Price:

AI & TECH

Anthropic hires Andrej Karpathy to speed AI's self-improvement

Monday, May 25, 2026 · from 3 podcasts
AI Daily BriefMoonshots with Peter Diam…This Week in AI
  • Andrej Karpathy's Anthropic move targets recursive self-improvement, not just image repair.
  • Coding is fading; AI orchestrates agents via orchestration layers.
  • Labs verticalizing forces a punk software resistance.

The narrative around Andrej Karpathy joining Anthropic is shifting from public relations to practical research. The AI Daily Brief notes Karpathy is explicitly focused on using Claude to accelerate Anthropic’s own pre-training research. This is a recursive self-improvement (RSI) effort. ‘Researchers believe this loop is the key to the endgame of model IQ increases,’ host Nathaniel Whittemore observed.

TMT Long Short argues the hire signals labs are nearing this breakthrough, which would cause compute value to explode. This Week in AI’s Jason Calacanis framed the recruitment as a cynical ‘golden cage’ tactic - using equity to lock elite talent up from competitors once a company hits a trillion-dollar trajectory. The consensus is Karpathy’s technical credibility matters, but the real prize is his focus on making the AI build the next AI.

“Karpathy is focused on using Claude to accelerate the research process for future versions of itself.”

- Nathaniel Whittemore, The AI Daily Brief

The industry’s focus has moved from writing code to orchestrating agents. Google’s Antigravity 2.0 effectively hides source code, forcing developers to interact with artifacts and agent orchestrators instead, a fast-follow to tools like Cursor and Claude Code. ‘In the future, nobody will want to look at a code editor,’ argued Dave on Moonshots.

This agentic shift creates a direct war for software revenue. Startups are building orchestration layers to prevent being cannibalized by frontier labs. Kanjun Q of Imbue warns labs will eventually come for every profitable niche. Her defense is ‘punk software’: tools built to tear down walled gardens, reliant on model agnosticism to swap underlying LLMs and avoid rug pulls. Jeremy Frankel of Fundamental argues the etiquette standard is Ownership of Output - AI work remains the user’s responsibility.

“Labs will eventually come for every profitable niche. Her defense is ‘punk software’: tools built to tear down walled gardens.”

- Kanjun Q, This Week in AI

Anthropic’s API pricing penalizes third-party providers, offering a 20x token savings plan only for customers using its first-party products. This subtle anti-competitive move aims to lock users into its ecosystem while the company, now profitable five years ahead of projections with a $44 billion revenue run rate, scales on a $45 billion SpaceX compute deal. The fight is over who controls the flow of intelligence.

#Models#Enterprise#Coding
Andrej KarpathyNathaniel WhittemoreAnthropicSpaceXClaudeGoogle

Source Intelligence

- Deep dive into what was said in the episodes

The AI Daily Brief: Artificial Intelligence News and Analysis
The AI Daily Brief: Artificial Intelligence News and Analysis

Nathaniel Whittemore

Anthropic Just Reset AI Expectations • May 21

  • The order instructs the Pentagon to harden critical systems within 30 days and tasks the Treasury to establish an AI clearinghouse.
  • Nathaniel Whittemore compares the order's protocol to Anthropic's Project Glasswing, suggesting it formalizes the existing rollout approach.
  • OpenAI's 'Guaranteed Capacity' program lets enterprise customers commit to 1-3-year budgets for discounts and service certainty.
  • Nathaniel Whittemore cites Uber burning its annual AI budget in 4 months and Box kicking token strategy to CFOs as evidence of enterprise budget challenges.
  • Artificial Analysis ranks Curse's Composer 2.5 model third on its coding agent index, costing 10-60x less than Opus 47 and GBT55.
  • TMT Long Short argues the Karpathy hire indicates labs are close to recursive self-improvement, which would cause compute value to explode.
Also from this episode: (12)

Big Tech (3)

  • OpenAI expects to file IPO paperwork confidentially this Friday, aiming to begin trading by September.
  • Nathaniel Whittemore notes OpenAI's haste could change the IPO race, as Anthropic was targeting October but is assembling a final private round.
  • The White House briefed CEOs from OpenAI, Anthropic, and Google on a new AI executive order, which could be signed this Thursday.

Regulation (1)

  • The executive order proposes a voluntary 90-day government review for frontier models, but labs are pushing for a 14-day timeline.

