Theme Studio
Pick a palette + text modeWriting
What I shipped, what failed, and what changed in the next iteration.
Showing all 52 posts.
Why I made Gumdrop browser-first, how the audio analysis feeds the scene system, and what makes the product feel lightweight instead of setup-heavy.
Automation has been compounding since the steam engine. AI will likely follow the same pattern: first-order labor narratives first, then harder-to-imagine layers of infrastructure, extension, and new kinds of agents.
A scrubbed look at the daily agent I use to triage recruiter outreach, check LinkedIn leads, avoid scheduling conflicts, and reduce the mental load of job-search inbox work.
As of April 28, 2026, both companies can do a lot. The more interesting difference is where each seems to think the center of the product universe is: Anthropic around technical workflows and OpenAI around a broad general assistant, especially voice and multimodal surfaces.
AI teams increasingly need hybrids who can move between product judgment, implementation reality, and operator-level AI workflow design. This is why PM/engineer-type roles matter more now.
AI agents increase leverage not by replacing judgment, but by letting one human operator run more repetitions, more parallel work, and tighter feedback loops.
AI can automate more output, but humans still define direction, care, meaning, and stewardship. A practical framework for building abundance without losing our humanity.
A practical breakdown of human-visible text versus machine-visible payloads, why hidden channels matter, and how to design AI products defensively.
AI is moving from text-only interaction into vision, audio, and action loops. The core tension: multimodal systems may approximate human sensing and expression before we understand consciousness biologically.
AI can simulate human outputs at impressive quality, but simulation does not erase origin. I argue for a future where automation serves biological life, not the other way around.
We went from copy-paste coding in 2023 to managing persistent AI workers in 2026. The next skill is organizational design for agent teams.
AI agents are most effective when the human operator understands their own goals, can delegate clearly, and can audit what the system is actually doing.
AI can approximate an idealized person at scale, but our limits, errors, and lived constraints are part of what make human meaning possible.
Managing AI agents is starting to look like running a factory: throughput, quality control, role design, and escalation paths. But the deeper question is whether that frame is temporary scaffolding.
A synthesis of Waterloo-era cognitive science questions: where consciousness might begin in biology, what computation captures, and what Chinese Room limits still imply for AI.
Naval's metaphor is useful: AI can be a high-leverage cognitive vehicle. The strategic question is whether labor mechanization expands human agency or concentrates control.
Andrew Ng's electricity metaphor still holds. The next phase is execution at scale: agent engines can drive major abundance, with real upside for broad access if costs keep collapsing.
Model quality improved fast, but bad UX still kills adoption. The interface is where trust, clarity, and repeat usage are won or lost.
In the AI era, PM leverage comes from implementation literacy: you cannot scope, steer, or evaluate what you do not understand.
How I think about behavior-shaping products without manipulation theater: agency, constraints, feedback loops, and measurable improvement.
A tamagotchi-inspired experiment in digital attachment: what nurture loops reveal about emotional connection with non-living systems.
What an overhead horse herd simulator reveals about local rules, global patterns, and why coherent behavior can appear without a central controller.
Horse Herd Simulator explores emergence: how local interaction rules produce coordinated global behavior without central command.
How to write prompts as contracts: role, goals, constraints, and failure behavior that hold up in production.
The principles I use to keep AI systems readable, reliable, and fast enough to ship every week.
The personal assistant project is built as an adaptive interface scaffold: fast iteration, explicit constraints, and durable architecture.
A practical design stance: less UI noise, more user confidence, and better decisions under uncertainty.
A practical philosophical stance on AI anthropomorphism, Turing-style behavior tests, and what they do and do not prove.
How I run lean loops for AI features: define, build, measure, and decide without fake certainty.
Life HUD applies RPG interface logic to real life operations: state visibility, actionable quests, and momentum loops.
A no-nonsense map for shipping AI features that survive real users, constraints, and messy context.
Family Tapestry treats lineage as living structure: editable graphs, narrative memory, and relational nuance.
LLMPrism is built on comparative thinking: side-by-side model behavior with privacy-first defaults and no provider worship.
ShipDojo is built on execution realism: move from prototype optimism to production evidence with explicit gates.
A token falls through a weighted machine: guided, partly stochastic, mathematically rigid, and still hard to predict path-by-path.
A compact operating model for shipping AI products without handoff theater or role confusion.
Divine Machine treats AI as a ritual interface: deterministic internals, interpreted meaning, and epistemic humility.
A grounded framework for agents: workflows with memory, tools, supervision, and operational accountability.
Why Bonzen is designed as a behavior loop system, not a content library: relevance, recovery, and repeat value.
This is not just model drama. It is a strategic split on product philosophy, safety posture, and developer lock-in.
How I scope one clear end-to-end AI feature for product teams so launches happen fast and hold up in real usage.
Open-ended chat feels flexible, but constrained interfaces usually produce better outcomes for real users.
A lightweight evaluation model to keep AI coaching experiences useful, safe, and measurable after launch.
A builder-focused map of the agent stack in 2025: models, tool APIs, safety posture, and deployment ergonomics.
How I converted messy customer conversations into a decision-ready roadmap copilot.
A framework for measuring whether AI features deliver repeat value instead of one-time novelty.
A practical operating system for moving fast on AI features while preserving user confidence.
What DeepSeek-V3 and R1 signaled about capability, inference economics, and the next competitive layer for AI products.
What the DeepSeek release window signaled about Chinese AI: faster iteration, stronger open-model pressure, and a more multipolar model market.
A practical explanation of RL and RLHF, and why reinforcement ideas are back at the center of AI product quality.
After IBM, I intentionally stayed in startup environments (Esper, SmartHalo, Unyte, Bonzen) to maximize agency, user proximity, and end-to-end product ownership.
Why the November 2022 and March 2023 moments reset product expectations, and how later jumps across Claude, Gemini, and DeepSeek repeated the same pattern.
Personal Assistant
Oli's Personal Assistant
Ask about projects, experience, or how I work.
Ready
