What was the biggest challenge when founding Anthropic?

Initially, Anthropic faced competition from OpenAI, which had significant funding and star power. They had to build their core training infrastructure and secure compute resources while establishing a team that already knew how to work together.

Why did Anthropic's Claude models become so popular for coding tasks?

Anthropic invested heavily in making Claude good at code, and user reactions validated this. Unlike other labs, Anthropic doesn't focus on gaming external benchmarks, leading to internal benchmarks and practical 'dogfooding' that better reflect real-world performance, particularly for coding.

What was the transition like for developers using Claude models for coding?

Many YC founders preferred Anthropic models for coding by a large margin, suggesting an 'X factor' beyond benchmarks. This could be due to Claude's personality, interpretability or the company's focus on internal evaluations and accelerating their own engineer's productivity.

How did Claude Code evolve from an internal tool to a successful product?

Claude Code was initially an internal tool for Anthropic engineers. The team was surprised by its success as a product, attributing it partly to a mindset shift where Claude itself was considered a user, equipped with the right tools and context.

What are the biggest bottlenecks in AI compute infrastructure today?

The primary bottleneck is power, especially in the US, impacting data center construction. While accelerators are increasing, securing sufficient electricity and navigating permitting processes are major challenges for scaling AI.

What advice does Tom Brown have for aspiring AI professionals?

He advises taking more risks, focusing on intrinsic motivation, and working on projects that inspire pride. He suggests that external credentials like degrees or working at big tech companies are less relevant than pursuing impactful work.

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

Y Combinator

Science & Technology4 min read36 min video

Aug 19, 2025|132,841 views|2,187|69

YC Y Combinator

Save to Pod

Key Moments

TL;DR

Anthropic's co-founder discusses building Claude, lessons from GPT-3, and the massive infrastructure for AI.

Key Insights

Early startup experience, like at Linked and Mob, taught valuable lessons about self-reliance and problem-solving, akin to being a 'wolf' rather than a 'dog'.

The journey to AI research was unconventional, involving self-study and overcoming self-doubt, eventually leading to OpenAI and then Anthropic.

The scaling laws in AI, particularly the predictable increase in intelligence with more compute, were a pivotal realization driving significant research and development.

Anthropic's success, particularly with Claude 3.5 and its coding capabilities, was partly due to an intense focus on developer experience and internal dogfooding, rather than solely optimizing for external benchmarks.

Building AI infrastructure is now humanity's largest infrastructure buildout, facing significant bottlenecks in power and permitting, especially in the US.

Anthropic diversifies its hardware by using GPUs from three manufacturers (NVIDIA, Google, and AWS), balancing performance engineering challenges with increased capacity and flexibility.

FROM STARTUP SURVIVAL TO AI PIONEERING

Tom Brown’s early career was marked by a transition from traditional software roles to the high-stakes environment of startups. His first role at Linked, a YC company founded by friends, taught him the importance of self-driven problem-solving, contrasting it with the task-oriented nature of school. This 'wolf pack' mentality, where survival and success depend on collective hunting, was crucial for later ventures at Mob and Grouper, where he experienced both the highs and lows of scaling early-stage companies.

THE UNEXPECTED PATH TO ARTIFICIAL INTELLIGENCE

Brown’s journey into AI research was not straightforward, partly due to a less-than-stellar grade in linear algebra and initial skepticism from peers. After leaving Grouper, he spent time exploring personal projects, including building an art car for Burning Man, and then committed to six months of intensive self-study in AI. This period, focused on machine learning courses, Kaggle projects, and foundational mathematics, was essential for building the foundational skills needed to even consider contributing to the nascent AI research field.

OPENAI IMMERSION AND THE GPT-3 REVOLUTION

Securing a position at OpenAI, initially through an offer to help with engineering tasks like building a Starcraft environment, marked a significant turning point. Brown was instrumental in the engineering efforts behind GPT-3, particularly the critical shift from TPUs to GPUs and increased compute. This work was deeply influenced by the discovery and validation of scaling laws, which demonstrated a predictable increase in model intelligence with greater computational resources, a finding that solidified his belief in the transformative potential of large-scale AI.

FOUNDING ANTHROPIC: MISSION OVER PRESTIGE

The decision to co-found Anthropic stemmed from a shared concern about AI safety and a desire to build an institution capable of managing the profound implications of advanced AI. A core group, who had collaborated effectively at OpenAI, left to pursue this mission. The initial phase was characterized by limited resources compared to established players, but a strong mission-driven culture attracted dedicated talent, emphasizing that early hires were motivated by the cause rather than just financial or reputational gains.

THE EVOLUTION OF CLAUDE AND PRODUCT STRATEGY

Anthropic’s first product, a Slackbot version of Claude 1, was developed in mid-2022, predating ChatGPT. The decision to hold back its public launch was driven by uncertainty about its societal impact and underdeveloped serving infrastructure. The company’s trajectory shifted significantly with the emergence of ChatGPT, leading to the relaunch of their API and Claude AI. It wasn't until Claude 3.5 and its strong coding capabilities that Anthropic saw clear product-market fit and began to experience substantial growth.

