YouTube 비디오를 PDF 또는 Kindle 준비 기사로 변환하세요.
18820자
12분 읽기
Michael Truell, CEO of Anysphere behind Cursor, discusses pivoting from AI-CAD to revolutionizing coding with AI, emphasizing superintelligent agents, taste's enduring role, and a future of magnified building productivity.
The end goal of Cursor is to replace traditional coding with a superior method, magnifying builders' abilities over the next decade and enabling massive gains by pushing frontiers faster than competitors.
Building a company is challenging, so founders should pursue projects that genuinely excite them, as demonstrated by Anysphere's shift to the future of code despite initial hurdles.
Cursor aims to invent a new programming paradigm where users describe desired software, which AI then builds, evolving from current AI-assisted coding to a higher-level abstraction.
Experienced programmers founded Anysphere, drawn to coding for its rapid building potential, but frustrated by the labor of editing esoteric languages to manifest simple ideas on screen.
Over 5-10 years, Cursor seeks to create a more productive software-building method focused on defining functionality and appearance without low-level coding details.
Current AI coding tools show early changes, especially in small codebases like startups, where users abstract above code and delegate changes to AI agents.
Vibe coding, or coding without deep code understanding, fails in large, long-term projects with millions of lines and multiple contributors due to unforeseen nth-order effects.
Cursor targets professional programmers, where AI currently writes 40-50% of code lines, but users must still review outputs, marking a shift from productivity aid to transformative tool.
Startups benefit most from AI coding initially with zero-line bases, but vibe coding is unsuitable for enduring codebases, better for short-lived experiments in early pivots.
AI coding operates in two modes: delegating tasks to agents or shoulder-surfing with tab-based interventions, both needing orders-of-magnitude improvement in the next 6-12 months.
Large language models can be viewed as human-like helpers or advanced compilers, requiring UIs that allow fine-grained control like pixel adjustments or logic edits.
As AI matures, programming languages may evolve to higher levels, reducing attention to code while maintaining written logic representations for editing.
Bottlenecks to superhuman AI agents include context windows limited to 1-2 million tokens, insufficient for 10-million-line codebases equating to 100 million tokens.
Effective attention to long contexts remains challenging, alongside continual learning for organizational knowledge, co-worker dynamics, and past attempts, lacking solid solutions.
AI progress in long-horizon tasks has advanced from seconds to about an hour in the last year, but sustained forward progress over extended periods is still limited.
Software engineering demands multimodal capabilities like running code, interacting with outputs, and tools like Datadog logs, presenting known and unknown challenges for superhuman agents.
Even with human-level coding AI, text-box interfaces are imprecise; future UIs may involve direct UI manipulation or evolved higher-level programming languages for precise control.
Models lack innate aesthetics understanding, improving via reinforcement learning on collected data rather than human-like teaching, hacking around continual learning issues.
Taste—defining what to build, both visually and logically—remains irreplaceable, as AI automates the human compilation of high-level ideas into low-level implementations.
Programming currently bundles product definition with mechanical details; AI will eliminate the latter, leaving logic designers to focus on useful, tasteful software architecture.
AI will boost professional developer productivity dramatically, accelerating large projects bogged down by existing code weight, enabling faster creation of tools like databases or AI models.
Niche software proliferation will surge, benefiting non-tech firms like biotech by providing accessible internal tools without hiring full engineering teams.
Anysphere's founders met at MIT in 2022, united by early AI projects like robotic learning and neural search engines, inspired by 2021's GitHub Copilot and scaling laws research.
Initial focus was AI for mechanical engineering CAD, training 3D autocomplete models by converting actions to method calls, but pivoted due to lack of excitement and insufficient data.
Early CAD work involved data scraping and model training at scale, providing valuable experience in infrastructure like forking Megatron LM for 10-billion-parameter models.
Pivot to Cursor stemmed from personal coding passion, observing slower progress in competitors, and belief in scaling laws predicting a radical coding transformation.
