Price:

AI & TECH

AI coding agents replace junior developers at startups willing to spend

Thursday, May 21, 2026 · from 4 podcasts, 5 episodes
This Week in AINerd SnipeAI Daily BriefThis Week in Startups
  • AI agents now autonomously handle tasks like code reviews, QA, and issue triage, replacing entry-level developer roles.
  • Venture-funded startups are spending over $1 million monthly on tokens to deploy these agents.
  • A duopoly between OpenAI and Anthropic captures 89% of all AI startup revenue.

AI has moved from a coding assistant to a direct replacement for junior engineering and QA roles. The evidence is in the budgets: OpenClaw creator Pete is spending $1.3 million monthly on tokens to power over 100 agents that autonomously manage a GitHub repository. On Nerd Snipe, hosts argued this 'token maxxing' eliminates human bottlenecks, with agents reviewing PRs, de-duplicating thousands of issues, and running security scans.

The economic pressure is intense for startups. On This Week in Startups, Jason Calacanis argued the only way to escape being a 'cost center' is to found a company, as AI automates away entry-level positions. Data from The Information reveals a stark consolidation: OpenAI and Anthropic now capture 89% of all AI startup revenue, creating a duopoly that sucks oxygen from the application layer.

"Pete is currently running a high-stakes research experiment on the future of labor. He spent $1.3 million on 603 billion tokens in 30 days to power over 100 autonomous agents that manage his GitHub repository."

- Theo and Ben, Nerd Snipe

This shift is triggering a 'SaaS apocalypse,' according to Nathaniel Whittemore on The AI Daily Brief. He notes the market narrative has flipped from questioning AI's viability to fearing its effectiveness at replacing entire software categories. Fully agentic companies like Pulsia are demonstrating the 'zero-employee' model, reaching $6 million in annualized revenue with a single founder.

Resistance is forming in the form of 'punk software.' On This Week in AI, Imbue co-founder Kanjun Q warned that frontier labs will vertically integrate into every profitable niche. The defense is building orchestration layers that can easily swap underlying models, avoiding lock-in. Cursor’s CEO declared 'wartime,' building its own model, Composer 2.5, to undercut API costs by 90% and escape dependence.

"The only way to escape a permanent underclass in the AI economy is to start a company, as being a worker for someone else makes you a cost center."

- Jason Calacanis, This Week in Startups

The capability overhang between companies is widening. Whittemore's data shows enterprise use is pivoting from time savings to new revenue streams, but adoption is uneven. While 91% of customer service departments use AI, legal and finance lag due to data quality issues. Companies that bridge this gap see compounding gains; those stuck in cost-cutting pilots fall behind.

For graduates entering this market, the mood is betrayal, not anxiety. Calacanis compared booing at commencements to 1960s anti-war protests, with students feeling drafted into an 'AI army' amid 100,000 tech layoffs. The tools they used to earn degrees are now seen as liquidating their career prospects before they begin.

#Agents#Coding#Startups
Nathaniel WhittemoreAnthropicOpenAICursorOpenClaw

Source Intelligence

- Deep dive into what was said in the episodes

This Week in AI
This Week in AI

Grads boo AI, Reese Witherspoon gets dunked + Karpathy joins Anthropic | TWiAI E14 • May 20

  • Imbue co-founder Kanjun Q bought a 10,000 H100 GPU cluster in 2022 as an investment to fund the company, which now generates substantial rental revenue. She avoided venture capital, taking investment from corporate arms and a non-profit.
  • Linear is shifting from project management to AI execution. The product now includes an agent that can research feedback, write proposals, examine codebases, and delegate tasks, with a native coding agent in development.
  • Karri Saarinen argues AI-generated design is often soulless and can worsen product quality. Founders who delegate design to AI without understanding the problem produce aesthetically pleasing but non-functional outputs.
  • Kanjun Q sees AI enabling bespoke, personalized user interfaces. She built her own agent UI for email and task management, stating that design principles shift when creating for a single user versus a mass audience.
  • Kanjun Q warns frontier AI labs will vertically integrate into profitable application layers. The defense for startups is building headless products with orchestration layers that can easily swap underlying models.
  • Jeremy Frankle calls graduating students hypocrites for booing AI while using ChatGPT for essays. He argues this is the best time to graduate, as AI is a powerful tool for creative expression and starting companies.
Also from this episode: (9)

