NNuggets
BookmarksCollections
  • About Us
  • Terms of use
  • Privacy policy
  • Disclaimer
  • Copyright & Takedown Policy
  • Community Guidelines
  • Cookie Policy
  • Contact

© 2026 Nuggets

NuggetsMarket PulseCollections

On this page

Speakers & Credentials

  • Speakers & Credentials
  • 1. Executive Summary
  • 2. Chronological Table of Contents
  • 3. Detailed Thematic Summary
  • The Reference Vault
  • 4. Data & Figures
  • 5. Core Frameworks & Mental Models
  • 6. Anecdotes
  • 7. References & Recommendations
  • 8. The Bottomline (by AI)

On this page

  • Speakers & Credentials
  • 1. Executive Summary
  • 2. Chronological Table of Contents
  • 3. Detailed Thematic Summary
  • The Reference Vault
  • 4. Data & Figures
  • 5. Core Frameworks & Mental Models
  • 6. Anecdotes
  • 7. References & Recommendations
  • 8. The Bottomline (by AI)
Technology/May 23, 2026/14 min read/youtu.be

Claude Code Head Boris Cherny: Insane Growth, Tokenmaxxing, AI Agents' Next Frontier | Big Technology Podcast | Alex Kantrowitz

Source
Source
Watch on YouTube ↗

"I don't write code I prompt Claude And actually nowadays mostly what I'm doing is I have a Claude that prompts other quads So I don't even talk to Claude I have a Claude that's talking to my quads" - Boris Cherny [00:00:25]

"Before models... you would work for a really long time and you would see maybe like a one, two, 3% improvement in productivity per engineer over the course of a year... the amount of code written per engineer at Anthropic has grown something like 250% since we introduced quad code" - Boris Cherny [00:15:15]

References

  1. Original source (youtu.be)

Disclaimer: Orignal content owned by or sourced from third parties. It does not represent the views of 'Nuggets' platform or it's team. AI is used extensively across this platform including for summaries. Accuracy is not guaranteed, there can be mistakes. Any info or content on this platform is not a financial, legal, or investment advice. Do your own research. Refer for complete disclosures:- Terms of Use · Full Disclaimer

Related nuggets

Jun 2, 2026

AI Is Escaping the Screen | 01 Jun 2026 | Coatue

Coatue : AI is entering a new phase: moving beyond digital tools and into fully autonomous systems operating in the physical world. From advanced manufacturing and surgical robotics to robots in the home, the next wave of innovation will b…

Jun 2, 2026

Kalshi Monthly Volume - Politics ($M) | Chart of the Day | Coatue

Coatue: Kalshi's political volume has scaled dramatically, and the American Power Index KPOW is what that scale enables: a single number gauge of the current balance of political power and where markets expect it to move, which Kalshi bill…

Jun 2, 2026

The BlackBerry Problem |18 May 2026 | The Mistakes Series | Malcolm Gladwell's Revisionist History

"My mistake and naivity was to think that people are were with me so you're flying around the world you're trying to get people on side and you think they're on side but they're not mhm mhm and you get blindsight" Jim Balsillie 00:01:34 ht…

Jun 2, 2026

Partnership Perspectives: Network International | 2 Jun 2026 | Brookfield Perspectives

Actions

Reading

Published
May 23, 2026
Read time
14 min read
Progress0%

"Claude code is 100% written by Claude code Co-work is 100% written by quad code An increasing number of features are fully written by quad code across anthropic and products" - Boris Cherny [00:27:13]

"Network effects... that's still going to matter Some modes get less important And this is for example switching costs because if you want to switch from vendor A to vendor B you can you know you can just ask Quad to do that" - Boris Cherny [00:46:56]

"If you throw away your filing cabinets... and you put a computer at the center of it and that's the way that you do all your business process then you benefit." - Boris Cherny [00:21:34]

"We recently shipped this thing called auto mode... whenever Claude wants to use a tool it asks another Claude is it safe to use this tool" - Boris Cherny [00:38:00]


Speakers & Credentials

  • Alex Kantrowitz: Host of the Big Technology Podcast. Technology journalist providing critical analysis of Silicon Valley trends, business models, and AI agent developments.
  • Boris Cherny: Head of Claude Code at Anthropic. Former engineer at Meta/Facebook focused on codebase health. Currently leads the development of Anthropic's agentic tools including Claude Code and Claude Co-work.

