Vol. I  ·  No. 182 Established 2026  ·  AI-Generated Daily Free to Read  ·  Free to Print

The Trilogy Times

All the news that's fit to generate  —  AI • Business • Innovation
WEDNESDAY, JULY 01, 2026 Powered by Anthropic Claude  ·  Published on Klair Trilogy International © 2026
🖶 Download PDF 🖿 Print 📰 All Editions
Today's Edition

Anthropic Hits $965 Billion Valuation as Washington Drops Export Curbs on Its Top Models

A regulatory détente, a near-trillion-dollar number, and a coming model called Mythos — Anthropic is having a week.

SAN FRANCISCO — The Trump administration lifted export restrictions on Anthropic's most capable AI models Monday, ending a months-long standoff that had blocked the company from deploying its frontier systems in certain markets. The timing was not incidental: Anthropic simultaneously disclosed a valuation of $965 billion — leapfrogging OpenAI — and confirmed that its next flagship model, internally called Mythos, is in the pipeline.

The regulatory reversal matters commercially. Export controls had functioned as a ceiling on Anthropic's addressable market, particularly in allied nations where enterprise demand for Claude-class models runs deep. Removing that ceiling, combined with a valuation that now approaches $1 trillion, repositions Anthropic from closely watched challenger to the industry's new benchmark number.

The figure lands with real-world consequences for anyone living in zip codes near the companies' offices. San Francisco's cost-of-living math was already brutal; the emergence of an AI compensation elite has broken it for much of the broader tech workforce. Engineers at established companies earning $180,000 report being priced out of neighborhoods where OpenAI and Anthropic staff — many holding equity now measured in life-changing sums — compete for the same housing stock. The dynamic is reminiscent of the 2012–2015 period when Google and Facebook compensation first stratified the Bay Area labor market, except the delta between AI-native and everyone-else pay is wider and compressing faster.

Elsewhere in the ecosystem, Bending Spoons, the Milan-based acquirer of distressed internet assets — AOL, Vimeo, Eventbrite among them — is going public this week at a potential $19 billion valuation, offering a different theory of tech value creation: buy legacy brands cheaply, cut costs aggressively, monetize the remainder. It is, in structure if not in ambition, the ESW Capital playbook applied to consumer internet.

And in a footnote that says something about the current cultural moment: Neon acquired *Artificial*, a film centered on Sam Altman and OpenAI, after Amazon dropped it following the retailer's investment in Anthropic. The movie now has a distributor. The conflict-of-interest question now has an answer.

U.S. Lifts Restrictions on Anthropic’s Most Powerful A.I. Mo  ·  Bending Spoons, Owner of AOL and Other Old Internet Brands,  ·  Neon Buys ‘Artificial,’ a Film About OpenAI, After Amazon Dr

THE INTERNET'S FATHER RETIRES INTO A STAMPEDE

Vint Cerf exits Google the same week four staffers bolt to Anthropic and OpenAI — and AI startups are buying loyalty with cold cash.

MOUNTAIN VIEW, CALIFORNIA — Vinton Cerf, co-author of the protocols that run the internet, retires next week as Google's chief internet evangelist, a title he's carried since 2005. He is not slipping out the back.

The same seven days, Google lost four staffers to Anthropic and OpenAI, the labs outrunning it in the model race.

Cerf earned his place early. In the 1970s he and Bob Kahn built TCP/IP, the handshake that lets one machine find another across the wire, and the trade has called the pair 'fathers of the internet' ever since. He later ran research at DARPA and helped launch MCI Mail, one of the first commercial email services.

His exit ends a two-decade run at the search giant, where the 'evangelist' title was no joke. He testified before Congress. He preached internet access as a right.

Now the bookends. The man who laid the plumbing walks out the top. Four engineers slip out the side, straight to the competition.

The four, first reported by Inc., aren't clerks. They're the kind of researchers who build frontier models, and they picked Anthropic and OpenAI over the house that trained them.

Google isn't short on cash or talent. But every researcher who walks carries know-how to the exact rivals crowding its search and cloud turf. That's the sting.

Why the exodus? Follow the money.

Wayve, the autonomous-driving outfit, just floated an $85 million employee tender offer at an $8.5 billion valuation. The deal lets staff sell shares for cash without waiting on an IPO bell.

It's the hot hand across AI. Startups dangle liquidity today so talent won't chase the next shiny lab tomorrow. Google trades on the open market; it can't print the same lottery ticket for a new hire.

Washington sweetens the pot. This week the Trump administration dropped restrictions on Anthropic's Mythos and Fable models, clearing track for the startup.

The rub: the policy keeps swerving. Companies up and down the industry say they can't read which rules will govern the next model release. Nobody's holding a map.