AI & Tech (6)

  • Andre Karpathy’s move to Anthropic is more about communication than research, according to Jason Calacanis. He argues Dario Amodei’s grim predictions make him a poor AI spokesperson, while Karpathy’s credibility can alleviate industry pressure.
  • Anthropic’s API pricing penalizes third-party providers. They offer a 20x token savings plan only for customers using Anthropic’s first-party products, a subtle anti-competitive move aimed at locking users into their ecosystem.
  • Fundamental builds tabular models for enterprise structured data, a modality poorly handled by LLMs. They have a confidential compute partnership with AWS, allowing models to be deployed and encrypted within a customer’s own VPC.
  • Jeremy Frankle defines poor AI etiquette as shifting the burden of reviewing AI-generated slop onto coworkers. He asserts all AI output is the user’s responsibility and must be reviewed before delegation.
  • The best use of AI is as a reflective surface to ask better questions, not just a solution generator. Imbue open-sourced ‘Blueprint,’ an agent skill tuned to ask high-quality questions to gather user context.
  • AI industry leaders are forming distinct cultural cults. Jason Calacanis categorizes them: SpaceX for tech libertarian monks, Anthropic for the left-leaning and earnest, and OpenAI for cutthroat capitalists.

AI Infrastructure (1)

  • A cost-effective local AI cluster can be built by daisy-chaining multiple Apple Mac Studios with high RAM. ExoLabs provides software to address multiple units as a single cluster.

Society (1)

  • Graduates are booing AI commencement speeches due to real fear and disempowerment. They perceive a future where entry-level jobs are automated and wealth creation excludes them, reacting against condescending advice.

Media (1)

  • A New York Times editorial attacked Reese Witherspoon for encouraging AI adoption. Kanjun Q argues this conflates two separate issues: using a helpful tool versus critiquing systemic power concentration.
#Agents#Startups#Coding
Nerd Snipe with Theo and Ben
Nerd Snipe with Theo and Ben

How the OpenClaw creator uses $1.3 million of tokens • May 20

  • OpenClaw creator Pete spent $1.3 million on tokens in 30 days, making 7.6 million requests totaling 603 billion tokens, as shown in his Codex Bar utility screenshot.
  • Pete's OpenClaw project leverages unlimited tokens to automate software development, running over 100 Codex instances in the cloud to review PRs, deduplicate issues, run security scans, create patches, and even generate PRs from meeting discussions.
  • Gary Tan claims token maxing with tools like OpenClaw and G-Brain at $10k per month provides a competitive advantage, offering 2028-level AI capabilities years early for those willing to invest heavily now.
  • Anthropic banned all non-interactive Claude Code usage, reclassifying tools like T3 Code and the Claude -p flag as 'programmatic' and removing their subsidized rates, effectively enforcing use only through their official interface.
  • The Bun runtime was rewritten from Zig to Rust using AI in about a week and a half, merging into main with all tests passing, representing a massive, agent-driven transformation of a critical JavaScript ecosystem project.
  • Theo spent approximately $10k-$12k on networking hardware, including a Synology NAS and drives, to build a secure private network for running AI agents, motivated by the increasing frequency of critical security vulnerabilities.
  • AgentMail provides email inboxes as an API for AI agents, enabling use cases like automated notifications, site sign-ins with 2FA, and customer service, becoming a critical primitive for Theo's agent projects.
  • Clerk offers a unified platform for user authentication, organization management, and billing, with components that deeply link subscription and payment management directly into application interfaces.
Also from this episode: (3)

AI Infrastructure (1)

  • Mark Cuban proposed a federal token tax of less than 50 cents per million tokens to incentivize efficiency, reduce energy strain, and generate revenue, but Theo argues it would disproportionately impact cheaper models and harm US competitiveness.

