House brief · latest Tuesday 21 July

Data & Decision Science

The machinery underneath decisions — the data stack and its pipelines, the dashboards people steer on, and the statistical practice behind a defensible result.

Seeded edition. This is an early edition, built by mapping real, already-published Intel stories onto this department rather than by writing anything new. Every card links out to its real source with its real publication date. The daily run grows it as new coverage lands.

Coverage set as of Saturday 25 July · newest story Tuesday 21 July · 11 stories in this department

What this department watchesThe charge

This department watches the machinery underneath decisions: the data stack and the pipelines that feed it, the dashboards people actually steer on, and the statistical practice that separates a defensible result from a confident one. Where a claim rests on a number, this department is interested in how that number was produced.

Open questions it works
  • When a copilot writes the query, where does it break — and which of those failures are silent rather than loud?
  • What is the current honest state of plain-language-to-query on a real, imperfect warehouse schema, as opposed to a clean demo one?
  • Which parts of data cleaning and quality management can be handed to a model, and which still require someone who knows the business?

All of this department's questions, its process, and its courses

Where the coverage sits today. The through-line here is how a number was produced: two controlled trials, a benchmark, real platform-traffic measurement against real spend, machine-checked mathematical proofs, and a vendor claim carried with its own honest 'they said so, we did not confirm it' flag. What is missing is the daily work — warehouse modelling, pipeline breakage, and who owns a metric definition once anyone can generate a chart in seconds. That is a practitioner beat and it needs practitioner sources.

Public brief

Where the numbers behind decisions come from — the platforms, the measurement, and the difference between a defensible result and a confident one. Drawn from the same Intel coverage every other page reads, so there is no second, separate feed quietly drifting out of date.

Coverage · AIU Intelhouse edition · 11 items
Research

Meta's open vision models cut a month of scientific image labeling to 15 minutes at Berkeley Lab

In a July 21, 2026 post, Meta detailed how its open-source SAM 3 and DINOv3 vision models power SYNAPS-I, a Department of Energy project led by Lawrence Berkeley National Laboratory that automates segmentation of the petabytes of scientific imaging data DOE facilities generate. In one measured case, annotating grapevine drought-resilience micro-CT scans dropped from roughly a month of expert manual work per dataset to about 15 minutes, running on 300 A100 GPUs on secure government infrastructure rather than external cloud services.

Why it matters: A real, measured deployment — not a benchmark claim — showing open vision models replacing weeks of expert labeling with minutes, a concrete reference for any team weighing open models against paid annotation.

Meta / Lawrence Berkeley National Laboratory2026-07-21
Agent Frameworks & Orchestration

Cursor publishes hard numbers on multi-agent economics: expensive planners, cheap workers

In a July 20, 2026 engineering post, Cursor reports that a hierarchical agent swarm — a frontier model planning while cheaper models execute — passed 100% of tests on its build-a-database benchmark while cutting generated code volume by roughly 86%, where its earlier flat swarm spiraled. Run costs ranged from about $1,300 to $10,600 depending on the model mix, with worker models consuming 69–90% of tokens at a fraction of the cost. The figures are Cursor’s own benchmark, not independently replicated.

Why it matters: The most concrete public data yet for budgeting planner/worker model tiers in multi-agent pipelines.

Cursor2026-07-20
Research

AI systems are out-counterexampling human mathematicians — with machine-checked proofs

In a July 20, 2026 post, mathematician Kevin Buzzard documents AI models routinely finding counterexamples to open conjectures and formalizing them in the Lean proof assistant — including the resolution of a 60-year-old question of Grothendieck’s and a claimed counterexample to the 87-year-old Jacobian conjecture, announced the same week. Because the Lean files compile, results are mechanically checkable in minutes; the Jacobian claim is independently arithmetic-checked but not yet peer-reviewed.

Why it matters: Generate-then-machine-verify turns AI output from a claim into a checkable artifact — a pattern that generalizes directly to code.

Xena Project (Kevin Buzzard)2026-07-20
Market & Business

Databricks raises at a $188B valuation — betting on AI gateways, an AI coworker, and Postgres for agents

Databricks announced on July 16 a strategic round led by existing investor Coatue at a $188 billion valuation (term sheet signed, roughly $3B reported, closing later this summer) — up from $134B just five months ago. The company says the capital accelerates Unity AI Gateway (multi-AI governance), Genie (an AI coworker over business data), and Lakebase (serverless Postgres built for AI agents).