Early Cursor decision to build a full editor, not an extension, anticipated AI's dominance, drawing from GitHub Copilot's internal challenges in modifying VS Code.
Cursor reached product-market fit after 9-12 months of iteration, focusing on custom models and metrics like paid power users employing AI 4-5 days weekly.
Dogfooding—internal use of the product—drove development, emphasizing speed, reliability, and usable demos over flashy external showcases.
First 10 hires were slow and deliberate, prioritizing polymath generalists blending product, commercial acumen, and large-scale model training for high talent density.
Engineer evaluation sticks to AI-free coding screens to test core skills, avoiding bias against AI novices while teaching tool use on the job.
Maintaining hacker mindset involves onsite project trials in hiring, encouraging bottoms-up experimentation, and sectioning teams for independent launches.
Moats for AI coding tools mirror search engines: high product ceilings, distribution for data feedback loops improving models via acceptance/rejection patterns.
Inspiration from consumer electronics underscores nailing breakthrough moments like ChatGPT, racing to push frontiers for outsized gains in an intelligence era.
AI will transform coding into descriptive building, eliminating low-level labor.
Vibe coding thrives in short-term startups but fails long-term projects.
Agents and tab interventions could handle 25-30% of pro development soon.
Context windows limit AI on massive codebases needing 100 million tokens.
Continual learning for organizational context remains unsolved AI bottleneck.
Long-horizon task progress jumped from seconds to hours in recent years.
Multimodal code running and tool integration essential for superhuman agents.
Text interfaces imprecise; future needs UI manipulation for control.
Aesthetics improved via RL data hacks, bypassing human teaching methods.
Taste defines irreplaceable human role in visual and logical design.
AI automates compilation, elevating programmers to logic designers.
Productivity surges will accelerate niche software in non-tech fields.
Founders' early AI projects built ambition for knowledge work evolution.
CAD pivot taught scaling model training at 10-billion parameters.
Building full editor anticipated AI's radical coding overhaul.
Dogfooding ensures reliable, speed-focused product over demo hype.
Slow first hires create immune system for talent density.
AI-free interviews test raw skills, teach tools later.
Moats via distribution data loops like search engine improvements.
Breakthrough moments like ChatGPT yield massive gains.
Decade magnifies building for pros and accessibility for all.
Scaling laws predict predictable AI capability explosions.
Personal excitement guides pivots in hard company-building.
Polymath hires blend models, product for hybrid company.
Bottoms-up experiments sustain hacker energy at scale.
Taste endures as AI automates mechanics, focusing humans on vision.
Distribution creates self-reinforcing data moats in high-ceiling markets.
Pivoting to passion leverages personal expertise for breakthroughs.
Slow hiring builds talent density accelerating future growth.
Dogfooding bridges demo hype to reliable professional tools.
Scaling laws demand betting on relentless model intelligence gains.
Multimodal integration unlocks superhuman agents beyond text.
Niche software explosion democratizes digital physics advantages.
Continual learning solves long-context organizational memory gaps.
Editor control anticipates imprecise text interfaces' evolution.
Polymath generalists thrive in AI-software hybrid environments.
Vibe coding's limits highlight need for reviewed AI outputs.
"The end goal is to replace coding with something much better."
"Building a company's hard and so you may as well work on the thing that you're really excited about."
"Vibe coding or coding without really looking at the code and understanding it it doesn't really work."
"There are lots of nth order effects... you can't really just avoid thinking about the code."
"AI write 40% 50% of the lines of code produced within cursor."
"The game in the next 6 months to a year is to make both of those... an order of magnitude more useful."
"Even if you had something you could talk to that was human level at coding... the UI of just having a text box asking for a change of the software is imprecise."
"One thing that will be irreplaceable is taste."
"Good people will help you hit... this bar, but the truly great... hit a bar that you can't even see."