Enterprise (2)

  • Hashimoto warns entire companies suffer from AI psychosis, prioritizing rapid bug fixes over system resilience, which risks creating incomprehensible, decaying architectures as seen in infrastructure's shift to cloud automation.
  • A severe two-week wave of security exploits included Copy Fail, 70 patched CVEs in macOS, a Windows BitLocker bypass, the Minishai Halad supply chain attack, and Google-confirmed AI-powered exploitations of zero-days.
#Agents#Coding#Big Tech#Safety#Enterprise
The AI Daily Brief: Artificial Intelligence News and Analysis
The AI Daily Brief: Artificial Intelligence News and Analysis

Nathaniel Whittemore

9 Codex Tips From the Codex Team • May 19

  • Cursor's Composer 2.5 coding model matches frontier model performance at a tenth of the cost, scoring 69.3% on Terminal Bench 2.0 and 79.8% on SweBench multilingual.
  • Cursor prices Composer 2.5 at 50 cents per million input tokens and $250 per million output tokens, half the cost of Opus 4.7 or GPT-5.5.
  • Jason Lou's first Codex tip advocates using durable, long-running threads for key work streams, relying on OpenAI's improved context compaction to maintain persistent memory.
  • Lou argues voice interaction with Codex unlocks richer context by allowing users to provide messy, uncertain backstory, letting the AI help clarify thoughts.
  • The steer feature in Codex lets users update prompts mid-execution, enabling parallel human-AI work instead of rigid turn-based prompting.
  • Lou built a structured Obsidian file vault for Codex memories, arguing work should leave behind inspectable artifacts, not just trapped chat history.
  • Codex's tool use - computer, browser, and connectors - transforms it from a chat interface into an evidence-gathering work system that needs full environmental access.
  • Lou uses heartbeats, or scheduled check-ins, to create autonomous feedback loops where Codex monitors tools like Slack and triggers actions without human intervention.
  • The side panel in Codex is where parallel processing happens, allowing users to inspect and annotate artifacts while the agent continues working.
Also from this episode: (4)

AI & Tech (4)

  • Cursor is training a new model from scratch using XAI's Colossus 2 cluster, which has a million H100 GPU equivalents.
  • Chamath Palihapitiya argues enterprises using OpenAI or Anthropic directly are letting competitors into their data, creating an opening for model-agnostic harness-first companies.
  • Cloudflare's review finds Anthropic's secretive Mythos model can create multi-step exploit chains and generate functional exploit proofs, working like a senior security researcher.
  • A jury dismissed Elon Musk's lawsuit against OpenAI in two hours, ruling his breach of charitable trust claim was barred by a three-year statute of limitations.
#Models#Startups#Safety#Agents#Enterprise

AI Inequality • May 17

  • Nathaniel Whittemore defines Q2 2026 as the AI 'second moment,' shifting from chatbot assistants to workable agentic systems, with stakes marked by billions of weekly users and $650 billion in planned capex.
  • Enterprise AI shifted from pilots to production, with 40% of enterprises predicted to have working agents by end of 2026. Pulsia, a fully agentic company, reached $6 million annualized revenue with a single founder.
  • Nathaniel Whittemore argues the capability overhang - the gap between AI's potential and deployed value - is widening, increasing the disparity between leading and lagging companies.
Also from this episode: (8)

Enterprise (1)

  • Claude Code revenue grew from $1 billion to $2.5 billion annualized in two months. Anthropic's enterprise share reached 70% of first-time buyers, and its overall revenue run rate hit $19 billion.

Models (1)

  • The quarter saw rapid frontier model releases: GPT-5.2 Codex, Genie 3, Opus 4.6, GPT-5.3 Codex, Sonnet 4.6, Gemini 3.1 Pro, Nano Banana 2, and GPT-5.4, with no single benchmark winner across common tests.