AI & Tech (4)

  • OpenAI offers 2 million tokens to Y Combinator startups for equity, which Tyler Bosman likens to headcount cash rather than free AWS credits.
  • Anthropic hired former OpenAI co-founder Andrej Karpathy, who will build a team focused on using Claude to accelerate pre-training research.
  • Nathaniel Whittemore observes Karpathy's move signals a pattern, following Yan LeCun and John Schulman leaving OpenAI for Anthropic.
  • Bloomberg columnist Connor Sen predicts the Anthropic IPO won't be for less than $2 trillion.

Business (2)

  • Anthropic forecasts Q2 revenue of $10.9B with an annualized $44B run rate and expects its first profitable quarter with a $559M operating profit.
  • Nathaniel Whittemore notes Anthropic's profitability, initially forecast for 2029, is partly due to compute constraints limiting spending.

Chips (2)

  • Nvidia reported Q1 revenue of $81.6B, beating estimates, with data center revenue growing 92% and 46% coming from hyperscalers.
  • Nvidia CEO Jensen Huang stated the company sells zero chips in China and conceded the market to Huawei, expecting a strong year for local chipmakers.
#Models#AI Infrastructure#Big Tech#Startups
Moonshots with Peter Diamandis
Moonshots with Peter Diamandis

Google I/O 2026, Karpathy Joins Anthropic, and Cerebras’ $95B IPO | EP #256 • May 21

  • Google's AI infrastructure now processes 3.2 quadrillion tokens monthly, a 7X jump from 480 trillion last year. Gemini has over 900 million monthly users, and AI Overviews has 2.5 billion monthly users.
  • Andrej Karpathy joined Anthropic to focus on using Claude to accelerate its own pre-training research. He stated that being outside a frontier lab causes one's judgment to drift from the actual pace of development.
  • Gemini Omni is Google's new multimodal AI family capable of generating videos from text, photos, videos, and audio. Dave notes that Google DeepMind is the only remaining American frontier lab seriously pursuing video as a modality.
  • Gemini 3.5 Flash is Google's new high-throughput model, four times faster than other frontier models in output tokens per second. Alex Weizenner argues it is solidly mid-tier in raw capability compared to GPT 5.5 High.
  • Andrew Feldman says Cerebras initially solved the wafer-scale engineering problem in 2019 but sold only 12 first-gen systems. Demand exploded in late 2024 when AI models became useful, leading to a $20 billion deal with OpenAI.
  • Google's new universal cart aggregates products from YouTube, search, Gemini, and Gmail across merchants like Nike and Target. Alex Weizenner sees this as Google's attempt to compete with Amazon in retail e-commerce.
  • Notebook LM has been used to create over 1.5 billion notebooks, podcasts, and slide decks. Alex Weizenner criticizes Google's fragmented branding and urges consolidation of Spark, Flash, and Anti-Gravity under a unified AI offering.
Also from this episode: (6)

Big Tech (1)

  • Google's annual CAPEX has grown 6X from $31 billion in 2022 to an estimated $180-190 billion in 2026, while its stock price increased - a scenario most considered impossible five years ago.

Startups (1)

  • Cerebras's IPO raised $5.5 billion and closed up 68%, reaching a $95 billion market cap - the largest US tech IPO since Uber in 2019.

Safety (1)

  • Synth ID has watermarked over 100 billion images and videos plus 60,000 years of audio. OpenAI, Kakao, and Levin Labs have now adopted the standard, marking an early step in industry self-regulation of AI-generated content.

Chips (1)

  • Cerebras founders bet that AI required dedicated silicon, not GPU derivatives, and that wafer-scale chips stuffed with SRAM would deliver superior memory bandwidth. Its WSE3 chip is 15-20 times faster than GPUs for inference.

AI & Tech (2)

  • Google's Audio Glasses, launching this fall with Samsung and eyewear partners, provide all-day Gemini assistance via private audio without a display. Dave notes this will force a societal rift over pervasive recording.
  • Google launched the $2 million Build with Gemini XPRIZE hackathon to solve real-world problems impacting at least 100,000 people. Peter Diamandis says the goal is teaching entrepreneurship over seeking traditional jobs.
#AI Infrastructure#Chips#Agents#Models#Big Tech
This Week in AI
This Week in AI

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

  • 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.
  • 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.
  • 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.
  • 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.
  • 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: (6)

AI & Tech (3)

  • 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.
  • 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.
  • 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