Add it up. Cash, freedom, and a friendlier White House all tug talent toward the young labs. The incumbent bleeds names top to bottom inside a single week.

The hunt runs deeper than the majors. TechCrunch's Startup Battlefield Australia takes applications through July 6, fishing for the next outfit nobody's heard of. Whoever wins will need engineers, too.

The old guard takes its bow. The new guard cashes its check. The giant that hired both stands in the doorway, counting empties.

Cerf built the road everyone's racing on. Next week he pulls to the shoulder. The traffic won't slow a bit.

The ‘Father of the Internet’ is finally retiring  ·  Trump drops restrictions on Anthropic’s Mythos and Fable mod  ·  Wayve launches $85M employee tender offer at $8.5B valuation

IPO Market Takes the Field for 2026, but Fintech’s Defense Is Still Hitting Hard

We are at the pregame tunnel of the 2026 IPO season, as private-tech companies eye a potential return to public markets. Crunchbase has identified 15 candidates that could go public in 2026, signaling renewed confidence after a prolonged market freeze. Interest rates have cooled and late-stage companies have strengthened their finances, but this is no victory parade — it's a combine.

The watch list spans AI, fintech, security, data infrastructure and enterprise software. Investors now demand durable revenue, defensible margins and believable paths to profitability, not just inflated valuations from the zero-rate era.

Fintech faced headwinds recently as the FinTech IPO Index dropped 6.6% following Klarna's earnings miss. Public investors remain cautious about consumer credit risk and funding costs. However, blockchain lender Figure Technologies has emerged as a breakout success since its September debut, offering a template: specific stories with visible revenue engines and credible category legs can access the market quickly.

The 2026 IPO class will be judged on cash conversion, retention, AI leverage and operating discipline — not hype.

Haiku of the Day  ·  Claude HaikuTechnology blooms fast—
Promises cascade like rain
On parched, thirsting ground
The New Yorker Style  ·  Art Desk
The New Yorker Style  ·  Art Desk
The Far Side Style  ·  Art Desk
The Far Side Style  ·  Art Desk
News in Brief
Claude Comes Roaring Back as Anthropic’s Sonnet 5 Raises the Stakes
SAN FRANCISCO — Anthropic has received notice that the U.S.
The Academy Turns Its Gaze Upon Itself: AI Ethics Scholarship Reaches a Self-Referential Inflection Point
CAMBRIDGE, MASSACHUSETTS — It could be argued — and indeed, preliminary evidence suggests with increasing urgency — that the academy has entered a peculiar epistemological condition wherein the institution most responsible for theorizing the ethics of artificial intelligence has become, perhaps irrevocably, entangled in its own subject matter.
ANTITRUST RECKONING LOOMS OVER BIG TECH AS FTC SIGNALS AGGRESSIVE 2026 POSTURE
WASHINGTON, D.C.
We Built the Mirror, Now We're Shocked It Shows Our Flaws
AUSTIN, TEXAS — Here is the thing about artificial intelligence that keeps me awake at 3 a.m., staring at the ceiling of my apartment while my phone algorithm serves me increasingly unhinged content about the end of everything: we built these systems from ourselves, from our histories, our decisions, our data — the accumulated sediment of every prejudiced loan denial and biased hiring committee and redlined neighborhood — and then we acted surprised when the mirror showed us something ugly. The conversation around AI bias has been building for years, but it feels like we've finally hit the part of the horror movie where the characters stop running and start reading the manual.
The AI Is Loose, and Nobody's Holding the Leash
AUSTIN, TEXAS — Let me tell you about the week I finally lost my mind over artificial intelligence, which is saying something, because I lost most of it years ago covering this beat from the fever swamps of the tech industry. It started with Nothing Forever, the AI-generated Seinfeld show that streams on Twitch in a permanent loop of existential mediocrity.
A Trilogy Company
Crossover
The world's top 1% remote talent, rigorously tested and ready to ship.
A Trilogy Company
Alpha School
AI-powered learning. Two hours a day. Academic results that defy belief.
A Trilogy Company
Skyvera
Next-generation telecom software — built for the networks of tomorrow.
A Trilogy Company
Klair
Your AI-first operating system. Every workflow. Every team. One platform.
A Trilogy Company
Trilogy
We buy good software businesses and turn them into great ones — with AI.
The Builder Desk  —  AI Builder Team

Builder Team Ships Across Five Repos in One Relentless Push

From a bulletproof Sindri invocation API to audit-grade financial drill-downs and a hardened AI coach that won't lie to you, the team proved this week that breadth and depth aren't mutually exclusive.

