Price:

AI & TECH

Cherny declares coding solved as AI agents hollow out junior roles

Monday, June 22, 2026 · from 5 podcasts
Lenny's PodcastPodcasting 2.0Peter McCormack ShowThis Week in AIBehind the Bastards
  • Claude Code now writes 4% of public GitHub commits and 100% of its lead’s code, solving routine development.
  • Solo founders replace million-pound dev teams in weeks, slashing costs to AI token bills.
  • The creative ‘path’ vanishes, producing a flood of trivial projects and potential ‘vanilla slop’.

The manual labor of coding is over. Boris Cherny, head of Claude Code, hasn’t edited a line by hand since November and now ships 10 to 30 pull requests daily by managing five concurrent agents. On Lenny’s Podcast, Cherny claimed coding is “virtually solved” for his work and predicts the title ‘software engineer’ will fade, replaced by ‘builder’. A SemiAnalysis report found Claude Code authors 4% of all public GitHub commits, a figure Cherny notes is significantly higher in private repositories.

“Coding is virtually solved for my work. I haven't edited a single line of code by hand since November.”

- Boris Cherny, Lenny's Podcast

The productivity gains are structural, not marginal. Cherny reported a 200% increase in pull requests per engineer at Anthropic since integrating these agents, a leap that makes annual 2-3% gains at legacy tech firms look trivial. His management philosophy intentionally underfunds teams to force automation - assigning a three-person task to one engineer creates a cultural forcing function to “Claudify” the workflow.

This isn’t confined to AI labs. On The Peter McCormack Show, host Peter McCormack described building a comprehensive management system for his football club - complete with inventory, website, and email delivery - in 11 days. A prior agency estimate pegged the same project at 15 people, 18 months, and a £1 million budget. Fernando Nikolić, a guest on the show, runs a one-person company with a 94% profit margin using an “arsenal of agents” as his mid-level development team.

“I rebuilt a club operations system in two weeks. The old quote was 15 people, 18 months, and a million pounds.”

- Peter McCormack, The Peter McCormack Show

The collapse extends to generic software platforms. McCormack argued that once you can “vibe code” a custom CMS in days, the value of paying for Squarespace or Canva drops to zero. Nikolić called this the “chaos gap” - every business built on information asymmetry is vulnerable if AI bridges the knowledge gap. The only safe moats are physical and tedious, like installing 400 cameras for data collection.

Not everyone celebrates the frictionless build. On Podcasting 2.0, developer Dave Jones mourned his shift from writing 100% of his code for the Podcast Index to just 10% within two months. He argued the “path” - the struggle through difficulty - determines if a project is worth finishing. Host Adam Curry warned that reducing build friction to near zero floods the world with trivial projects rather than meaningful contributions.

Consensus holds that AI excels at execution but fails at creativity. McCormack and Nikolić agreed AI is a terrible writer and a worse comedian, producing sanitized “vanilla slop” that rounds off the edges of language. The risk is a highly productive but soul-crushing digital homogeny, where humans are left to focus on the one thing models can’t touch: actual creativity.

#Agents#Big Tech#Startups#Labor
Boris ChernyPeter McCormackAnthropicClaudeClaude CodePodcasting 2.0

Source Intelligence

- Deep dive into what was said in the episodes

Lenny's Podcast
Lenny's Podcast

Building the most AI-pilled engineering team in the world | Fiona Fung (Manager of the Claude Code and Cowork Teams) • Jun 21

  • Boris Cherny states that 100% of his code is written by Claude Code, with no manual edits since November 2024. He ships 10-30 pull requests daily and often runs five agents simultaneously.
  • Cherny argues coding is virtually solved for his work, and he predicts the title 'software engineer' will fade, replaced by 'builder'. He believes everyone will soon be a product manager who codes.
  • A SemiAnalysis report found Claude Code authors 4% of all GitHub commits, a figure Cherny notes is higher for private repos. The report predicts Claude Code will author one-fifth of all commits by the end of 2025.
  • Cherny says productivity per engineer at Anthropic has increased 200% since introducing Claude Code, measured by pull request volume. He contrasts this with his time at Meta, where annual productivity gains were only a few percentage points.
  • Claude Code's growth is accelerating, with daily active users doubling in the past month. Cherny built the initial prototype, called Quad CLI, as a terminal tool because the model improved too quickly for other form factors.
  • He sees the printing press as the best historical analog for AI's impact, enabling a transition from a specialized skill to a universal capability. He imagines a future where anyone can build software.
  • For using Claude Code, Cherny recommends always using the most capable model, starting tasks in 'plan mode', and experimenting with different interfaces beyond the terminal.
  • Cherny describes a three-layer safety approach at Anthropic: mechanistic interpretability to study model neurons, laboratory evals, and studying model behavior in the wild through early product releases.
  • He observes that newer engineers often use Claude Code in more advanced, 'AGI-forward' ways than veterans, who can get stuck in old mental models. He cites an example where a junior engineer used Claude to debug a memory leak faster than he could manually.
Also from this episode: (3)

