The Frontier

DHH argues that aesthetically beautiful software is more likely to be correct, a principle he finds true in mathematics, physics, and other domains.

DHH switched from skeptical of AI coding tools to using them extensively, driving a 180-degree turn in his workflow after a few weeks of experimentation.

AI agents allow his team to tackle internal projects they would never have started before, making engineers more ambitious and productive than ever.

He finds supervising AI agents for one hour can be highly effective and intoxicating, leading people to work harder than before.

DHH built the Linux distribution Umachi from scratch on Arch and Hyprland as a personal itch-scratching project, and it quickly gained a community.

He sees Ruby on Rails having a renaissance due to its token efficiency, making it ideal for AI agent workflows that still require human-readable code.

DHH started programming on the internet in 1994 and began building Ruby on Rails in 2003 when he chose Ruby to build Basecamp without external mandates.

He believes your unique spin on an idea matters more than its novelty, proven by projects like Rails, Kamal, and Umachi finding large audiences.

Anthropic's new 'Mythos' model is so adept at chaining together 3-5 security vulnerabilities to create sophisticated cyberattacks that the company is withholding its public release, labeling it a potential 'cyber-weapon of mass destruction'.

Anthropic's 'Project Glass Wing' gives select partners like NVIDIA, AWS, and Azure early access to Mythos to find and patch vulnerabilities before bad actors can exploit them, while also establishing a $100 million compute credit fund for system hardening.

Hosts argue the potential power of Mythos raises the prospect of nationalization, as its capabilities could be considered too powerful and dangerous for a private entity to control.

Rob May defines small language models (SLMs) as sub-20 billion parameter models that can run on high-end laptops and are improving in 'intelligence density' via techniques distilled from larger models.

Rob May's company, Neurometric, offers a 'Claw Pack' of 39 task-specific SLMs for unlimited inference at $8 per month, using automated distillation and 'harness engineering' to keep models on-task and reduce costs.

Rob May cites an AT&T case study where rearchitecting AI workloads to use frontier models for 10% of tasks and SLMs for 90% resulted in a 90% cost reduction, proving the economic case for model orchestration.

Jason Calacanis predicts the rise of hyper-specialized SLMs could lead to 'hyperdeflation,' collapsing the value of frontier models for many tasks as 'good enough' verticalized models become free or nearly free.

Hosts analyze Meta's new 'Muse Spark' model, which ranks fourth on the Artificial Analysis benchmark but criticize Meta's lack of a clear strategic vision beyond improving ad recommendations and user addiction.

Guest Gani's tool 'Death by Claude' critiques startups' defensibility by generating a 'death score' and replacement code, identifying hardware, network effects, and regulated/scientific work as key moats against AI replacement.

Polymarket prediction markets in April 2026 show a 95% chance Anthropic reaches a $500 billion valuation and only a 28% chance Mythos is released by June 30, indicating a belief in extended restricted access.

A host notes SpaceX's Mars ambition unlocked commercial opportunities like Starlink by driving down launch costs, creating a virtuous cycle where commercial profits fund NASA's ultimate goals.

Brett models SpaceX's Starlink revenue potential between $100 billion and $200 billion, driven by its unmatched up-mass capacity and the pending cost reductions of a reusable Starship.

He states SpaceX's growth constraint is satellite deployment speed, not demand. A shift to Starship could drop launch costs by an order of magnitude, massively accelerating revenue.

Brett claims the AI compute opportunity in orbit requires up to 60x more up-mass than Starlink, citing SpaceX's filings for one million AI satellites versus 40,000 for Starlink.

A host counters that SpaceX's decade-long lead in rocket reusability, with Blue Origin just landing its first orbital rocket, creates an unassailable moat that justifies its premium valuation.

Brett estimates the foundation model market could reach $2 trillion in revenue by 2030, supporting a $15-$20 trillion aggregate enterprise value for providers like OpenAI, XAI, and Anthropic.

He cites reports that OpenAI expects $250 billion in revenue and notes the entire AI agent space has roughly 1.1-1.2 billion weekly actives today, projected to reach 4-5 billion by 2030.

Nick says OpenAI's advertising business is at $100 million ARR but actual spend is lower. He projects AI could facilitate $9 trillion in global commerce by 2030, representing 25% of online sales.

A host argues Google's real advantage over OpenAI is deep integration into consumer products like Gmail and Drive, not just subsidy power, while Apple lacks a coherent integration playbook.

Brett states Uber's strategy of numerous autonomous vehicle partnerships is viable while the market remains supply-constrained at a $3-per-mile price point, but fails if prices drop to $1 per mile.

He contrasts Uber's network model with Tesla's potential robotaxi future, where consumer-owned FSD cars could supplement ride-hail supply, solving the peak-demand utilization problem.

Morgan Stanley’s spot Bitcoin ETF (MSBT) began trading with a 0.14% expense ratio, undercutting BlackRock's iShares Bitcoin Trust (IBIT) at 0.25%.

Anthropic's AI model Claude Mythos Preview can autonomously find and exploit decades-old software vulnerabilities for under $50 in compute, raising potential security risks for open-source DeFi protocols.