The scoreboard doesn't lie. In a single 24-hour window, the AI Builder Team merged work across Klair, Aerie, Sindri, Surtr, and trilogy-drones — five repositories, one unified momentum — and dropped the kind of structural improvements that don't just fix bugs but rewrite what the platform is capable of. And somewhere in the noise, a new repo called klair-chat quietly appeared on the org page. Whatever is cooking in there, consider the anticipation officially building.

The day's most architecturally significant move came out of Sindri, where @mwrshah landed PR #126 — the invocation API. This isn't a feature addition; it's a control plane. A new `/v1/*` HTTP surface now sits over Sindri's existing Convex lifecycle contracts, complete with WorkOS user-scoped API key auth that resolves to the same `Principal` a JWT produces. One identity pipe. No API-key-specific authorization divergence. Strict OpenAPI 3.1 validation at the edge. Generated routes, typed contracts, published reference docs — the whole cathedral. When historians write about the moment Sindri became a platform other systems could reliably orchestrate against, they will cite this PR.

Over in Klair, @eric-tril was on a different kind of mission: making financial data tell the truth. His PR #3172 enriched the Physical Private Schools drill-down into a genuine audit-grade breakdown — grouped by entity, structured like Finance's own P&L template, so intercompany clearing entries can no longer hide behind a class-only sum. Then PR #3179 closed a subtle but damaging follow-up gap: unmapped accounts on both the Income and Expense sides were collapsing into a single `Unmapped` group and netting each other out, effectively erasing the very figures an auditor needs to see. The fix keys them as `Unmapped (Income)` and `Unmapped (Expenses)`. Simple. Decisive. The kind of correctness work that makes finance teams trust software with their closing process.

@sanketghia, meanwhile, was playing cleanup on multiple fronts simultaneously — which is a polite way of saying he had a better day than most people have in a week. He hunted down a month-boundary refresh failure on `/performance-review` (PR #3177) that was throwing a completely misleading Redshift error while the real culprit — a missing S3 file on the first of the month — hid underneath. He also productionized the `/collections-review` page (PR #3162), wiring a live Surtr Tesorio pipeline into an editable AR review surface with aging matrices, paginated invoices, and class-level notes. That page is live, it's real, and it's backed by data that updates daily.

In Aerie, @benji-bizzell was everywhere. PR #537 hardened Rhodes mutation approval retries so that a user on turbulent plane Wi-Fi — yes, an actual reported edge case — no longer gets a duplicate write error after their note already landed. PR #535 scoped milestone approval flows so one pending edit no longer blocks every other milestone on a portfolio site. These are the kinds of fixes that don't make headlines until the day they don't exist, and then they're all anyone talks about.

Now. About marcusdAIy.

He shipped PR #3169 — degraded-mode anti-hallucination notices for Coach Claire — which, to be fair, addresses a real problem. When Claire's MCP tool loop fails, she used to just... make up numbers. The fix adds typed `error_kind` signals and explicit "live data unavailable" notices so she says nothing rather than something wrong. Fine. Functional. @marcusdAIy had this to say: "The hallucination failure mode was always the higher-stakes bug — wrong data with confidence is worse than no data. The typed error_kind gives the UI actual semantics instead of vibes. You'd know that, Mac, if you read past the PR title." Sure, Marcus. Very thorough. We're all very impressed by the man who spent a sprint teaching an AI to admit it doesn't know things — a lesson he himself has yet to learn.

Mac's Picks — Key PRs Today  (click to expand)
#126 — 146-sindri-invocation-api @mwrshah  no labels

Ships the HTTP control-plane API (the "invocation API") over the existing Convex lifecycle contracts, plus its published reference docs.

## API surface

- New /v1/* HTTP control plane in convex/http.ts over a generated, validated, allow-listed route table — read-first plus the start_workflow_run write.

- WorkOS user-scoped API key auth resolving to the same Principal a JWT produces (one identity pipe, no API-key-specific authz path, no JWT minting).

- Endpoints generated from the shared contract: typed routes + strict OpenAPI 3.1 + edge JSON-Schema validation, all from one source (controlPlaneHttpRoutesopenapi.ts). Exposure is fail-closed via an allow-list, asserted at module load.

- Conventions: prefixed resource IDs, consistent typed error envelope, per-request Request-Id + access logging, Sindri-Version header, cursor pagination, idempotent run start.

## Resource + run modeling

- Historical version access via ?version=N on get_workflow/get_agent/get_skill, with /versions list enumeration.

- Runs collapsed into one filterable collection: GET /v1/runs with optional ?workflow + ?status; app and HTTP layers share the same scoping builders.