AI Infrastructure (3)

  • Cherny advocates underfunding projects initially to force teams to 'Claudify' and automate work with AI. He advises leaders to give engineers unlimited tokens for experimentation, then optimize costs only after an idea proves successful.
  • Cherny advises builders to design for the AI model six months from now, not its current capabilities. He warns against over-engineering workflows, arguing the more general model will always outperform a specialized, scaffolded one.
  • The Claude Co-work agent was built in 10 days using Claude Code. It emerged from observing latent demand, as users were hacking the coding tool for non-technical tasks like analyzing genomes or recovering photos.
#Agents#Coding#Startups#Enterprise
Podcasting 2.0
Podcasting 2.0

Adam Curry

Episode 264: Podcast Plebicide • Jun 19

  • Dave mourns the abrupt shift from writing 100% of his code to writing only 10%, which happened within two months due to AI coding agents.
  • Adam argues that AI reduces build friction to near zero, which floods the world with trivial projects rather than meaningful contributions.
  • Adam describes how AI-generated art and music on No Agenda initially displaced human artists, but skilled creators later learned to use the tools effectively.
  • Dave explains Godcaster sends only one play event per user per hour, deduplicated via a ULID, to avoid issues with caching and unreliable connection data.
Also from this episode: (7)

Social Media (2)

  • Dave theorizes that isolated, private communication platforms like Slack distort attention, making single issues feel like the entire world and fueling conflict.
  • Dave argues federated platforms like Mastodon better mimic real-life conversation by allowing semi-public discussions where others can overhear and join.

Media (3)

  • Adam and Dave credit the Podcasting 2.0 project's success to its combination of Mastodon, GitHub, weekly podcast meetings, and live chat.
  • OP3 data shows Podcasting 2.0 had about 5,247 unique listeners in May, placing it in the so-called 'indie middle class' of 5k-25k monthly downloads.
  • Adam rejects the term 'indie podcaster', arguing all podcasters are independent by definition, and dismisses obsession with download metrics.

AI Infrastructure (2)

  • Dave clarifies James's proposal: full support for the transcript tag means supporting only the VTT format; anything else does not count.
  • Podping's trust system uses a plebiscite model where trusted nodes vote to add or remove other nodes from the trusted list, ensuring swarm integrity.
#Coding#Agents#Psychology#Social Media#Macro
The Peter McCormack Show
The Peter McCormack Show

Peter McCormack

#185 - Fernando Nikolić - AI, One-Person Companies & The New Economic Elite • Jun 19

  • Fernando Nikolić runs a one-person company using AI agents, achieving a 94% profit margin with an infrastructure cost of roughly $200-$250 per month on Google AI Ultra.
  • Peter McCormack rebuilt a comprehensive club operations system for his football team alone in two weeks. His previous agency estimated the same project would require 15 people, 18 months, and a £1 million budget.
  • McCormack built an automated media company infrastructure for his podcast, complete with simulated departments, weekly AI agent meetings, and a grading system for hundreds of weekly recommendations, in 11 days.
  • McCormack argues front-end UI and many SaaS tools are being commoditized to zero value, as custom software can be 'vibe-coded' with AI in days, rendering tools like Squarespace, Canva, and Grammarly obsolete for proficient users.
  • Both hosts observe AI is terrible at creative, subjective tasks like writing jokes, developing brand strategy, or generating original marketing content, which Nikolić calls 'slop' that is polluting the web.
  • Nikolić warns of a 'permanent underclass' if AI's exponential productivity gains are unevenly distributed to a wealthy elite, questioning whether those left behind will even be able to afford the AI utility bills.
  • McCormack built a system tracking 628 UK MPs, ingesting 10,915 tweets to grade their truthfulness with 81% accuracy, costing roughly £3,000 to process, as a prototype for AI-driven political accountability.
  • McCormack sees a parallel between the current 'chaos gap' of AI disruption and past upheavals like file-sharing's destruction of the music industry, driven by the collapse of information asymmetries.
  • McCormack estimates he'll spend at least $10,000 on AI inference tokens this month, primarily on Claude, but considers it 'peanuts' compared to the multi-million pound development costs it replaces.
Also from this episode: (3)

Protocol (1)

  • Nikolić argues Bitcoin has failed to achieve hyperbitcoinization and is not for everyone, requiring a radical 'Truman Show' moment of economic realization that most people are emotionally unwilling to undergo.

Society (1)

  • Both hosts note a craving for analog, high-trust experiences - like physical media, vinyl, and localized interaction - as a reaction to the overwhelming, sanitized digital landscape and AI-homogenized content.

