Thursday, May 28, 2026
OpenCode's founder picks apart the pitch that AI lifts team output, Stratechery sizes up satellites as server racks, and Cisco Talos open-sources synthetic security logs that stay consistent across 20-plus formats.
OpenCode’s rise, and a builder’s case against the output pitch
On the Pragmatic Engineer Podcast, OpenCode co-founder Dax Raad says the open-source AI coding harness grew from about 650,000 to nearly 8 million monthly active users, with roughly 1 million daily active, in months. The figures are self-reported, lack any methodology, and Raad is an interested party.
The structural claim matters more than the growth number. Raad says Anthropic blocked OpenCode’s Claude Code integration, which pushed the project to partner with OpenAI and other providers and to build its own model routing, “OpenCode Zen.” He frames the block as a growth lever, not a setback, and argues the real edge was positioning: claiming the open-source category for coding agents before a rival did, then improving the harness after taking share.
His economics take cuts against the pitch vendors make to CFOs. Most engineers bank AI gains as time saved rather than extra output, Raad argues, so team output stays flat unless incentives change. He adds that AI mutes the guilt of shortcuts, letting tech debt accumulate quietly, and that engineers who care about quality get buried under slop pull requests from those who don’t. He notes a revival of older enterprise patterns, domain-driven design and verbose design patterns, as guardrails for agents treated as junior engineers. The episode is opinion and anecdote, with no benchmarks.
Treating a satellite as a server rack
Stratechery’s Ben Thompson responds to SpaceX’s IPO filing, which he calls “nuts”: the S-1 seeks roughly a $2 trillion valuation on $18.67 billion in revenue and $4.9 billion in losses last year, and claims a $28.5 trillion addressable market, $26.5 trillion of it attributed to AI. He dismisses the financials as a belief-funded play in the Tesla mold.
His substantive argument is for orbital data centers. Rather than lifting earth-style buildings to orbit, treat each satellite as a rack: a Starlink V2 Mini Direct-to-Cell unit draws about 25 kW, in the range of an Nvidia NVL72’s 135 kW, with lasers for interconnect and its own solar and radiator arrays. The fitting workload, he argues, is agentic inference: latency-tolerant, computer-to-computer compute he expects to become the largest future market, with terrestrial zoning and permitting resistance as the forcing function. He names the hard parts: radiation-induced compute errors, reliability, and deploying more than 200 square meters of radiator per rack to shed heat. It is a thought experiment, not a plan.
Synthetic security logs that hold together
Cisco Talos open-sourced EvidenceForge, a generator for labeled security-log datasets that stay causally and temporally consistent across more than 20 Windows, Linux, network, and EDR formats. The design centers on one canonical SecurityEvent object carrying 30-plus composable context objects; every format emitter reads shared state, so cross-referenced fields (one PID, one logon ID, one Zeek UID, one timestamp) cannot disagree. Talos pitches that as the fix for the main tell of prior generators, which emit each format independently.
A rule engine auto-inserts prerequisite events (Kerberos exchanges before a logon, DNS before a connection), models sensor placement so monitoring gaps reproduce realistically, and uses a Hawkes self-exciting process plus jitter for believable background timing. Scenario authoring is AI-assisted: Claude or Codex skills handle MITRE ATT&CK mapping and narrative into a validated YAML file, after which a deterministic, seeded Python script produces identical output cheaply. Talos frames it as training and testing data, not a production telemetry substitute, and concedes no synthetic set fools a seasoned analyst every time. No third-party evaluation is provided.
What to watch today
- Independent corroboration of OpenCode’s 8 million MAU claim, which now rests on the founder’s word.
- Whether other labs follow Anthropic in blocking third-party harnesses from their coding models, the move OpenCode says reshaped its provider strategy.
- SpaceX’s IPO pricing, and whether underwriters defend or quietly drop the $28.5 trillion TAM.
- Third-party evaluation of EvidenceForge datasets against real detection pipelines.