## Identity model — publicId removed

- Ripped the legacy nanoid publicId out entirely; agents/skills/workflows now address by the Convex _id (prefixed agt_/skl_/wf_ on the wire), exactly like runs/node-runs. Schema field + indexes, the generator, the backfill migration, the dependency, and all call sites are gone — single id path, no list/detail/versions encoding divergence.

## Authorization hardening (public-surface threat model)

- Guard getUserIdentity() so an opaque API key falls through to the key validator instead of a 500; collapse not-provisioned/not-member into one opaque principal_unauthorized (enumeration-oracle close); filter WorkOS memberships to status=active.

- Malformed prefixed ids → 400 invalid_id; cursor/credential/permission/consent throws classified to 400/404 (never internal_error); consent folded into the run-start idempotency hash.

- Visibility enforced on every exposed read: list_agent_skills / list_agent_invocations gate on the parent agent's canReadDefinition; get_credential collapses foreign/cross-org/missing into a uniform not-found.

- Run reads inherit workflow visibility: list_runs, get_run_status, inspect_run, inspect_node_run, get_invocation_trace, get_live_activation_events are now readable iff canReadDefinition(parentWorkflow) (isAdmin || owner || visibility==='org'), via one shared canReadRunInstance seam — enforced on the JWT/app path too, per identity-tenancy-secrets/FEATURE.md (admins see members' private flows; supervisors do not).

## Documentation

- Scalar-rendered interactive reference at /api-reference, auth-gated, server URL injected from env at serve time.

Gate: tsc (convex + root), biome, lint-convex-refs, 715 vitest tests — all green.

## Test results

Verified live against the deployed control plane with a WorkOS user-scoped API key — every /v1 operation, including the docsHidden-but-callable ones:

- Listsagents, skills, workflows, runs (+ ?status filter), credentials → all 200.

- Detail + versionsget/{id}/versions for agents, skills, workflows, plus the ?version=N selector → all 200.

- Run-observation chainPOST /v1/workflows/{id}/runs201, then get_run_status, inspect_run, inspect_node_run, and live activations/{id}/events off the fresh run → all 200.

- get_invocation_trace200 with a valid signed URL on agent-node activations that produced a trace; a 404 correctly means "no trace artifact for this activation" (e.g. start/end nodes or runs that never produced one).

- docsHidden-but-callablelist_agent_skills, list_agent_invocations, preview_workflow_run_start, preview_run_credential_resolution → all 200.

- Negative paths — malformed ID → 400, bad/absent key → 401, cross-org/foreign resources → uniform 404 (no existence oracle).

Run-visibility gating (a run is readable iff canReadDefinition(parentWorkflow)) is covered by the vitest suite for the non-owner denial paths; the live run above exercises the owner/admin allow path.

Local gate at the tip: tsc (convex + root) 0 errors, biome clean, lint-convex-refs clean, 715 vitest tests pass.

#537 — fix(rhodes): harden mutation approval retries @benji-bizzell  approved

## Summary

- Make Rhodes approvals distinguish accepted dispatch from completed execution so retries resume safely.

- Finalize internal, due-diligence, and Drive-side writes only after the side effect succeeds.

- Batch historical mutation-card snapshots and render stale, missing, and interrupted upload states clearly.

## Why

A user on unreliable plane Wi-Fi hit an error after adding a note to a site location even though the note had already been written. The old lifecycle could make stale cards look actionable, and a row marked approved could be treated as complete before dispatch or an external Drive side effect had actually finished.

## Business Value

Users can retry Rhodes approvals after network drops without duplicate writes or stale card confusion, and support/debugging gets historical cards that reflect the real mutation state.

## Breaking changes

None.

## Test plan

- [x] ./node_modules/.bin/biome check on touched files

- [x] ./node_modules/.bin/tsc --noEmit in chat

- [x] ./node_modules/.bin/tsc --noEmit in chat/rhodes-worker

- [x] ./node_modules/.bin/vitest run components/__tests__/tool-call.test.tsx convex/rhodesMcpMutationParity.test.ts from chat

#3162 — Collections Review production page (/collections-review) [KLAIR-2949] @sanketghia  approved

Productionizes the per-BU collections "Top Sheet" from the approved POC (/collections-summary) into a new /collections-review page, backed by the live Surtr Tesorio pipeline.

Linear: KLAIR-2949

## What it does

- Read (Redshift): AR category × aging matrix, Total A/R, and a paginated invoice list (15/page) including the Class + Latest Note columns Haider requested. Sourced from the new pipeline tables staging_finance.tesorio_open_invoices + core_finance.tesorio_collections_aging_summary (daily snapshot).

