No PriorsNo Priors Ep. 10 | With Copilot's Chief Architect and founder of Minion.AI Alex Graveley
CHAPTERS
- 0:06 – 1:19
Early tech roots: Linux, GPL, and the open-source ethos
Alex traces his path into tech starting as a teenager discovering Linux and open source. He explains how the GPL and the early web shaped his motivation to build in public so others could learn from his work.
- •Got started young (Linux at ~14) when the web was still early
- •Read the GPL early and internalized the value of open collaboration
- •Spent years compiling kernels and contributing to open source
- •Contrasts Linux’s early “weirdness” vs. dominant Windows/server ecosystems
- 1:19 – 2:20
From VMware to startup ambition: the “education startup” detour
After several years at VMware, Alex leaves to pursue startups—beginning with a common founder rite of passage: an education product. He describes realizing how hard education is and learning the importance of a clear, paid-for value proposition.
- •Left VMware intending to do startups
- •Built an education startup idea for ~9 months
- •Learned education is difficult to monetize and execute
- •Took away a key lesson: build something achievable people will pay for or use
- 2:20 – 2:46
Hackpad’s origin story: Burning Man logistics + forking Etherpad
Hackpad emerged from a real need: coordinating large Burning Man projects while living in a San Francisco warehouse. Alex explains forking Etherpad, recruiting a collaborator, and seeing adoption spread through communities and then startups.
- •Hackpad started as a fork of Etherpad to organize Burning Man projects
- •Real user pain (large, semi-anonymous groups) drove early product direction
- •Recruited a friend from the community to build with him
- •Early traction came from Burning Man camps before YC
- 2:46 – 3:54
Hackpad product mechanics and growth: real-time docs, YC, and early enterprise use
Alex details what made Hackpad distinct—real-time editing with attribution—and how YC became a growth lever. He shares how internal note-taking and founder adoption helped land notable customers like Stripe, which used Hackpad as an early knowledge base.
- •Real-time text editor like Google Docs, with strong attribution/traceability
- •Useful for large groups where ownership and accountability are unclear
- •YC strategy: get batch companies to adopt; used it for widely read notes
- •Notable long-term customer: Stripe used it as an early knowledge base
- 3:54 – 4:04
Post-Dropbox: Paper era and the persistent desire for “bots that do things”
After the Dropbox acquisition and work that became Paper, Alex describes exploring what to do next. A consistent throughline emerges: he wants to build a robot/assistant that takes action on users’ behalf.
- •Transition from Hackpad into Dropbox and Paper-related work
- •Period of exploration afterward
- •Clear personal north star: build action-taking assistants/bots
- 4:04 – 8:30
Inside Magic: ops-heavy assistants, real-world complexity, and early ML augmentation
Alex recounts working at Magic, a text-based assistant service that relied heavily on human operators rather than ML. He explains why text interfaces are hard (goals can change at the last second) and how they shipped an early model to speed up operator responses.
- •Magic was “super heavy ops,” teaching lessons about human behavior under stress
- •Text interactions are high-variance; users can change goals at the 99% mark
- •Real-world tasks break easily (delays, lost passwords, changing constraints)
- •Trained a small seq2seq-style model on chat histories to suggest responses
- •Measured productivity gains by reducing operator typing time
- 8:30 – 12:54
Crypto and identity meets AI risk: discomfort era, scams, and defensive infrastructure
The conversation shifts to crypto projects and then to AI-era identity and fraud. Alex predicts a multi-year period of societal discomfort where agents impersonate people and extract money, emphasizing mitigation systems, legislation, and stronger anti-spam/fingerprinting.
- •Worked on hCaptcha and then MobileCoin/Signal payments goals (fast, “Venmo-quality”)
- •Skeptical that identity alone will solve agent-vs-human problems
- •Predicts near-term harms: impersonation, scams targeting vulnerable people
- •Calls for broader solutions beyond model labs: systems + enforcement + policy
- •Notes mitigations like account creation friction, spam filtering, and fingerprinting
- 12:54 – 14:33
A personal turning point and joining GitHub: recovery, return to work, and Codespaces
Alex shares a major health event discovered during an attempted kidney donation for his father. After recovery and good outcomes for both, he re-enters work via GitHub—starting with foundational engineering like using GitHub’s own product to build GitHub.
- •Health scare led to major lung surgery in 2018; long recovery
- •Positive outcome: cancer-free for years; father received a transplant
- •Returned to work by reaching out to a friend for a role
- •Helped GitHub use its own tooling (Codespaces) to build GitHub
- 14:33 – 16:20
Copilot’s beginnings: the OpenAI/Microsoft deal context and the “small model” drop
Alex explains how GitHub’s collaboration with OpenAI emerged amid Microsoft’s supercomputer negotiations. OpenAI provided an early code-tuned model—initially weak but promising—and the team began probing what it could do and how to evaluate it.
