Frontier labs are re-litigating what actually constrains them, chips or data, while the sharpest edge in coding agents has moved from the underlying model to the system wrapped around it.
Inside the Machine: Interpretability Gets Practical
Anthropic’s newest interpretability work suggests Claude does more thinking than it ever shows on the page. Tencent’s latest open release raises the opposite question: how much do we actually know about what a model can do before someone independently tests it.
- Anthropic Finds a Workspace for Deliberate Thought in Claude. A new interpretability paper identifies internal patterns, nicknamed J-space, that let Claude reason deliberately and can reveal thoughts it never writes down in its visible output.
- Tencent Ships Hy3, a 295B Open Model, Free Through July 21. Tencent says Hy3 rivals larger open flagships, but so far the only public test is Simon Willison’s pelican-on-a-bicycle SVG check.
Data Over Chips: The New Scarcity
Two operators inside the industry are converging on the same read: the bottleneck is shifting from compute to data, and from frontier models to fine-tuned open ones.
- The Next AI Buildout Isn’t Chips, It’s Data. Will Depue argues frontier labs are moving from a compute-limited regime to a data-limited one, with spend set to top $100 billion a year by 2030.
- Decagon’s Jesse Zhang: Open Models Beat Frontier Ones at Scale. The customer service AI startup runs most of its production workloads on fine-tuned open-source models, and its CEO expects that pattern to spread as enterprise AI matures.
The Harness Beats the Model
Four stories today point to the same shift: the scaffolding around a coding agent, not the base model inside it, is where the competitive edge now lives.
- Replit Builds Two Systems to Make Agents Learn Without New Weights. Michele Catasta says ViBench and Telescope let Replit improve its production coding agents through harness and context changes instead of retraining.
- Claude Code, Codex, and Omp Are Now Roughly Tied on Quality. A mid-2026 survey argues the coding agent harness, not the underlying model, now decides which tool wins a given task.
- Claude Developers Maps the Four Variables Behind Agent Loops. Anthropic’s builder account breaks agent loops into trigger, stop condition, primitive, and task fit, then names the tradeoff between output quality and token spend.
- Why Top Agentic Engineers Are Pulling Away From the Median. A post from systematicls argues the current model era rewards orchestration skill over raw coding speed, widening the gap between the best operators and everyone else.
Infrastructure and Capital: Betting Years Ahead
The capital behind AI keeps flowing into chips and power, not just models, with multi-year commitments locking in supply years before the demand arrives.
- Broadcom Locks In Apple Silicon Deal Through 2031. The chipmaker will co-design custom AI processors with Apple across several product generations as Apple prepares to run its own AI servers by 2027.
- TeraWulf Signs 20-Year, $19B Anthropic Lease, Exits Texas JV. TeraWulf will supply roughly 401 megawatts of Anthropic’s compute starting in 2027, while selling its stake in a separate Texas project to Fluidstack.
Policy and Geopolitics: Talk Versus Rules
Governance moved at UN pace this week, while a rebrand at xAI showed how fast a corporate narrative can shift without any underlying product changing.
- UN Opens First AI Governance Dialogue, No Binding Rules Attached. Geneva talks drew all 193 member states plus industry and researchers, producing positions on AI safety and liability, not commitments.
- xAI Renames Itself SpaceXAI. Nothing Operational Changed Yet. The rebrand tightens Musk’s AI-and-rockets narrative right after SpaceX’s IPO, but Grok’s problems predate the new name.