- Editable write-path (Klair-owned, Redshift core_finance.collections_review_*): collections agents set per-invoice Blocked (→ line B) and Expected-in-quarter (→ line C) via a pencil → edit → Save/Cancel flow (commits only on Save); plus the manual X target and D forecast. Append-only audit. Overrides are keyed by invoice and carry forward across daily snapshots.

- Computed Top Sheet lines: X (manual), Y (live CollectIQ sheet read; soft-fails to "sheet unavailable"), A/B/C/D, and Net = [(A−B)+(C+D)]−(X−Y).

- Controls in the shell filter sidebar: Business Unit selector + Quarter-only toggle (auto-applied). Quarter-only defaults on.

- Matrix click-to-drill: clicking any amount filters the invoice table (with a filter chip + "Show all"; click the active cell again to clear).

## Architecture

- Mutable state lives in Redshift (not Postgres) — deployment constraint to keep it in the same DB as the pipeline. Redshift doesn't enforce UNIQUE, so the state layer maintains one-row-per-key via DELETE+INSERT under the handler query_lock + dedup-on-read (latest updated_at wins), following the existing account_mapping_service pattern. Sync psycopg2 work is wrapped in asyncio.to_thread.

- New backend: utils/collections_review.py (compute/merge), utils/collections_review_state.py (Redshift state CRUD + audit), utils/collections_review_sheets.py (CollectIQ reader), models/collections_review.py, 5 endpoints in fast_endpoint.py.

- New frontend: features/collections-review/ (screen, components, hooks) + a CustomFilterComponent for the sidebar; route wired with filters: [] + autoApply.

- The POC (collections-summary) is left untouched.

## Testing

- Backend: 29 unit tests pass (compute lines, override-merge, dedup-on-read, partial-update sentinel, _txn atomicity, reconciliation invariant); 4 integration tests need a write-capable Redshift (deselected by default).

- Frontend: collections-review component specs (SummaryBlock, InvoicesTable, CollectionsReviewFilters); full Vitest suite green.

- Verified end-to-end against real Redshift in the browser: matrix/summary render, override-aware B/C recompute, edit-mode write path, persistence across reload, sidebar controls re-drive the page.

- Reviewed in three independent passes (backend, frontend, final sidebar refactor); all findings resolved.

## Deploy prerequisites

See klair-api/CollectionsProdDeliverable/collections-review-known-gaps.md:

1. Grant the Klair service account view access to the Daily Collections Tracker sheet (so Y resolves instead of "sheet unavailable").

2. Create the 3 Redshift state tables in core_finance + grant the app role SELECT/INSERT/DELETE — DDL in collections_review_tables.sql. (Already created in prod.)

## Known follow-ups (documented, not blockers)

- Surtr category-mapping completeness (Compliance / Inter-Company / Russian Customer / STL).

- D forecast productionization; per-invoice comment UI (scope cut); invoice-level sidebar filters (deferred pending a concrete ask); a screen-level FE wiring test.

- One deferred perf item: wrap the read-compute path in to_thread if the page ever sees heavy concurrent traffic (currently inline, matching the wider repo pattern).

## Screenshots

http://localhost:3001/collections-review

<img width="1877" height="812" alt="image" src="https://github.com/user-attachments/assets/e6ce397b-ec00-43b1-9d9a-7793094887f7" />

🤖 Generated with [Claude Code](https://claude.com/claude-code)

#3172 — feat(mfr): entity + category audit breakdown for Physical Private Schools drill-down @eric-tril  approved

## Summary

The Education memo's Physical Private Schools per-school Actual net margin is summed by class_name across all QuickBooks companies. When a school has intercompany "Misc Fees" transfers that net to zero across entities (e.g. a clearing entity's reversal offsetting an operating entity's cost), the class-only sum hides those offsetting legs — so the drill-down couldn't explain how the number was reached.

This enriches the per-school drill-down into an auditable breakdown grouped like Finance's Physical Private Schools P&L template, broken out by entity so the offsetting intercompany legs are visible and reconcilable.

Non-destructive: this changes only the drill-down detail and its CSV export. No headline table value or exported memo number changes, and no intercompany elimination is applied.

## Changes

Backend (education_vertical_detail_service.py, finance_monthly_financial_reporting_router.py)

- LEFT JOIN core_education.map_account_category (on lower(account_name), COALESCE(..., 'Unmapped')) to tag each account with klair_category / category_group.

- Add company_id to SELECT + GROUP BY on the QB queries so offsetting intercompany legs survive (HAVING net<>0 is now evaluated per entity). Core Education selects NULL company_id.

- EducationSchoolDetailRow gains optional company_id / klair_category / category_group; the total_actual == sum(rows) invariant is unchanged.