"People need to become logic designers."
"Many more pieces of niche software will exist."
"GitHub copilot was honestly the moment where... we really felt like now it was possible to make just really useful things with AI."
"Follow the scaling laws."
"What do you believe that nobody else believes."
"Building for yourself doesn't work in a lot of spaces. For us, it did."
"Optimizing for the demo... there's a long line... between... great looking demo and... a useful AI product."
"If you really nail the first 10 people... they will both accelerate you in the future."
"Programming without AI is still a really great timeboxed test for skill and intelligence."
"The market... mirrors search at the end of the 90s where the product ceiling is really high."
"This is going to be a decade where just your ability to build will be so magnified."
Pursue exciting projects despite building challenges.
Dogfood product internally for rapid iteration.
Hire slowly for first 10 to ensure talent density.
Conduct onsite project trials in hiring process.
Encourage bottoms-up team experimentation.
Teach AI tools to new hires on the job.
Monitor paid power user metrics weekly.
Avoid demo optimization, focus on reliability.
Maintain personal coding practice at scale.
Interview without AI for core skill tests.
Section teams for independent project launches.
Bet on scaling laws for long-term planning.
Conduct user interviews for domain understanding.
Train models at scale for infrastructure mastery.
Pivot based on personal domain excitement.
Use RL on data for aesthetics improvement.
Build full editors anticipating UI evolution.
Follow predictable AI progress lines.
Blend product and model work in hires.
Review AI outputs thoroughly in production.
Cursor hit $9 billion valuation, $100 million ARR in 20 months.
AI writes 40-50% of Cursor code lines.
Context windows recently reached usable 2 million tokens.
10 million code lines equal about 100 million tokens.
AI long-horizon progress from seconds to hour.
Codex training cost estimated at $90k-$100k.
Cursor processes over half-billion model calls daily.
Initial Cursor beta took 3 months from code start.
First public iteration grew slowly for 9-12 months.
GitHub Copilot beta in 2021, general availability 2022.
ChatGPT from GPT-3 via 1% training cost increase.
Anysphere founded in 2022 by MIT co-founders.
Early CAD used 10-billion-parameter models.
InstructGPT fine-tuned GPT-3 on instructions.
DALL-E launched in summer 2022.
PALM and Stable Diffusion followed in 2022.
RHF and GPT-3.5 improved without cost spikes.
Biotech firms often lack internal software tools.
Scaling laws show predictable AI improvements.
First useful AI products emerged in 2021.
Y Combinator (YC) as forefront for AI coding changes.
GitHub Copilot as inspirational first useful AI product.
OpenAI research on scaling laws for data and compute.
Codex papers as early coding model influences.
SolidWorks and Fusion 360 as CAD tools.
Megatron LM and Microsoft DeepSpeed for training.
VS Code as base for editor modifications.
Datadog for logs and tool integration.
Peter Thiel's "Zero to One" for unique beliefs.
Sam Altman's talks on betting against model smarts.
ChatGPT as iPod/iPhone-like breakthrough moment.
Apple's dogfooding process for product development.
DeepMind's work on CAD autocomplete.
Robotic reinforcement learning projects.
Neural network-based search engine competitor to Google.
Biotech company's internal software needs.
InstructGPT, DALL-E, PALM, Stable Diffusion launches.
Identify personal excitement in domains before committing resources.
Conduct deep user immersion like undercover work for understanding.
Train models at scale using forked infrastructure like Megatron.
Scrape and prepare domain-specific data for AI improvement.
Pivot when science or passion doesn't align with goals.
Build full editors instead of extensions for future-proof UI.
Iterate publicly at small scale to refine product details.
Monitor paid power user frequency as key growth metric.
Dogfood internally, focusing on speed and reliability demos.
Hire slowly, prioritizing polymath generalists for early team.
Use AI-free coding screens followed by onsite projects.
Encourage team sections for bottoms-up experimentation.