Open Source (1)

  • OpenClaw became the most starred open-source project on GitHub. Nvidia launched Nemo Claw as an enterprise-grade wrapper, and Anthropic integrated its features into Claude Code and Claude Co-work.

AI & Tech (5)

  • Survey data shows 71% of practitioners used vibe coding, 62% used agentic automation, and the average respondent uses 3.5 models. ROI shifted from time savings (13.6% of use cases) to increased output and new capabilities.
  • Customer service AI adoption is mature with 91% of businesses experimenting, but 64% of customers prefer no AI in interactions. Legal AI adoption lags, with only 15% of tasks using AI despite 80% capability.
  • HR AI deployment grew 320% in 12 months, from 19% to 61%. Finance AI adoption faces data quality obstacles, with 91% of firms reporting low impact.
  • Generative Engine Optimization (GEO) market was under $1 billion in 2025, projected to reach $34 billion by 2034. Sales AI use cases are mature, with 63% categorized as 'primetime' for most organizations.
  • Anthropic and the Pentagon clashed over terms for Claude's use, leading to Anthropic being designated a supply chain risk and a subsequent lawsuit. ChatGPT's agreement with the Department of War triggered a 775% surge in one-star reviews.
#Agents#Enterprise#AI Infrastructure
This Week in Startups
This Week in Startups

Why is Gen Z hates AI? • May 18

  • Calacanis highlights that tech companies cut 100,000 jobs in the first part of the year, a figure already surpassing the total data center jobs projected for 2030.
  • Calacanis believes the only way to escape a permanent underclass in the AI economy is to start a company, as being a worker for someone else makes you a cost center.
  • Jason Calacanis advocates for 'delightful scale' companies making $500k to $5M annually as a viable path, arguing necessity and modern tools like ChatGPT lower the barrier to entry.
  • The Information reports Anthropic and OpenAI generate 89% of all AI startup revenue, creating a duopoly that is consolidating power.
  • Hosts note that six months ago, Anthropic and OpenAI represented 4.5% less of total startup AI revenue, indicating they are rapidly gaining market share.
  • Calacanis argues AI model providers like OpenAI and Anthropic are likely selling tokens at a massive loss currently, similar to Uber and Lyft's early subsidized growth phases.
  • Hosts discuss the risk for application-layer AI startups, as major labs gaining access to their usage data could quickly build competing features and erase their market.
Also from this episode: (9)

AI & Tech (5)

  • Jason Calacanis argues college graduates fear AI not out of simple anxiety but from feeling betrayed by tech leaders like Eric Schmidt, who they believe have bad intent about job displacement.
  • Calacanis says this generation has used ChatGPT for two years to complete degrees and understands the technology well, which fuels their cynicism about future job prospects.
  • Hosts note a stark gap between business excitement about AI and average consumer sentiment, particularly among recent graduates entering the workforce during economic uncertainty.
  • Jason Calacanis advises against using AI alone to file patents due to the high-stakes, specialized legal nature of IP, though it can be used with a human-in-the-loop for initial work.
  • Calacanis believes running AI models locally on hardware like Mac Studios is the future for privacy and cost reasons, not a fad, especially for sensitive corporate data.

AI Infrastructure (1)

  • A University of Utah study projected US data center construction jobs will peak and then decline, with total direct operations jobs reaching only about 65,000 by 2030.

Education (1)

  • A viral New York Times essay by Stanford senior Theo Baker claimed cheating with AI is omnipresent on campus, dissolving the foundations of liberal arts education faster than the workforce.

Society (2)

  • The shooting incident in Austin highlighted the privacy-versus-safety debate around Flock Safety's license plate readers, which the city council had removed but a neighboring county used to capture suspects.
  • Calacanis suggests potential safeguards for surveillance tech like Flock include strict audit trails, biometric access logs, and enforced data retention policies of up to 36 months.
#Startups#Labor#Big Tech#Regulation

Related Stories