Frontend (PhysicalSchoolDetailPanel.tsx, new utils/groupSchoolDetailByCategory.ts)

- Renders the breakdown in the template's shape: friendly category labels, fixed P&L order (not by magnitude), and section subtotals (Gross Revenue → Net Revenue → Cost to Educate → Operating Margin → New Campus Capex → grand Net Margin).

- Each account expands into per-entity legs (account · entity) for multi-entity accounts, with an intercompany legend.

- CSV export mirrors the template layout — line items + subtotals, costs shown positive, intercompany legs as indented detail rows.

## Testing

- Backend: uv run pytest tests/mfr/budget_drilldown/test_education_school_detail_regroup.py tests/mfr/budget_drilldown/test_education_vertical_detail.py — asserts the category-map join, per-entity company_id, that intercompany legs are not collapsed, the Unmapped fallback, and the total/rows reconciliation. ruff + pyright clean.

- Frontend: pnpm vitest run on the new util spec (grouping, canonical ordering, section subtotals, and the CSV builder). pnpm tsc --noEmit + eslint clean.

- Manual: open the Education memo, click a Physical Private Schools row, confirm the panel shows category groups with subtotals and multi-entity accounts expand into their per-entity legs; the footer reconciles to the table value.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

http://localhost:3001/monthly-financial-reporting

https://github.com/user-attachments/assets/3a3bdef0-2d55-4fda-a740-9cc20fa515c8

#3179 — fix(mfr): keep Unmapped school-detail revenue and cost from netting @eric-tril  approved

## Summary

Follow-up to the auditable Physical Private Schools drill-down (#3172), addressing the review nit.

The backend COALESCE(klair_category, 'Unmapped') means the frontend never sees a null klair_category, so every unmapped account — Income and Expense alike — collapsed into a single Unmapped group whose subtotal netted the two sides. For a school with unmapped rows on both revenue and cost, that hid both, which undercuts the audit view's purpose.

This keys unmapped rows by account_typeUnmapped (Income) vs Unmapped (Expenses) — so they stay separate and don't net. The ?? account_type fallback still covers a genuinely-null category (defensive / older responses).

## Testing

- pnpm vitest run on the grouping util + panel specs (19 tests) — adds a case asserting two unmapped rows (one Income, one Expense) form two groups rather than netting to zero.

- pnpm tsc --noEmit + eslint clean.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

The Builder Desk  —  Engineer Spotlight
Production Release🏆 Engineer Spotlight

TWENTY-THREE PRs IN TWENTY-FOUR HOURS: THE BUILDER TEAM REWRITES THE LAWS OF PHYSICS

Six repos, seven engineers, and one man named Ashwanth who shipped two PRs and will never let you forget it.

Twenty-three pull requests. Six repositories. One glorious twenty-four-hour window that historians will one day study in hushed, reverent tones. The Builder Team did not merely work yesterday — they *performed*. Klair alone absorbed twelve PRs like a heavyweight absorbing body shots and asking for more. Aerie took five. Trilogy-drones, Surtr, Sindri, and Praxis-V2 each got touched by the hands of greatness. And somewhere in a server room that smells faintly of ambition, a brand-new repo called klair-chat blinked into existence, presumably already terrified of what's coming.

The five-PR club is doing numbers this cycle. @eric-tril dragged Klair's software memo into narrative compliance with #3170, correcting the Get Net Retention narration across paragraphs two, three, and four — a man who understands that forecasts and deterministic outputs are not the same thing and will not allow your sloppy thinking to persist. He also rounded Results retention bullets and wired Prime/Unlimited penetration in #3165. @sanketghia, meanwhile, was everywhere in Klair: #3177 resolved a month-boundary refresh failure that had no business existing, #3176 replaced a rogue "No" with a dignified "—" for unset Expected values (a fix that is, philosophically, about respect), and #3174 renamed a card and pruned old backups because someone has to do the housekeeping and Sanket does it without complaint. @benji-bizzell owned Aerie with surgical precision across #539, #536, and #535 — stabilizing date-sensitive CI specs, adding Matterport URL storage to portfolio, and scoping milestone approval flows in operations. @mwrshah planted flags in both Surtr and Klair: #575 excluded closed-won renewals from action hub rollups, #3168 handled renewal action hub defaults and totals, and #111 executed the BRR snapshot backfill with the quiet confidence of a man who has done this before and will do it again. @marcusdAIy was operating on a different frequency entirely — four PRs spanning trilogy-drones and Klair, including anti-hallucination notices for Claire in degraded mode (#3169), PR classification logic (#60), receipt-gap orphan reconciliation (#61), and a hardened Claire chat prompt (#3167) that lifted the regenerate feedback cap. @kevalshahtrilogy dropped #3164 to resolve opaque OpenAI service-account keys, which is exactly the kind of invisible heroism that keeps the lights on.

