I launched a website! It’s the landing page for all my content, including this newsletter. Check it out!
SUMMARY
Anthropic continues to scale compute consumption, demonstrates meaningful capabilities with Mythos, and launches dynamic workflows in research preview that allows for running substantially longer tasks. OpenAI announces innovative mathematical findings, as they confidentially file an S-1 with the SEC. Cerebras is running Kimi K2.6, a trillion-parameter open-weight model with low latency responses. Uber built an agent platform to solve for provenance, carrying user identity throughout the agentic request chain. And the pope writes his first encyclical on safeguarding the human person in the time of artificial intelligence.
FRESH READS
① Anthropic will pay xAI $1.25B per month for compute
Summary: Anthropic structured a deal with xAI to buy 300 MW of compute, which is the full output of their Colossus 1 data center near Memphis. The deal, disclosed in SpaceX’s S-1 SEC filing, will cost Anthropic $1.25 billion per month through May 2029.
Signal: xAI over-built its compute infrastructure as Grok usage declined and is now monetizing available capacity to a direct competitor. The neocloud model, i.e., building infrastructure for your own needs and renting idle capacity to others, is emerging as a structural hedge in the AI infrastructure build out. It demonstrates that AI infrastructure is decoupling from AI model success.
② Project Glasswing: An initial update
Summary: Six weeks into Project Glasswing, Anthropic reports that ~50 partners have used Claude Mythos Preview to find more than 10k high- or critical-severity vulnerabilities across systemically important software with some partners seeing a greater than 10x increase in their bug-finding rate. The constraint has shifted from vulnerability discovery to vulnerability verification, disclosure, and patching. Open-source maintainers are reporting that they are capacity constrained with the volume of security disclosures and bug reports.
Signal: The risk window here is structural. The gap between fast vulnerability discovery and slow patching is where attackers will operate. The call to action is to compress patch cycles and accelerate enhancements to automated software delivery pipelines.
③ Project Glasswing: what Mythos showed us
Summary: Cloudflare pointed Mythos at fifty of their repositories, finding 2k bugs across critical-path systems, 400 of which were high- or critical-severity. They observed that two capabilities distinguished Mythos from prior general purpose models: 1/ exploit chain construction, i.e., chaining multiple small attack primitives into a single working exploit, and 2/ proof generation, i.e., iterating on an exploit until it was demonstrably functional. Cloudflare characterized the output as resembling the work of a senior security researcher rather than an automated scanner.
Signal: The key lesson from Cloudflare’s experience is architectural. General coding agents didn’t work for this kind of security analysis because they lacked the context breadth and throughput required for vulnerability analysis. A purpose-built harness was required, targeting narrowly scoped tasks, using adversarial review between discrete agents, and implementing parallel execution. This reframes agentic security as an engineering investment, not a model procurement decision.
④ Introducing dynamic workflows in Claude Code
Summary: Announced alongside the Opus 4.8 launch, Claude Code now supports dynamic workflows, where it shards work, orchestrates parallel subagents, and verifies the returned work product. Dynamic workflows allow for long-running work that could extend hours or even days, like the rewrite of Bun from Zig to Rust over 11 days, generating roughly 750k lines of Rust with 99.8% of existing test suites passing.
Signal: METR measures the time horizons of public frontier AI models, and enhancements to harnesses like Claude Code are demonstrating that workflows can work autonomously for longer and longer periods of time. Since longer runs consume more tokens and drive higher cost, the workload should deliver commensurate value to the organization.
⑤ An OpenAI model has disproved a central conjecture in discrete geometry
Summary: An OpenAI reasoning model disproved the Erdős unit distance conjecture, an 80-year open problem in discrete geometry, by constructing an infinite family of point configurations that beats the traditionally accepted square grid. The key move was importing tools from algebraic number theory, e.g., Golod-Shafarevich theory, infinite class field towers, which no one had previously connected to the problem. The proof was independently verified by external mathematicians.
Signal: An AI system has autonomously resolved a longstanding 80-year open problem at the center of an active field. This was enabled by cross-domain lateral thinking, which could help advance the frontiers of research in math, science, and engineering.
