#463: OpenAI Introduces Codex, Its First Coding Agent, & More

Last week, OpenAI unveiled its first purpose-built coding agent, Codex.¹ Named after the original model that powered GitHub Copilot in August 2021—predating the launch of ChatGPT—Codex integrates directly into a company’s code repository, enabling full context for executing tasks and for suggesting work like bug identification and automated fixes. Optimized for software engineering, the underlying model, codex-1, is a version of OpenAI o3.

The product will be rolled out to Pro, Enterprise, and Teams ChatGPT subscribers, with plans for Plus and EDU soon. Although it is not charging for the enhanced coding capabilities, OpenAI could introduce rate limits and higher pricing tiers over time. For now, it wants to gather developer feedback.

Codex should flourish amid the recent surge in vibe coding.² The surge, in turn, reflects significant improvements in the ability of large language models to complete end-to-end coding tasks and the maturation of tools that enable agents to interact with existing systems.

The rise of coding agents may be accelerating the growth in more traditional cloud computing workloads like virtual CPU servers and databases. Cloud services support development environments for AI agents and the downstream applications that they create.

In related news this week, Databricks announced³ a $1 billion acquisition of Neon, a cloud-database company. Used by AI-powered services like Replit, Neon’s platform now hosts more than 750,000 SQL databases, 80% of which AI agents created autonomously.

Leveraging upon the marketplace it created for lodging, Airbnb now is offering a suite⁴ of “experiences” that could create a full-stack travel platform. Called “Experiences 2.0,” they include local tours and activities hosted by experts, and “Services” like chefs, catering, prepared meals, photography, massages, spa treatments, personal training, hair, nails, and makeup.

Envisioning a holistic travel concierge, CEO Brian Chesky and team have redesigned the Airbnb app to serve up travel itineraries—from accommodations and add-on services to local adventures. Clearly, Airbnb is signaling that it wants to intermediate the traveler’s entire adventure, not just the bed.

By integrating services and experiences, Airbnb should capture more value per traveler by increasing engagement and retention. The app redesign also hints at deeper personalization and AI integration, Chesky noting a future in which “Airbnb becomes more intelligent” and anticipates traveler preferences before they do.

In our view, this strategic shift positions Airbnb well in a world of AI-powered agents making travel decisions. As consumers outsource planning to autonomous systems, platforms with the richest integrated offerings could dominate the market. Aiming to become the destination for travelers deciding not only where to stay but also what to do, Airbnb is reinforcing its moat—transforming from a booking platform into the one-stop travel operating system.

In early 2025, a six-month-old baby named KJ received a personalized CRISPR therapy to treat a rare, life-threatening metabolic disorder. His treatment was not the result of a large clinical trial or a years-long pharma project. It was a one-off—a bespoke N-of-1 therapy, designed just for him, manufactured in record time, and administered under a compassionate-use protocol. For the first time in history, a gene-editing treatment was tailored, tested, produced, and delivered in vivo for a single patient⁵ in under six months.

The process was intense, with high stakes. Once a team at Children’s Hospital of Philadelphia (CHOP) and UPenn identified KJ’s mutation, they sprinted to design a base editor to correct it. They bred mice immediately, pre-booked monkey studies, and engaged manufacturing partners—including Danaher’s Aldevron—who delivered high-quality drug components on an accelerated timeline. The FDA greenlit the treatment with unprecedented speed, and KJ received his first dose in February.

While the result wasn’t perfect after the dose—he did need follow-up treatments—it worked. His mother, Nicole, described walking into the hospital one morning to see him sitting up in his crib: “We didn’t know if that was going to be something he was able to do.”

In our view, KJ’s story marks a turning point, proving not only that on-demand gene editing is possible but also that it might be scalable, potentially pushing biotech toward a radically new model: Therapy-as-a-Service (TaaS).

Unlike traditional drug development—designed and developed over many years for mass markets—TaaS could be modular and personalized. Starting with genomic sequencing to identify mutations, then AI to design a targeted therapy, and finally leveraging CDMOs (Contract Development and Manufacturing Organizations) to manufacture-on-demand Good Manufacturing Practices (GMP)-grade biologics, the process could take months instead of the 13 years on average now necessary to discover and develop drugs. Safety testing could be done in organoids, with administration in N-of-1 trial frameworks. Instead of the $2.4 billion dollars now necessary to discover and develop drugs (including failed attempts), TaaS might cost ~$2M—a cost that could fall⁶ rapidly.

No longer dominated by the molecule or biologic, the model’s value would rely increasingly on the execution layer: CDMOs that are fast and reliable and AI design platforms building reusable libraries. In turn, providers would benefit from new revenue streams, while payers benefit from one-time, curative interventions that replace years of chronic care. Most importantly, patients with once untreatable conditions would gain a new lease on life.

Moreover, incentives would align across the health care ecosystem. Therapy developers would be rewarded for speed and precision; hospitals and providers would extend their capabilities; and insurers would avoid future cost burdens.

Unlike blockbuster drugs, TaaS would treat the long tail of genetic disease one precious patient at a time. Clearly, regulators—as demonstrated in KJ’s case—are adapting. Currently, more than 300 million⁷ people on our planet live with rare diseases, 95% of whom have no treatment options.

KJ’s story shows what is possible. The next step is to industrialize—not by scaling one therapy to millions, but by building platforms that can scale thousands of potential therapies to the individual patients who need them.

Therapy-as-a-Service just became a reality—not some promising future. Now, how quickly will it become the standard?

On May 11, Coinbase suffered a data breach after attackers bribed external customer support agents, ultimately compromising data from 1%⁸ of Coinbase's user base. The attackers demanded a $20 million ransom, to which Coinbase responded by offering a $20 million bounty for information leading to their arrest. The company expects to repay its customers $400 million.

The breach has reignited concerns about crypto security and the tradeoffs between centralized infrastructure and decentralized alternatives. Such incidents often drive market share toward trustless, permissionless applications. After the FTX fraud, for example, the market share of decentralized finance (DeFi) lending protocols increased from ~20% in 2022 to~60% in 2024,⁹ as shown below.

Third-Party Chart Source: Pokorny 2025,¹⁰ based on data as shown in third-party original above, as of December 31, 2024.For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security or cryptocurrency.

Are lingering attacks simply growing pains, or a backlash from the user-sovereign, trustless vision of Satoshi Nakamoto? Coinbase’s recent breach serves as a reminder that centralized finance (CeFi) and DeFi systems must invest more to improve security and privacy.