Anthropic Co-founder: Building Claude Code, Lessons From GPT-3 & LLM System Design

Early Career and Founding History

• 2009 Career Start: Mirchandani joined the startup "Linked In" (linked.com) as its first employee at age 21 after graduating from MIT, preferring the high-stakes "wolf" mindset of early-stage startups over the structured tasks of big tech.
• Grouper Launch (2012): After leaving school, Mirchandani founded "Grouper," a YC-backed dating app that manually matched groups of three men and three girls for social events in bars before AI or modern swiping apps existed.
• Grouper Challenges: The startup faced competition from Tinder, which solved the same social matching problem with a superior user interface (requiring mutual consent), causing Grouper's growth to flatten and eventually decline.
• Grouper Closure: Mirchandani left Grouper in June 2014 due to burnout and lack of revenue, spending three months building an art car before pursuing AI research.
• Self-Study Period (2014): To bridge the gap between software engineering and AI research, Mirchandani took a three-month contract with Twitch to fund six months of self-study in machine learning, linear algebra, and distributed systems.
• OpenAI Entry: He joined OpenAI roughly a year after launching in 2014, initially working on the StarCraft environment and infrastructure rather than machine learning research, due to his background in distributed systems and Greg Brockman's recognition of a "paucity" of engineers with dual expertise.

OpenAI Tenure and the Scaling Laws Breakthrough

• GPT-3 Migration: Mirchandani led the critical infrastructure transition from TPUs to GPUs for GPT-3 (2018–2019), a shift driven by the superior software stack of PyTorch over TensorFlow which enabled faster iteration.
• Scaling Laws Discovery: He published a pivotal paper in 2020 observing a linear relationship between compute investment and model intelligence across 12 orders of magnitude, convincing him that brute-forcing scale would yield transformative results.
• Initial Skepticism: Early reactions to the scaling laws strategy from within and outside OpenAI included accusations of "wasting money" and criticism that the approach lacked elegance compared to algorithmic efficiency.
• Anthropic Split: A core group from OpenAI, including Mirchandani, split to form Anthropic after recognizing the high stakes of AI alignment and the need for an institution specifically built to handle the weight of transformative AI.

Anthropic Founding and Product Evolution

• Initial Team: Anthropic launched in 2021 with seven co-founders and approximately 25 engineers from OpenAI, relying on a pre-existing working culture where all internal communication occurred on public Slack channels.
• Infrastructure Focus: The first year of operations was dedicated to building training infrastructure and securing compute capacity rather than shipping immediate consumer products.
• Product Timeline:
  • A Slack bot version of "Claude 1" launched in Summer 2022, nine months before ChatGPT, but was not released as a full product due to uncertain serving infrastructure and impact theory.
  • Following the ChatGPT launch (Fall 2022), Anthropic relaunched its API and Cloud AI, with market viability becoming clear around the release of Claude 3.5 Sonnet.
• Coding Dominance:
  • Anthropic intentionally invested in coding capabilities before 3.5 Sonnet, a decision validated when 3.5 Sonnet became the default choice for coding in YC batches, capturing 20–30% market share.
  • Mirchandani estimates 80–90% of current YC founders prefer Anthropic models for coding over competitors.
  • Claude Code: Developed initially as an internal tool to assist engineers, it was released externally after demonstrating "agentic" capabilities, such as decomposing compiled binaries into editable C code in under 10 minutes.
• Evaluation Strategy: Unlike competitors, Anthropic does not maintain a dedicated team to "game" public benchmarks; instead, they rely on internal benchmarks and prioritize qualitative "dogfooding" to ensure models improve their engineers' productivity.
• Personality and Safety: A dedicated team led by Amanda Askell evaluates "personality" and safety by simulating diverse social interactions to ensure conversations leave users feeling positive, treating interpretability as a long-term necessity for future AI safety.

Infrastructure and Market Trends

• Compute Growth Trajectory: Humanity is projected to undergo the largest infrastructure buildout in history, with AGI compute spending increasing by roughly 3x annually, outpacing the combined scale of the Apollo and Manhattan projects.
• Hardware Strategy: Anthropic utilizes a multi-vendor strategy, running workloads on GPUs (NVIDIA), TPUs (Google), and Tranium (AMD) to maximize capacity flexibility, though this requires splitting performance engineering teams across different platforms.
• Power as the Primary Bottleneck: The limiting factor for AI growth is electricity supply, particularly in the US, driving policy goals to expedite data center permitting and a shift toward nuclear and renewable energy sources.
• MCP Success: The Model Context Protocol (MCP) succeeded because it was built with an "empathy for the model," treating the AI as a primary user who needs specific tooling and context to be effective.
• Developer Advice: Mirchandani advises younger engineers to take risks and pursue projects based on intrinsic pride or excitement rather than chasing extrinsic credentials like degrees or FAANG employment.