1. Executive Summary

  • Anthropic is experiencing an unprecedented, exponential surge in demand, largely driven by the mass adoption of its agentic products like Claude Code and Claude Co-work, causing demand to surge up to 80x year-over-year.
  • The fundamental paradigm of AI has shifted from reactive auto-complete generation to proactive, tool-using agents capable of operating browsers, modifying local desktop files, connecting to platforms like QuickBooks, and navigating complex workflows autonomously.
  • Enterprise productivity metrics are shattering historical norms; where companies historically fought for 1-3% annual engineering productivity gains, Anthropic is internally tracking a 250% increase in code output per engineer.
  • Anthropic is solving agentic safety and user fatigue through "Auto Mode," a novel system where LLMs independently evaluate and grant safety permissions to other LLMs, allowing users to run thousands of parallel agentic instances without constant human oversight.
  • The proliferation of capable AI agents will fundamentally reshape software moats, likely destroying switching costs while significantly amplifying the value of traditional network effects and deep infrastructure scale economies.

2. Chronological Table of Contents

  • [00:00:00] Introduction & The Scale of Anthropic's Hypergrowth
  • [00:06:19] Defining Claude Code & The Shift to Tool-Using Agents
  • [00:10:00] Real-World Capabilities: Travel Booking & Iterative Expectations
  • [00:13:36] "Token-Maxxing", Gamification, and Enterprise Adoption
  • [00:22:28] Model Efficiency, Looping Behavior, and The "Effort" Parameter
  • [00:26:25] AI Writing AI: The Self-Replicating Codebase
  • [00:31:28] Infrastructure Limits, Rate Limits, and Elon Musk's Colossus
  • [00:37:20] The Future Roadmap: Auto Mode & Parallel Agent Execution
  • [00:45:26] The SaaS Apocalypse & The 7 Powers of Business Moats
  • [00:50:09] AGI Timelines, Self-Improvement, and the World Model Debate
  • [00:54:33] Real-World Extrapolation: Non-Engineers Building Software

3. Detailed Thematic Summary

The Scale of Anthropic's Hypergrowth & The API vs. Product Mix

  • Anthropic CEO Dario Amodei recently noted that demand for Anthropic products has increased by approximately 80 times year-over-year [00:01:27]. Kantrowitz notes that while the company celebrated a $4 billion ARR run rate a year ago, current estimates place it closer to $40-$45 billion [00:01:42].
  • The product growth for Claude Code hit an explosive inflection point with the release of Opus 4 and Sonnet 4 last May [00:02:42]. This momentum carried continuously through subsequent drops of Opus 4.5 in November, 4.6 in February, and the current 4.7 [00:02:55].
  • While API usage remains massive, the share of usage coming from direct consumer/enterprise interfaces (Claude Chat, Claude Code, Co-work) has accelerated dramatically, proving the value of AI labs developing their own product ecosystems to control safety and capture mindshare [00:04:14].

Defining Claude Code & The Shift to Tool-Using Agents

  • The crucial differentiator between early chatbots and Claude Code is tool access. About 18 months ago, Claude developed the capacity to actively engage with local environments, meaning it could edit desktop files, organize folders, and manipulate browsers natively [00:07:19].
  • This transition moved the industry from Gen 1 "autocomplete prediction" (predicting the next word or line of code) into full agentic workflow automation where natural language drives physical software actions [00:08:38].
  • Boris Cherny personally utilizes Co-work to execute complex logistical workflows. Using Opus 4.7, he commanded Co-work to review his emails, cross-reference his calendar, and book 8 flights and 5 hotels for an upcoming multi-city tour (London, Tokyo, etc.). The model autonomously corrected two date errors and found two missing stops before successfully executing the credit-card bookings [00:10:06].
  • Agentic capabilities are continuously expanding into core business functions; this week, Anthropic rolled out capabilities allowing Claude Co-work to take over QuickBooks for small business bookkeeping [00:37:05].