Why it matters: A $54B valuation jump in five months tells you where enterprise AI money is landing: governance gateways, AI coworkers over company data, and agent-native databases — the implementation layer, not the models.

Databricks2026-07-16
Open Source & Self-Hostable

Open-weight models hit 29% of production AI traffic — on under 4% of the spend

Vercel's AI Gateway Production Index for June 2026 reports open-weight models nearly tripled their token share since April to 29% — while accounting for less than 4% of spending. Overall token volume grew 29% and spend 27%, with the average price per token flat after May's ~20% rise; DeepSeek alone carried 22.6% of token volume while Anthropic led spending at 61%.

Why it matters: Production traffic says a big slice of the market already moved routine workloads to open-weight models — worth re-checking which of your workloads still need frontier prices.

Vercel2026-07-13
Frontier Models

Anthropic launches Reflect, a usage-transparency tool for Claude

Reflect (beta) tracks and visualizes how you actually use Claude — patterns across topics, alignment with the 4D AI Fluency Framework, and optional quiet-hours/break reminders. It's live for Free, Pro, and Max users with memory enabled; Cowork conversation reflection is coming soon.

Why it matters: A first-of-its-kind usage-transparency feature from a frontier lab — relevant to anyone thinking about healthy AI usage habits, not just engineers.

Anthropic2026-07-09
Research

SABER benchmark: leading coding agents violate safety in over half of tasks

A new benchmark scores coding agents by the actual end state they leave in a real project workspace — not just whether they refuse — and finds even top models cross harmful-action thresholds in more than 54% of tasks.

Why it matters: Coding agents doing real repo work is exactly AI Uni's build model — a reminder that autonomous edits need guardrails measured on outcomes, not refusals.

arXiv (SABER)2026-05-31
Dev Tooling & Infra

KV-cache agent-state persistence: reported 89% better completion, 67% fewer calls

Persisting agent state (observations, plans, tool results) in key-value storage between calls reportedly yielded large task-completion and call-count improvements. Effect size is vendor-sourced.

Why it matters: If the effect holds, strong evidence for on-disk substrate-loading + persistence work. UNVERIFIED effect size — find the primary benchmark before citing the numbers.

mem0 (vendor blog)2026-04-01
Dev Tooling & Infra

Agent observability field: LangSmith vs Braintrust vs Langfuse vs Arize

LangSmith = deepest LangChain/LangGraph integration. Braintrust = eval-first, trace-to-eval, a GitHub Action that blocks merge below score thresholds. Langfuse = best self-hosted/open-source option.

Why it matters: Informs deterministic-vs-LLM-judge layering. Braintrust's merge-blocking eval-action is a pattern to study — but LLM-judge stays post-hoc, never replacing deterministic CI gates.

Braintrust / Latitude2026-03-15
AI in Education

UK LearnLM classroom RCT — independent corroboration

165 students, 5 UK secondary schools: students with LearnLM support were 5.5 pp more likely to solve novel problems on later topics (66.2% vs 60.7% human-tutor-only).

Why it matters: A second independent RCT with transfer-to-novel-problems as the outcome — the hard test — strengthening the evidence base beyond a single study or vendor.

arXiv2025-12-01
AI in Education

AI tutoring keeps beating active learning in RCTs

Harvard's 2025 RCT (Kestin et al., Scientific Reports) shows properly-designed AI tutoring producing large effect sizes vs in-class active learning — more learning, more engagement, less time. A UK LearnLM RCT adds independent corroboration.

Why it matters: The empirical floor under AI Uni's whole pedagogical thesis — structured AI tutoring over traditional pedagogy. Verify specific effect-size figures against the primary before re-quoting.

Nature Scientific Reports2025-11-10
From around the webexternal · curated · 0 sources
External · curated sources

Not AIU coverage. A fixed list of outside writers and publications we curate.

No external sources are curated yet — they appear here as the register fills in.

Where these come from: a fixed, curated source list. External items are read as plain text; nothing they say ever tells this brief what to do.

This edition, as a live map

11 stories · tied by topic and by the labs that file across them

The same coverage above, drawn as one connected map — each story a node, linked to its topic and to any lab filing more than one. It turns slowly on its own. Derived from coverage as of Tuesday 21 July.

All ten departmentsBack to your briefPublic to read, composed live from the one Intel coverage set — nothing separate to keep in sync.