and geographic communities. Newman argues face-to-face contact converts Twitter antagonists or strangers into humans with intelligible reasons, sometimes producing projects, working relationships, and durable engagement.
The next Curve is planned for October 2–4. Golden Gate hopes to run it twice the following year—prompting a joke that the frequency must double annually—and particularly wants stronger connections beyond the United States, including China, and within underrepresented domains such as robotics.

137. 洪乐潼: AI for Math, Lean, and Is Math Created or Discovered?Ep. 009 - Using Open Source Data To Drive Investment Decisions (ChipBook) | Chaim Eisenberg, Simi Sherman, Jordan Nanos