Collect acceptance/rejection data for model feedback loops.
Bet on scaling laws by planning for capability explosions.
Teach AI tools to hires, mining beginner insights.
Maintain hacker energy through passion-driven hiring.
Race to breakthrough moments like consumer electronics launches.
Review AI outputs thoroughly to catch nth-order effects.
Evolve UIs for direct manipulation beyond text boxes.
Focus on taste in defining visual and logical software.
AI will replace coding with descriptive building, magnifying human taste and productivity.
Bet on AI models getting exponentially smarter via scaling.
Build full editors anticipating radical UI evolutions.
Hire polymaths blending models and product expertise early.
Dogfood rigorously to prioritize reliability over demos.
Pivot to domains matching personal passion and data abundance.
Monitor power user metrics for sustainable professional growth.
Use AI-free interviews to test raw engineering skills.
Encourage bottoms-up experiments for sustained innovation.
Collect usage data for self-improving model moats.
Immerse in user workflows via undercover or interviews.
Focus on taste as irreplaceable in logic design.
Race toward breakthrough product moments aggressively.
Teach tools to AI novices for fresh insights.
Avoid vibe coding in long-term, complex codebases.
Train on RL data hacks for aesthetics mastery.
Plan for multimodal agents integrating code running.
Break growth rules when execution demands speed.
Follow scaling laws for predictable progress bets.
Evolve interfaces for precise, non-text control.
Proliferate niche tools for non-tech industries.
In the bustling world of AI-driven innovation, Michael Truell, co-founder and CEO of Anysphere—the powerhouse behind the Cursor AI coding platform—offers a visionary blueprint for software's future. Valued at $9 billion and achieving $100 million in annual recurring revenue just 20 months post-launch, Cursor exemplifies explosive growth. Truell, a seasoned programmer alongside his MIT co-founders, envisions replacing tedious coding with intuitive description, where AI agents construct software from high-level intents. This shift, he argues, will magnify builders' capabilities over the decade, turning ideas into reality without the drudgery of esoteric languages.
Yet, Truell tempers optimism with realism. Current "vibe coding"—delegating without deep review—falters in sprawling professional codebases riddled with nth-order effects. AI already authors 40-50% of Cursor's code, but professionals must scrutinize outputs. Bottlenecks persist: context windows cap at 2 million tokens, far short of the 100 million needed for massive repositories, while continual learning for organizational nuance remains elusive. Long-horizon tasks have progressed from seconds to hours, but superhuman agents demand multimodal prowess, like running code and parsing logs. Even then, imprecise text interfaces necessitate evolved UIs, perhaps direct screen manipulation.
Taste, Truell insists, endures as humanity's edge. As AI automates the "human compilation" of ideas into loops and variables, programmers evolve into logic designers, wielding aesthetic and structural intuition irreplaceable by models. This promises seismic productivity gains, accelerating tools from databases to AI frameworks and spawning niche software for sectors like biotech, where custom tools once required full engineering hires. The physics of digital creation, already potent, will amplify exponentially.
Anysphere's journey underscores strategic pivots. Inspired by 2021's GitHub Copilot and OpenAI's scaling laws, the team abandoned an AI-CAD venture—plagued by data scarcity—despite mastering 10-billion-parameter training. They built Cursor as a full editor, not a mere extension, foreseeing AI's dominance. Dogfooding drove refinements, hitting product-market fit after a year's iteration, tracked via paid power users. Slow hiring of polymaths forged a high-density team, blending model labs with software agility, while AI-free interviews preserve hacker ethos amid scaling.
Looking ahead, Truell sees moats in distribution-fueled data loops, akin to search engines, and breakthrough moments like ChatGPT. In this intelligence dawn, building becomes accessible to all, not just pros. "What a time to be alive," he reflects, urging founders to chase excitement in a field where following the puck—via relentless scaling—yields unparalleled gains.