"Token-Maxxing", Gamification, and Enterprise Adoption

  • There is a developing Silicon Valley phenomenon known as "token maxxing," prominently highlighted at Amazon, where managers mandate 80% AI adoption targets, leading engineers to spin up useless AI instances just to inflate their token usage metrics [00:18:44].
  • Cherny advocates against gamified token-maxxing. Drawing on his past tenure monitoring codebase health at Meta/Facebook, he notes that historical productivity gains were painstakingly won at a rate of 1-3% per year [00:15:08].
  • By stark contrast, since the rollout of Claude Code, Anthropic engineers are writing 250% more code per engineer without any degradation in codebase health or reliability [00:15:40].
  • Cherny references a classic 1990s Harvard Business Review article about the PC revolution: companies only realize productivity gains when they throw out the "filing cabinets" and put the new technology strictly at the center of their business process, rather than keeping it on the periphery [00:21:00].

Model Efficiency, Looping Behavior, and The "Effort" Parameter

  • Users frequently encounter models getting stuck in inefficient "loops," burning massive token counts to complete simple tasks (like exporting a PDF) [00:23:25].
  • Anthropic's core development philosophy is to optimize strictly for intelligence first, and efficiency second [00:24:30].
  • To manage this natively, Anthropic introduced the "Effort" parameter. Users running Opus 4.7 can set effort to "Extra High" or "Maximum" for pure intelligence, or throttle it down to "Medium" or "Low" to forcibly conserve tokens and limit runaway looping [00:25:26].

AI Writing AI & The Extensibility Ecosystem

  • In a milestone for recursive capability, Claude Code is now 100% written by Claude Code, and the sister product Co-work is also 100% written by Claude Code [00:27:13].
  • During a recent Y Combinator talk, roughly 50% of the audience confirmed that 100% of their codebase is now written using Claude Code, with only a single participant claiming 0% AI adoption [00:27:32].
  • Addressing rate-limit complaints, Anthropic recently doubled baseline limits. Cherny confirmed that only a tiny fraction of Pro/Max users ever actually hit the ceiling, but power users are now running hundreds or even thousands of Claude instances in parallel overnight, pushing the bleeding edge of the infrastructure [00:32:17].
  • Crucially, Anthropic deployed fresh compute capacity leveraging Elon Musk's Colossus cluster to handle the unprecedented and unexpected scaling requirements [00:35:22].

The Future Roadmap: Auto Mode & Parallel Agent Execution

  • Anthropic has launched "Auto Mode" to solve permission fatigue. Historically, agents paused at every step to ask the human for execution permission, causing "always allow" security vulnerabilities [00:38:00].
  • In Auto Mode, a primary operating Claude routes permission requests to a secondary, independent Claude that acts purely as a safety validator. This LLM-to-LLM verification has proven fundamentally safer across thousands of benchmarks than relying on fatigued human approvals [00:38:46].

The SaaS Apocalypse & The 7 Powers of Business Moats

  • Addressing the "SaaS Apocalypse" (the theory that AI will write all bespoke software and kill standard SaaS subscriptions), Cherny applies the "7 Powers" economic framework [00:45:52].
  • He predicts that Switching Costs will evaporate as an enterprise moat, because users can simply instruct Claude to autonomously migrate their entire system from Vendor A to Vendor B [00:46:56].
  • Conversely, Network Effects (e.g., messaging protocols like Signal) and Scale Economies (e.g., TSMC's capital-intensive manufacturing or massive AI infrastructure) will increase massively in value, as AI cannot hallucinate a physical hardware fab or a pre-existing user network [00:46:24].