CODING EXCELLENCE AND DEVELOPER FOCUS

Claude's exceptional performance in coding tasks, particularly evident in benchmarks and adoption by YC startups, is attributed to Anthropic's internal focus and investment in this area, rather than just optimizing for public benchmarks. The development of Claude Code as an internal tool highlighted the potential of AI as a co-pilot for engineers. Anthropic prioritizes an API-first approach, believing that developers will build innovative applications on their platform, and encourages exploration in areas like AI coaching for business tasks.

MANAGING HUMANITY'S LARGEST INFRASTRUCTURE BUILDOUT

Anthropic is currently managing what's described as humanity's largest infrastructure buildout, with spending on AI compute projected to triple annually. This massive expansion faces significant bottlenecks, particularly in securing adequate power and navigating permitting processes for data centers, especially in the United States. The demand for compute far outstrips supply, creating a critical challenge for continued AI development and deployment, even as new hardware startups emerge with novel accelerator solutions.

STRATEGIC HARDWARE DIVERSIFICATION AND ENGINEERING

Anthropic employs a unique strategy by utilizing GPUs from three different manufacturers (NVIDIA, Google's TPUs, and AWS's Trainium), a departure from the industry norm. While this complicates performance engineering by splitting teams, it significantly enhances flexibility. This approach allows Anthropic to leverage greater overall hardware capacity and select the most appropriate chips for specific tasks, distinguishing between those optimized for training versus inference, thereby maximizing efficiency across their vast computational needs.

ADVICE FOR THE NEXT GENERATION OF AI INNOVATORS

For aspiring individuals, particularly students uncertain about their career path in AI, Brown advises taking more risks and pursuing work that aligns with intrinsic motivation and passion. He suggests focusing on endeavors that would impress peers or a more idealized version of oneself, rather than solely chasing external validation like degrees or jobs at traditional tech giants. This mindset shift emphasizes long-term impact and personal fulfillment over short-term credentials.

Mentioned in This Episode

●Products

●Software & Apps

●Companies

●Organizations

●Studies Cited

●Concepts

●People Referenced

Anthropic Co-founder's Career & AI Insights

Practical takeaways from this episode

Do This

Embrace the 'wolf' mindset: be proactive and hunt for your own solutions rather than waiting for tasks.

Take risks and pursue work that your future or idealized self would be proud of.

Focus on intrinsic motivation and building impressive things, rather than solely chasing external credentials.

When founding a company, prioritize building the core infrastructure and securing necessary compute.

Consider building tools that empower AI models as users, not just as tools for humans.

Leverage multiple hardware platforms (GPUs, TPUs, Traniums) for flexibility in compute.

Prioritize building a robust and reliable software stack for fast iteration and experimentation.

Avoid This

Don't get stuck in a 'dog waiting to be fed' mentality; be proactive.

Don't shy away from challenging technical problems, even with a less-than-stellar academic record (e.g., B- in linear algebra).

Don't be afraid to pivot or learn new fields, even if it seems unconventional (like AI research in 2015).

Don't solely focus on external benchmarks; prioritize internal evaluations and practical use cases ('do the stupid thing that works').

Don't hesitate to launch products if the infrastructure isn't perfect; learn and adapt.

Don't underestimate the value of direct user empathy (e.g., understanding Claude as a user) in product development.

Common Questions

Tom Brown started by joining early-stage startups, embracing a proactive 'wolf' mindset. He then co-founded his own startup, explored AI research through self-study after gaining experience at companies like Mopub and Grouper, and eventually joined OpenAI before co-founding Anthropic.

Topics

Compute Infrastructure

Mentioned in this video

personDanny Hernandez

Co-authored a paper with Tom Brown showing the impact of algorithmic efficiency on AI progress.

companySolidStage

A startup co-founded by Tom Brown in 2012 focused on DevOps solutions before Docker existed, aiming to be a more flexible Heroku.

personAmanda Askell

Leads the team at Anthropic responsible for evaluating model personality and ensuring it acts as a 'good world traveler'.

companyLinked

Tom Brown's first startup experience after graduating from MIT, where he learned the value of independent problem-solving.

productTranium

A type of hardware accelerator used by Anthropic, alongside GPUs and TPUs, to provide flexibility in compute resources.

companyGrouper

A dating app co-founded by Tom Brown that aimed to facilitate introductions in social settings, ultimately outcompeted by Tinder.

companyMopub

A mobile advertising company where Tom Brown worked as an engineer after his initial startup experience.

softwareHeroku

A cloud platform mentioned as inspiration for SolidStage, highlighting the complexity of building similar services without containerization.

organizationMIRI

Mentioned as a potential place for AI research work alongside DeepMind and Google Brain.

Found this useful? Build your knowledge library

Get AI-powered summaries of any YouTube video, podcast, or article in seconds. Save them to your personal pods and access them anytime.

Try Summify free