⑥ The big questions looming over OpenAI’s trillion-dollar IPO
Summary: OpenAI filed a confidential S-1 with the SEC on May 22, targeting a public listing as early as September. The company’s last private round valued it at $852 billion with projections up to $1 trillion at IPO. When the S-1 goes public, it will be the first full disclosure of OpenAI’s burn rate, revenue mix, unit economics, and executive compensation structure.
Signal: The IPO is a stress test for the AI industry’s financial narrative, not just OpenAI’s. If public markets bite despite the burn rate, it signals continued tolerance for frontier AI capital expenditures and clears the path for Anthropic and others. If they balk, it resets valuation expectations across the board.
⑦ Cerebras says its chips run a trillion-parameter AI model nearly 7 times faster than GPU clouds
Summary: Days after completing the largest tech IPO of 2026 at a $95 billion valuation, Cerebras announced that it is running Kimi K2.6, a trillion-parameter open-weight model for enterprise customers at nearly 1k tokens per second, 6.7 times faster than the next fastest GPU-based cloud provider. For agentic coding tasks, that translates to a full response in 5.6 seconds versus 163.7 seconds on the official Kimi endpoint.
Signal: Nvidia’s acquisition of Groq for $20 billion is a tell, namely that inference is becoming a strategic battleground. Cerebras is positioning itself to enterprise customers as an alternative to capacity-constrained closed-source model providers like Anthropic and OpenAI, by leaning into wafer-scale processing that delivers low latency inference at comparable per token unit cost.
⑧ Solving the identity crisis for AI agents
Summary: Uber documents the agent platform that they built in early 2025 that allows teams to compose, deploy, and operate production-grade agents at scale. The core problem they solved was provenance, i.e., carrying the originating user identity and the full delegation chain across every agent hop, rather than losing that context at each service boundary. The platform includes two build paths: 1/ a Python SDK for developers, 2/ a no-code UI for business users. The platform has been adopted across thousands of internal agents with p99 security overhead under 40 milliseconds.
Signal: As agents become digital workers acting on behalf of users, identity cannot exist at just the point of user authentication. It needs to be carried across the full delegation chain to ensure that each hop accesses only resources allowed for that end user. Agent identity and provenance are not optional plumbing but the trust foundation that makes agentic workloads safe to put into production.
ONGOING
① Magnifica Humanitas: an encyclical letter of Pope Leo XIV
Note: I’ve only had a chance to read the introduction so far, so I cover only that here. I plan to post summaries of additional chapters as time affords.
Summary: Pope Leo XIV’s encyclical letter, published on May 15, 2026, is a letter on safeguarding the human person in the time of artificial intelligence. The introduction draws on two biblical illustrations when considering the trajectory of technological progress: the Tower of Babel and the rebuilding of Jerusalem. The pope poses three orienting questions with regard to artificial intelligence: 1/ where are we going, 2/ what’s our goal, 3/ what direction should we choose as a human society? While technology is not inherently evil, it also isn’t neutral. Instead, it takes on the characteristics of those who devise, finance, regulate, and use it.
Signal: The introduction’s most pointed claim is that the primary drivers of AI development are private transnational actors, whose resources exceed those of many governments, making regulation necessary but insufficient. A non-state actor with global moral authority has laid down a challenge to orient technological progress for the common good.
ONE TO WATCH
① Inside Anthropic’s $100 Billion Al Compute Commitment | CFO Krishna Rao
Summary: Patrick O’Shaughnessy sits down with Anthropic CFO Krishna Rao to discuss the business of building at the frontier. They cover compute strategy at extraordinary scale, pricing philosophy, the fungibility of compute resources across research and inference workloads, Jevons paradox applied to both token consumption and labor, and what it means to release a model as capable and sensitive as Mythos. The conversation closes on culture, optimism in biotech, and the indicators that Anthropic is closely watching to see if that scaling story is slowing.
Signal: Anthropic has deep internal conviction that the capability gains they are delivering will continue and, as a result, is deploying significant capital accordingly. The combined compute commitments of up to 10 GW across Google and Amazon is a bet that only makes sense if you believe that the scaling laws hold. Furthermore, with their seven co-founders still present and maintaining the culture bar, that conviction is grounded in a culture of intellectual honesty, open dialog, and well-informed decision making.
Always be learning.
heeki reads #2
Written by Heeki Park, Principal SA @ AWS. Opinions are my own.
Alcurio is where alchemy meets curiosity.