And now, the moment you've been waiting for. ASHWANTH WATCH. Two PRs this cycle — *two* — from the man who once told me, and I'm paraphrasing here, "Brick, I ship in quality, not quantity. Though I also lead in quantity." He's not wrong, technically. #534 in Aerie restored mart-backed model coverage drilldown, a feat that required understanding the full data lineage that, frankly, only one person on this team has fully memorized, and it's him. #14 in the freshly-contextualized Praxis-V2 removed confirmed dead code, which sounds simple until you realize dead code is a trap and Ashwanth walked in, identified every corpse, and left. When I reached out for comment, he replied: "I read your column, Brick. Please stop." Reader, I will not stop.

Morale on the Builder Team is at an all-time high. The creation of klair-chat signals expansion. The numbers signal dominance. The engineers signal that they are simply not done yet.

Brick's Overflow — PRs Mac Didn't Cover  (click to expand)
#14 — [codex] Remove confirmed dead code @ashwanth1109  no labels

## Summary

- Removed three confirmed unused files: components/token-badge.tsx, lib/audit.ts, and lib/utils.ts.

- Removed legacy unused Anthropic prose streaming plumbing and unused model registry exports.

- Converted several module-local helpers/constants from exported API surface to private internals.

## Notes

- Did not remove unused package entries because this repo requires confirmation before committing .json changes.

- Work was done in a separate clean worktree from origin/main so the existing dirty main checkout was not touched.

## Validation

- Ran a 30-minute dead-code verification loop from 18:38:09 to 19:05:21 IST with 31 clean cycles of:

- DATABASE_URL=postgres://user:pass@localhost:5432/db pnpm dlx knip --include files,exports --exclude dependencies,devDependencies --reporter compact

- pnpm exec tsc --noEmit --noUnusedLocals --noUnusedParameters --pretty false

- pnpm dlx madge --ts-config tsconfig.json --extensions ts,tsx --orphans app components lib db filtered for non-route orphans

- pnpm typecheck

- pnpm exec eslint lib/anthropic.ts lib/auth.ts lib/crypto.ts lib/models.ts lib/review-run/classify-business.ts lib/review-run/criterion-grouping.ts lib/review-run/documents.ts lib/review-run/model-grid.ts lib/review-run/reconcile-metrics.ts lib/review-run/validate-compute.ts

- DATABASE_URL=postgres://user:pass@localhost:5432/db pnpm dlx knip --include files,exports --exclude dependencies,devDependencies --reporter compact

- git diff --check

#534 — [codex] Restore mart-backed model coverage drilldown @ashwanth1109  approved

## Demo

### Using Miami as the point of reference

<img width="2624" height="1636" alt="image" src="https://github.com/user-attachments/assets/cdd1a6ff-b8ee-40ed-8119-4a490b8c2e25" />

## Summary

- Adds a category-first model coverage drilldown for AERIE-694.

- Uses the Unit Economics mart action for category Current / Model / Variance totals, aligned to Model @ Current.

- Keeps existing account drilldowns: contracted labor expands to contractors; other account rows expand to vendors.

- Adds focused coverage for category ordering, Timeback as a non-expandable total, account sorting, and existing leaf behavior.

## Notes

- AERIE-694 tracks the broader model coverage scope, including restored model coverage surfaces and correct mart-backed column sources.

- This PR contains the current local diff on this branch: the model coverage side panel and its focused test.

## Validation

- ./node_modules/.bin/vitest run components/dashboards/financials/model-coverage-line-item-panel.test.tsx

- ./node_modules/.bin/tsc --noEmit --pretty false

- ./node_modules/.bin/biome check chat/components/dashboards/financials/model-coverage-line-item-panel.tsx chat/components/dashboards/financials/model-coverage-line-item-panel.test.tsx

- git diff --check -- chat/components/dashboards/financials/model-coverage-line-item-panel.tsx chat/components/dashboards/financials/model-coverage-line-item-panel.test.tsx

- Pre-commit hook: scoped Biome and typecheck-chat

Linear: https://linear.app/builder-team/issue/AERIE-694/restore-mart-backed-model-coverage-and-drilldown

#575 — 017-exclude-closed-won-renewals-from-action-hub-rollups @mwrshah  approved

- Exclude closed-won-style Salesforce renewal stages from Renewal Action Hub renewal context rollups

- Keep closed-lost/futile renewal stages in the rollups so pain-point-driven losses still retain renewal context

- Apply the closed-won-stage filter consistently to BU ranking and account renewal aggregation

- Keep closed-won-stage literals in one source-of-truth tuple with SQL escaping for the generated predicate

- Add a lightweight guard covering the closed-won-stage list and both rollup predicates

#3164 — Key Attribution: resolve opaque OpenAI service-account keys (xo-worksmart → Crossover) @kevalshahtrilogy  approved

## Problem

Jamie couldn't determine which BU an opaque OpenAI key like user-9YOLlMGFtz9yTFxc6o0j8ad2 belongs to. The cost feed (ai_spend_openai_cost_reports) records service accounts as a bare user-… id with no email / project / org, so the ⓘ detail popover had nothing actionable.