OpenSats announced its sixteenth wave of grants, funding projects like Nostr Mail, a decentralized email system built on the Nostr protocol.

Actress Mila Jovovich, in collaboration with Bitcoin coder Ben Sigman, developed an open-source AI memory tool called Mem Palace, inspired by ancient mnemonic techniques.

Aaron Levie argues that the diffusion of AI capability across enterprises will be slower than Silicon Valley expects, citing entrenched domain knowledge in systems like SAP and new security and operational complexities.

The central enterprise question is how to build software for a future where AI agents outnumber human users by factors of 100 or 1000 to one. This shifts focus to designing robust APIs, access controls, and monetization for agents.

A successful emerging paradigm gives coding agents access to SaaS tools and internal workflows, enabling them to both read information and use APIs or write code to execute tasks. This is exemplified by tools like OpenAI's 'super app' and Perplexity Computer.

Steve Sinofsky observes that agents do not seek simpler interfaces but choose backends based on cost, durability, and reliability. He contends the industry's focus on marketing to agents via APIs is wrong, as agents select systems based on underlying quality, not interface polish.

A major operational challenge is coordinating thousands of autonomous agents acting on shared systems, like a Box repository, which risks creating conflicting operations, performance issues, and security vulnerabilities that CFOs and CIOs must manage.

The permission model for agents is complex. While the 'end-to-end argument' suggests treating them like separate humans with their own accounts, agents are legally extensions of their users, requiring full oversight and lacking a right to privacy, which breaks traditional RBAC models.

Current AI agents struggle with information containment, as data in the context window can potentially be extracted via prompt injection. This makes it difficult to securely grant agents access to highly confidential resources like M&A data rooms.

Sinofsky predicts a widening gap in adoption speed between startups, which can adopt agents freely, and large enterprises like JP Morgan, which face significant legacy system and risk constraints, slowing AI diffusion.

There is tension between legacy SaaS vendors and the agent ecosystem, as agents want unlimited API access to data for operations, while vendors have traditionally monetized intelligence and domain expertise through UI-based subscriptions, not pure data licensing.

Martin Casado notes that every infrastructure company in his portfolio of about 50 has seen asymptotic growth in the last six months due to an unprecedented increase in software being written, driven by AI agent development.

The engineering compute budget for AI tokens is becoming a critical financial debate. CFOs must decide what percentage of R&D spend should go to tokens, a decision that directly impacts earnings per share given R&D typically constitutes 14% to 30% of tech company revenue.

A key friction is the current high cost of tokens, which pushes the industry toward usage-based pricing. This creates a short-term budgeting nightmare for engineering teams deciding between experimental waste and perfect optimization.

Sinofsky contends the token cost issue is transitional, comparing it to historical transitions like mainframe MIPS pricing. He believes the cost will plummet due to increased supply, algorithmic improvements, or hardware changes, making compute abundant.

Andy Miller describes AI-generated prose as often flat, lurid, and clunky, prone to repetitious metaphors, verbless sentences, and triadic adjectives - flaws evident in the withdrawn novel 'Shy Girl'.

Miller argues that while AI cannot match the profound originality of human literary genius, it can compete with formulaic commercial fiction, and some romance novelists already openly use LLMs to generate genre tropes.

Pavel began contributing to Samurai's Dojo software in 2019 because it was written in JavaScript, a language he knew, allowing him to add features to the open-source node software.

Ronin Dojo remains active despite setbacks, with Pavel finishing a UI update that will reintegrate a transaction privacy analysis tool, similar to the defunct kycp.org site.

Ashigaru is a fork of Samurai Wallet that demonstrates open-source code cannot be stopped by arrests; its team recently relaunched Whirlpool as an act of defiance.

Pavel notes Ashigaru's team communicates only via email, making public trust reliant on their transparency in documenting code changes and their rationale.

Anthropic's annualized revenue has surged to $30 billion, representing a 3x increase since the end of last year and a 58% increase since February.

OpenAI expects to spend $30 billion on model training this year, which is triple last year's cost. Anthropic projects its training costs will reach $28 billion by 2028.

Excluding training costs, both OpenAI and Anthropic are on track to generate a small profit this year. Anthropic forecasts a traditional profit by 2028, while OpenAI expects to turn cash flow positive by 2030.

Anthropic's enterprise customer base with annual spends over $1 million doubled from 500 to 1,000 in less than two months.

Anthropic signed a compute deal with Google and Broadcom for 3.5 gigawatts of capacity set to come online from 2027. Anthropic uses Google TPUs exclusively for inference, while its training clusters are operated by AWS.

Google's Gemma 4 model was downloaded 2 million times in its first week. In contrast, Gemma 3 saw 6.7 million downloads over a year, and Alibaba's Quen 3.5 achieved 27 million downloads since mid-February.

Meta's forthcoming model 'Avocado' will have an open-source version, contradicting prior speculation. Alexander Wang views Meta as a democratizing force for US-trained open-source models, while rivals focus on government and enterprise.

Meta employees have an internal leaderboard called 'Claudeonomics' tracking token consumption, with top users earning ranks like 'Session Immortal'. The practice is driven from the top, with Andrew Bosworth endorsing high token spend for efficiency gains.