Lenny's PodcastThe future of AI in software development | Inbal Shani (CPO of GitHub)
CHAPTERS
- 0:00 – 0:45
Cold open: Developers must shift from “writing code” to systems thinking
Inbal frames the core mindset change AI will force in software development: moving from code-centric work to understanding connected systems and architecture. She highlights how AI assistance can let junior developers focus earlier on product and environment context instead of syntax and boilerplate.
- •AI tools require a different way of thinking for software builders
- •Focus shifts from typing code to understanding systems and architecture
- •Junior developers can spend more time learning the system/product earlier
- •AI changes what “being successful as a developer” looks like
- 0:45 – 4:25
Show setup: What GitHub’s CPO sees from the center of the Copilot shift
Lenny introduces Inbal Shani and sets expectations for a wide-ranging conversation about Copilot’s end state, what’s over/under-hyped in AI for devs, and how GitHub measures success. The episode positions GitHub as a driver of the current AI tooling wave.
- •Inbal’s background (GitHub CPO; AWS, Microsoft, Amazon Robotics)
- •Episode focus: future of software development with AI
- •Under/over-hype in AI-driven developer tooling
- •How GitHub evaluates Copilot and builds for developers
- 4:25 – 6:53
Overhyped vs. underhyped: AI won’t replace developers, but testing is the sleeper topic
Inbal argues the biggest overhype is that generative AI will replace humans soon, emphasizing the need for human creativity and “human in the loop.” She calls AI-driven testing significantly under-discussed, especially as code generation accelerates and raises the testing burden.
- •92% of developers already use AI tools (AI is becoming table stakes)
- •Overhype: “GenAI will replace humans” (innovation/creativity still human)
- •AI needs humans generating and validating data/usage feedback loops
- •Underhype: AI-driven testing across unit, integration, performance, security
- 6:53 – 7:43
AI-driven testing: Scaling quality, security, and reliability as code volume explodes
The conversation drills into testing types and why AI could expand test coverage dramatically. Inbal connects faster code creation to an even greater need for automated validation, including areas teams often don’t treat as “testing” but should.
- •Testing spans unit, integration, load, infra, security, and penetration
- •AI could generate larger, more complete test suites than humans can
- •As more companies become “software companies,” testing becomes critical
- •Teams often skip important validation due to time/effort constraints
- 7:43 – 10:35
The next 3–5 years: Copilot stays a copilot, and developers level up to architecture
Inbal predicts developers will increasingly operate at the systems level, with AI handling more routine coding tasks while humans guide direction and decisions. She expects this shift to affect junior developer growth paths and increase the importance of hardware/compute optimization as AI workloads expand.
- •Core principle: “Copilot is a copilot, not a pilot”
- •Developers focus more on system design, architecture, and big-picture thinking
- •Junior devs can learn product/system context earlier with AI help
- •Hardware footprint and compute efficiency/optimization grow in importance
- 10:35 – 12:07
Copilot by the numbers: Adoption, speed, confidence, and review cycle improvements
Inbal shares adoption and impact metrics that show Copilot’s scale and perceived productivity gains. She highlights speed improvements, confidence in code quality, faster reviews, and reduced frustration—positioning these as early signals of a broader shift.
- •37,000+ organizations and 1.5M+ developers using Copilot
- •Surveyed developers report writing code ~55% faster
- •85% report more confidence in code quality; 88% less frustration
- •Code reviews completed ~15% faster; high code suggestion retention in some orgs
- 12:07 – 13:58
Why efficiency doesn’t mean fewer engineers: Reclaiming time for collaboration and innovation
Lenny raises the concern that productivity gains could drive headcount cuts; Inbal rejects this framing and re-emphasizes the need for humans in the loop. She explains how developers spend a minority of time coding today, so AI time-savings can reduce burnout and create space for higher-value work.
- •Companies shouldn’t translate gains into “25% fewer engineers”
- •Developers often spend <20–25% of time actually writing code
- •Recovered time can go to rest, collaboration, and creative problem-solving
- •Goal: happier developers, better retention, and more innovation capacity
- 13:58 – 17:04
Adopting AI without flailing: Change management and problem-first thinking
Inbal outlines common mistakes teams make when rushing into AI: expecting instant transformation and treating AI as a checkbox. She recommends working backward from real customer/workflow problems, then applying AI where it reduces friction or manual effort.
- •Mistake: “Here’s a tool—use it” without change management
- •Mistake: starting with “What should we do with AI?” instead of the problem
- •Best practice: work backward from the user’s pain and workflow friction
- •GitHub’s origin story: productivity need drove integration of foundation models
- 17:04 – 18:47
Operationalizing dogfooding: GitHub runs on GitHub before customers see anything
Inbal describes how GitHub tests and validates new capabilities internally across many functions—not just engineering. She explains the “months cooking inside GitHub” approach as a quality bar and adoption filter: if GitHub teams can’t use it, customers likely can’t either.