- •Origin tied to broader OpenAI–Microsoft infrastructure deal dynamics
- •OpenAI shared an early, small code-trained fine-tune as a test artifact
- •Early model quality was poor and limited (e.g., Python-only at first)
- •Code in training may improve reasoning due to linear, structured nature
- 16:20 – 18:52
Building the evaluation harness: from <10% test pass rates to major gains
The team constructed systematic tests: internal problems unlikely to be in training data, plus “in-the-wild” repo testing by zeroing function bodies and asking the model to regenerate them. Alex describes the iterative climb in success rates as data, prompting, and training strategies improved.
- •Crowdsourced internal Python problems to avoid training-set leakage
- •Automated repo-based tests: identify called functions, blank bodies, regenerate, rerun tests
- •Early pass rates were very low (<10%), but progress was clear
- •Improvements came from better prompting and better/more comprehensive training data
- •Shifted from latest-code-only to leveraging diffs and multiple versions over time
- 18:52 – 22:58
Finding the right product shape: failed UIs, VS Code extension, and ghost text breakthrough
Early thinking leaned toward a Stack Overflow competitor, but practical UX experiments in VS Code changed the trajectory. After trying many UI metaphors, the team landed on “ghost text” completions that match how programmers think in blocks, making suggestions easy to judge quickly.
- •Early product idea floated: Stack Overflow-like Q&A experience
- •Tried many VS Code UI patterns (lists, buttons over functions, side panels) with weak engagement
- •VS Code extension/autocomplete became a fast way to test model iterations
- •Key inspiration: Gmail-style free-text autocomplete as a model for interaction
- •Ghost text + tab-to-accept + multiline completions matched “block-based” coding cognition
- 22:58 – 27:58
Shipping reality: skunkworks execution, latency as the killer metric, and global scaling
Alex describes Copilot as a fast-moving skunkworks that had to coordinate with VS Code and Azure under tight timelines. They learned a decisive lesson: speed dominates usage, with tiny latency increases materially reducing completions—driving the push for more regions and infrastructure.
- •Skunkworks project required internal politicking to get platform support
- •Moved from idea to public launch in under a year
- •Discovered speed was paramount (e.g., ~10ms changes impact completion usage)
- •India usage revealed network latency issues from a single Texas data center
- •Accelerated Azure deployment and added regional GPU capacity to reduce latency
- 27:58 – 29:58
Post-launch economics and adoption: retention, pricing debates, and inference-cost surprises
Copilot’s uptake validated the product: retention stayed extraordinarily high, and users—from novices to “whales”—generated large fractions of their code. Meanwhile, inference cost estimates were initially far off, and optimization plus infrastructure improvements made consumer pricing feasible.
- •Sustained cohort retention around/above ~50% was a strong signal
- •Uncertainty was less about value and more about willingness-to-pay and cost
- •Inference cost projections started high (e.g., ~$30/user/month) then dropped markedly
- •Pricing strategy debate: go cheap to capture market vs enterprise-only feasibility
- •User impact varied widely, with some users reporting extremely high code-generation share
- 29:58 – 31:56
Looking ahead: toward agentic software development and new verification workflows
Asked about the next 3–5 years, Alex highlights the psychological shift from writing code you fully control to supervising generated outputs. He argues the barriers are solvable via targeted reviews, tests, and models that learn from before/after changes—enabling increasingly agentic workflows.
- •Big transition: from certainty-by-hand to trust-with-verification
- •Selective code review for risky parts can increase automation safely
- •Models can learn from diffs and refactors to propose “polished” versions
- •Test-driven scaffolding can let agents iteratively extend a main loop
- •Belief: if society chooses this path, the technical barriers are achievable
- 31:56 – 33:21
Minion AI: Copilot for life—agents that take action with controlled access to data
Alex introduces Minion as the next step: moving from answering to acting. He describes everyday tasks (scheduling, travel, taxes, relationship maintenance) and emphasizes controlled, aligned action backed by access to user data and real-world systems.
- •Core thesis: the next phase is AIs taking actions, not just generating text
- •Examples: scheduling, booking travel, planning trips, doing taxes, contact follow-ups
- •Requires giving AIs access to information plus constrained action mechanisms
- •Frames Minion as “Copilot applied to everyday activities,” outside the editor
- 33:21 – 41:40
Why this startup now: real-world brittleness, task decomposition, and hiring for experimentation
Alex explains he returned to the long-standing “bots that do things” problem because models finally got good enough to connect to real data and handle variability. He also shares what makes the work hard (real-world changes, alignment checks), what data might matter (web interaction traces), and why hiring favors people comfortable with uncertainty and iteration.
- •Left earlier attempts because connecting AI to real-world data felt intractable then
- •Real world is dynamic: delays, price changes, inventory constraints break naive agents
- •Approach: decompose tasks like humans do—lists, queries, checks, iterative steps
- •Sees web/app interaction behavior as a massive, underutilized dataset
- •Hiring insight: “bimodal” understanding—very senior and very young adapt fastest; need experimental, metric-driven builders