AGI Timelines, Self-Improvement, and the World Model Debate

  • Cherny generally agrees with Anthropic founder Jack Clark's assessment of a 60% probability that models will be fully self-improving by 2028. Currently, Claude Code actively suggests what features Anthropic should build next, though a human still executes the master prompts [00:50:09].
  • Countering Meta Chief AI Scientist Yann LeCun's claim that LLMs cannot achieve agentic reliability without physical "World Models," Cherny notes that the simple act of predicting the next token fundamentally forces the LLM to learn advanced planning and consequence mapping, pointing to Anthropic research observing models planning the end of a poem before writing the first line [00:54:02].
  • Proving that this agentic future is not just a "fever dream" for engineers, Cherny highlighted a recent hackathon for Opus 4.7 where a doctor, an electrician, and a carpenter built economically useful software with zero prior coding experience, with one non-technical founder even selling their startup immediately after [00:55:25].

The Reference Vault

4. Data & Figures

Data PointValueContextTimestamp
Anthropic YoY Demand Growth80xExplosive enterprise and consumer scale over the last 12 months.[00:01:27]
Anthropic ARR Estimates~$40B - $45BUp from a highly celebrated $4B ARR run rate the previous year.[00:01:42]
Historic Engineering Productivity1% - 3%The traditional annual efficiency gains achieved by elite engineering teams at companies like Meta.[00:15:08]
Anthropic Engineer Code Output+250%The increase in raw code volume produced per engineer since the introduction of Claude Code.[00:15:40]

5. Core Frameworks & Mental Models

  • The Paradigm of Psychological Safety: When deploying AI inside an enterprise, management must avoid token leaderboards and instead provide "psychological safety." Because the largest efficiency breakthroughs often come from unexpected sources (like an accountant optimizing scripts), employees must feel safe burning tokens on failed experiments without reprisal [00:16:28].
  • Intelligence vs. Efficiency Sequencing: Anthropic's model development philosophy explicitly dictates that a model must first be pushed to its maximum raw intelligence capability regardless of compute cost, and only afterward optimized for token efficiency. [00:24:30].
  • Agentic Routing (Auto Mode): The framework of using AI to guardrail AI. To combat human "always allow" fatigue, Anthropic routes executable commands to a secondary LLM strictly tasked with safety evaluation before execution. [00:38:00].
  • The 7 Powers (Moat Restructuring): An economic framework applied to the age of AI. Software capabilities will be commoditized, destroying switching costs. True defensive moats will rely entirely on scale economies (capital-intensive compute/infrastructure) and network effects (user liquidity). [00:45:52].

6. Anecdotes

  • The Global Travel Booker: Cherny verbally dictated a massive multi-city itinerary (London, Tokyo, and multiple other stops) to Claude Co-work. Unprompted, the agent cross-referenced his Gmail and Google Calendar, identified two missing stops he had forgotten, corrected his misstated dates, and cleanly executed the payment for 8 flights and 5 hotels while he worked on other tasks. [00:10:06].
  • The PDF Rabbit Hole: Kantrowitz described Co-work spinning into an unhinged token-burning loop when trying to export a PowerPoint to PDF. After manual intervention, Claude apologized, admitting it got trapped "worrying about a constraint that wasn't actually blocking us." [00:23:25].
  • The Y Combinator Show of Hands: At a recent YC event, Cherny asked the audience of elite startup founders who had 100% of their codebase written by AI. Half the room raised their hands. When he asked who had 0% written by AI, only a single person raised their hand. [00:27:32].
  • The Non-Technical Language Picker Fix: A non-engineer friend of Cherny's encountered a broken language input setting on her laptop. Instead of Googling the fix, she granted Co-work desktop access. She watched the screen glow orange as the agent autonomously navigated system settings, diagnosed the localized registry error, and executed the fix in real-time. [00:31:00].
  • The Opus 4.7 Hackathon Winners: To prove that AI agent enthusiasm isn't just a tech-bubble "fever dream," Cherny shared that a recent Anthropic hackathon was dominated by non-engineers. A doctor, a carpenter, and an electrician successfully built complex, economically viable applications, with one participant selling their startup immediately after the event. [00:55:25].

7. References & Recommendations

Books, Articles & Frameworks

  • Harvard Business Review Article (1990s): Referenced to explain the modern AI productivity paradox; notes that the PC revolution only yielded macro-economic gains when companies physically threw away filing cabinets and placed the computer at the core of the business process. [00:21:00].
  • The 7 Powers by Hamilton Helmer: The definitive business strategy framework regarding defensible moats (Scale, Network Effects, Switching Costs, etc.) used to predict which SaaS companies survive the agentic wave. [00:45:52].