- •Strong dogfooding culture: “GitHub is running on GitHub”
- •Adoption spans finance, legal, HR, and product—beyond engineering
- •Internal trials validate chat, summarization, search, and workflow fit
- •Nothing ships before sustained internal use proves it works
- 18:47 – 20:24
Copilot’s product philosophy: Seamless, low-friction assistance developers actually want
Lenny calls out Copilot’s “background” magic; Inbal ties this to a deliberate design philosophy: reduce friction and avoid context switching. She describes how early collaboration between engineers, designers, and model capabilities shaped an intuitive tool that fits into real developer workflows.
- •Design goal: developers shouldn’t have to “work” to use AI
- •Lower friction increases adoption; extra steps create churn
- •Built by putting themselves in developers’ shoes and testing internally
- •Copilot aims to improve productivity, collaboration, and happiness
- 20:24 – 24:44
Measuring success: Beyond “time saved” to code quality, security, and time-to-value
Inbal explains why there’s no single metric for AI impact and warns that speed alone can produce bad code faster. She advocates a portfolio of measures—quality, security outcomes, collaboration, and ultimately developer happiness—plus business-facing “time to value.”
- •No single “one metric to rule them all” for AI productivity
- •Time saved is insufficient; it can incentivize low-quality output
- •Security-oriented AI should be measured by prevented leaks/issues found and fixed
- •Shift framing toward “time to value” tied to business outcomes
- 24:44 – 29:35
What we gain (and lose): AI as a collaboration translator, and choosing your own flavor of coding
Discussing sketch-to-app demos, Inbal sees AI as a powerful collaboration and communication amplifier more than a fully autonomous production machine. She also addresses concerns about losing joy in writing small functions, arguing developers should selectively delegate what they dislike and keep what they love.
- •Sketch-to-code is framed as improving clarity and collaboration cycles
- •AI can become a “universal conversation language” across roles
- •Developers can choose how and when to use Copilot (no single right way)
- •Analogy: past abstraction shifts (C → Java → Python) changed work, not eliminated it
- 29:35 – 32:53
From niche AI to GenAI and back to hybrid: Inbal’s long-view of what comes next
Inbal reflects on the surprising democratization of AI and predicts the future won’t be purely LLM-centric. She expects hybrid and multi-model systems, where general-purpose models are combined with specialized, safety-critical, or highly tuned niche models.
- •She came from “expert-only” niche AI; GenAI’s mass adoption was surprising
- •GenAI has limitations tied to data and generalization tradeoffs
- •Prediction: hybrid world of multiple LLMs plus specialized custom models
- •Safety-critical domains (e.g., automotive/aerospace) likely require niche models
- 32:53 – 39:20
Creating the next Copilot: Experimentation culture, customer-driven ideas, and GitHub Next
Inbal details how GitHub creates space for innovation through customer feedback loops, internal pitching, flexible resourcing, and a dedicated research org (GitHub Next). She explains what makes such teams succeed: the right mindset plus tight pathways from research to production rather than “papers only” or overly tactical demands.
- •Innovation requires bandwidth, experimentation, and comfort with failure-as-learning
- •Customer/community conversations are a primary source of product ideas
- •GitHub Next focuses on a 3–5 year horizon with applied research and POCs
- •Success factor: strong synergy with product/engineering to ship real outcomes
- 39:20 – 45:34
Leadership lessons: Becoming a CPO, influencing without authority, and learning from change mistakes
Inbal shares what helped her develop as a product leader: systems thinking, influencing skills, and learning from leaders and teams around her. In “Failure Corner,” she reframes failure as learning and recounts moving too fast when driving change—emphasizing communication, culture, and bringing people along.
- •CPO scope: business thinking, GTM, sales, engineering—not just product vision
- •Key PM/CPO skill: influencing cross-functional teams without direct control
- •Early leadership learning: explain the “why” and manage change deliberately
- •Cultural context matters; speed without alignment creates resistance
- 45:34 – 50:03
Closing + lightning round: Books, interview questions, risk-taking motto, and how to reach Inbal
Inbal closes on the excitement of her first year at GitHub and GitHub Universe, then moves into a lightning round covering recommendations, hiring questions, and personal motto. She points listeners to LinkedIn and invites shared stories and tips on how people use GitHub.
- •Book recs: Failing Forward; Good to Great (Flywheel); Dare to Lead Like a Girl
- •Favorite interview prompts: “most innovative thing you’ve done” + disagreement with manager
- •Motto: taking risks is necessary to create a future; growth requires discomfort
- •Find her on LinkedIn; she welcomes user stories and product/usage feedback