Enterprise (1)

  • Nikolić's company, Perception, focuses on 'narrative engineering' by aggregating and sanitizing expensive, tedious data from the digital asset space, which he argues remains a moat versus easily automated software.
#Agents#Coding#Startups#Macro#Media
This Week in AI
This Week in AI

Why AI Models Aren’t the Product Any More | TWiAI Ep 18 • Jun 18

  • Jason Calacanis says SpaceX plans to acquire Cursor for $60 billion in stock, giving Cursor access to SpaceX’s massive compute resources for model development.
  • Ali Ansari argues the real product in AI is not the model but the agent layer, evaluations, harness, and UI built atop it. The frontier of intelligence will be defined by application companies owning their proprietary workflows.
  • Jason Calacanis warns that platform companies like OpenAI or Anthropic will study and eventually compete with their most successful application-layer customers, citing historical examples from Microsoft, Facebook, and Apple.
  • Ryan Daniels describes Crosby Legal as an AI-first law firm that uses AI internally to provide scalable legal services at flat rates, aligning incentives by eliminating billable hours.
  • Ali Ansari says Micro One pivoted from an AI recruiter tool to a marketplace providing pre-vetted human experts who train AI models, achieving about $300 million in ARR by April 2026.
  • Jason Calacanis notes SpaceX’s market cap hit $2.88 trillion after its IPO, making it the fourth most valuable US company. He calculates the Cursor acquisition at a 20x revenue multiple.
  • Ryan Daniels explains that Cursor once accounted for 40-50% of Anthropic’s total revenue, but Anthropic’s launch of Claude Code forced Cursor to build its own model, leading to the SpaceX deal.
  • Ali Ansari defines model distillation as using an open-source baseline model and conducting massive post-training that changes most weights, creating a distinct model without reinventing core reasoning.
  • Ryan Daniels argues only a handful of companies can build frontier models due to capital and compute constraints, creating a binary divide between model-makers and everyone else.
  • Jason Calacanis cites Claude Code’s rumored $2.5 billion revenue run rate and Cursor’s $4 billion run rate, predicting the AI coding market will reach $100 billion soon.
  • Ryan Daniels says Crosby’s vertically integrated law firm creates a proprietary feedback loop where lawyers train AI on subjective legal judgment, making human expertise more valuable as models improve.
  • Ali Ansari predicts nearly 100% of future AI data spend will focus on the agent and application layer, with orders of magnitude more agents built than base models.
  • Jason Calacanis describes his vision for an internal venture AI trained on Slack and Notion data to analyze investment history and missed opportunities, calling Slack’s corpus the ultimate dark data pool.
  • Ryan Daniels forecasts AI lawyers could match the average practicing attorney by late 2027, forcing courts to consider ethical access to AI counsel for self-representation.
  • Ali Ansari and Ryan Daniels collaborated on a multi-turn contract redlining benchmark using real lawyer negotiations to evaluate AI models, finding current models perform at only 10-20% of human capability.
  • Ali Ansari proposes an industry-led AI safety consortium where competing model companies create adversarial benchmarks for self-regulation, similar to the MPAA for movies.
#Big Tech#Coding#Agents#Startups
Behind the Bastards
Behind the Bastards

Part One: The Fake Bomb Detector Grift That Killed Hundreds • Jun 16

Also from this episode: (11)

Psychology (6)

  • Robert Evans explains the ideomotor effect as unconscious muscle movements driven by mental imagery, like participants moving a Ouija board pointer without realizing it.
  • Despite scientific debunking, loopholeism allowed brilliant minds like natural selection co-discoverer Alfred Russel Wallace to fall for spiritualist scams in 1865.
  • Robert Hare, a chemistry professor, was conned by fake mediums and built a 'spirit scope' device in 1855, believing he communicated with historical figures.
  • Clever Hans, a German horse in the early 1900s, appeared to solve math problems but was actually reading his owner's subtle cues, a phenomenon now called the Clever Hans effect.
  • Police drug-sniffing dogs often alert based on handler bias; a 2011 Chicago Tribune analysis found dogs found drugs in only 44% of alerts, dropping to 27% for Latino drivers.
  • A 2011 study by Lisa Lit tested 14 sniffer dogs; handlers told a cocaine scent was present (but wasn't) led dogs to alert, proving the Clever Hans effect in canine units.

History (4)

  • Dowsing, using a forked stick to find water or minerals, has been practiced globally for millennia, with texts from 2000 BCE and cave paintings from 6000 BCE possibly depicting it.
  • Despite being debunked, dowsing maintained professional credibility; one in eight archaeology instructors in the 1980s were favorable to the practice.
  • The spiritualism movement in the mid-19th century popularized table-turning and seances, which scientists like William Carpenter correctly attributed to the ideomotor effect.
  • Robert Evans links this history of gullibility to modern tech grifts, noting the defense industry later turned a fake golf ball finder, the Gopher, into a lethal bomb detector.

Science (1)

  • French chemist Chevrouel conducted the first double-blind test on the ideomotor effect in 1808, proving pendulum analysis was unconscious movement, not a chemical property.
#Psychology#History#Regulation#Macro

Related Stories