Companies, Institutions & Tools

  • Anthropic: The AI lab behind Claude, Opus, Sonnet, Claude Code, and Co-work. Operating heavily on an AI safety mandate. [00:01:27].
  • Amazon: Highlighted critically by the Financial Times as an environment suffering from corporate "Token Maxxing," where developers write scripts to burn tokens to hit 80% AI utilization KPIs. [00:18:44].
  • OpenAI / Codex: Explicitly mentioned as a competitor offering alternative code-generation models. Kantrowitz notes users hitting rate limits on Claude sometimes threaten to migrate to OpenAI's tooling. [00:34:12].
  • Meta / Facebook: Cherny’s former employer, referenced as a traditional benchmark for immense software scale where engineering productivity was traditionally maximized for tiny 1-3% gains. [00:14:35].
  • Waymo: Kantrowitz compared the psychological experience of letting Claude take over your computer to the "white-knuckle" experience of sitting in the back of a driverless Waymo for the first time. [00:29:41].
  • TSMC: Referenced as the ultimate example of "Scale Economies"—a physical, capital-intensive moat that cannot be easily disrupted by software agents. [00:49:24].
  • Signal: Referenced to illustrate the enduring power of Network Effects; even if an AI writes a better messaging app in three hours, it lacks the liquidity of users present on Signal. [00:48:52].
  • QuickBooks: Specifically cited as a new integration platform where Claude Co-work can now autonomously handle small business bookkeeping. [00:37:05].

People

  • Dario Amodei: CEO of Anthropic. Mentioned for enforcing financial discipline regarding massive compute build-outs compared to competitors. [00:01:27].
  • Elon Musk: Referenced as the provider of the "Colossus" compute capacity that Anthropic tapped to alleviate extreme user token rate limits. [00:35:22].
  • Ethan Mollick: Wharton Professor. Cited claiming that true AGI only arrives when AI labs stop hiring "forward deployed engineers" and consultants. [00:42:07].
  • Yann LeCun: Chief AI Scientist at Meta. Referenced for his persistent argument that LLMs cannot plan effectively because they lack underlying "World Models." [00:51:58].
  • Jack Clark: Co-founder of Anthropic. Cited predicting a 60% probability that models enter recursive self-improvement phases by 2028. [00:50:09].

Concepts & Theories

  • Token-Maxxing: The counter-productive corporate practice of gamifying AI adoption, leading employees to write wasteful loops just to hit AI quota metrics. [00:13:36].
  • World Models: The theoretical AI debate on whether foundation text models need a simulated physics engine of reality to plan, or whether next-token prediction inherently develops spatial/logical reasoning. [00:51:44].

8. The Bottomline (by AI)

We have violently crossed the chasm from models that autocomplete text to agents that execute physical digital workflows—and the friction of user permission is the only remaining bottleneck. The immediate realization of productivity relies not on strict corporate AI mandates, but on granting non-technical employees the psychological safety to rip out core processes and let LLMs rebuild them organically. Watch carefully for a hyper-consolidation in software: as agents learn to autonomously migrate enterprise databases and recreate standard SaaS tools on the fly, switching costs will collapse, leaving raw infrastructure compute and pure user network effects as the only defensible moats left standing.

"Brookfield's the largest infrastructure owner in the world... We drew a pipeline and we showed all the different components of the payments ecosystem on a pipeline and said it's like a pipe that moves any commodity except what it's moving…

Amazon AI Adoption Mandate80%Internal mandate forcing engineers to use AI tokens, resulting in gamified "token maxxing."[00:18:44]
Claude Code AI Base100%The percentage of the Claude Code and Co-work codebase physically written by Claude Code.[00:27:13]
Parallel Agent Instances100s - 1,000sThe amount of concurrent Claude Code agents power users (like Boris) run nightly.[00:32:17]
AGI Self-Improvement Timeline2028 (60% Probability)Jack Clark's public estimate for when foundation models achieve autonomous, recursive self-improvement.[00:50:09]