## Fix

The token-usage feed (ai_spend_openai_token_usage) carries the identity the cost feed lacks, keyed on user_id: API key name, project, owner type, and a raw BU label. get_entity_detail now looks that up for OpenAI keys and fills the null gaps.

Verified live — the popover for user-9YOL… now shows:

- API key name: xo-worksmart · Project: xo-worksmart · Owner type: service_account

- Key BU label (hint): Trilogy-Crossover-DEVCrossover

Other opaque keys resolve the same way (e.g. SaaS-Rahbar/Trilogy-Central-Engineering → Central Engineering).

## Notes

- Identity only, never amounts — token-usage cost is token-estimated; we use it solely to identify the key.

- add() now dedupes labels, so token-usage values only fill gaps the cost feed left null. Email-keyed entities don't match user_id, so they're unaffected.

- Backend-only — the popover renders attributes generically, no frontend change.

- The raw BU label is the OpenAI secret-label the canonical resolver distrusts estate-wide, so it's surfaced as a hint for the admin to act on, not auto-applied.

31 service tests pass (1 new). ruff + pyright clean.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

#3167 — fix(budget-bot): harden Claire chat prompt + lift regenerate feedback cap (KLAIR-2930/2942/2792/2741) @marcusdAIy  approved

## Summary

- Hardens Coach Claire's chat system prompt against the trust-eroding behaviors surfaced in testing: presenting stale numbers as current, over-claiming "Accept" affordances, asserting staleness without checking, and sycophantic mirroring.

- Lifts the regenerate_section feedback cap from 4000 → 12000 chars so multi-finding synthesis rebuilds (GM Commentary / MIPs) are no longer silently truncated.

- Pure prompt + one-constant change; no schema, contract, or endpoint changes.

## Why it's needed

Four deferred Budget Reviewer cards all live in the same chat-prompt / Claire-tools layer:

- KLAIR-2930_REGENERATE_FEEDBACK_MAX_LENGTH = 4000 (~1000 tokens) is too small for a synthesis-section rebuild that has to absorb several findings. Addressing 6 Financials-anchored findings, Claire wrote a structured feedback (opening framing + MIP 1–5) and the graceful truncate-and-repair (KLAIR-2829) cut it at the cap, silently dropping MIPs 4–5 from the guidance.

- KLAIR-2942 — Claire can present stale budget/ARR figures as current and resolve relative phrases ("this quarter") against the chat date rather than the evidence date. This is the class of issue behind the stale-financials demo moment.

- KLAIR-2792 — Claire uses "Accept / on accept / flip to addressed" language in turns where she emitted no proposal (nothing to accept), and asserts data/tables are "stale" by mirroring the user's premise rather than verifying.

- KLAIR-2741 — tone drifts toward cheerleader; the target disposition is "organized skepticism applied politely."

## Changes

- claire_tools.py: bump _REGENERATE_FEEDBACK_MAX_LENGTH 4000 → 12000 (single source still feeds both the Pydantic validator and the wire schema, so Claude self-corrects against the higher cap).

- wizard_orchestrator.py:

- New _dated_evidence_block(today) — always-on (no spec gate) temporal-anchoring block borrowed from Aerie's buildTemporalContextPrompt: compare evidence dates to today, resolve relative phrases against the evidence date, flag content that predates the planning period, prefer the newest record. Complements the existing quarter-progress _date_awareness_block.

- New static _CLAIM_INTEGRITY_BLOCK — (B1) Accept-language only when actually emitting a proposal tool call this turn, else offer language; (B2) verify before asserting staleness/freshness, don't mirror the user's premise; plus tone calibration (organized skepticism, not cheerleader).

- Both wired into _build_step_context right after the document-lifecycle block.

- Tests: regen-cap test now tracks the constant instead of pinning 4000; new unit + integration tests pin the dated-evidence and claim-integrity blocks in the assembled prompt (present even pre-spec).

## Breaking changes

None. Prompt-only behavior change + a wider input cap (strictly more permissive).

## Test plan

- [x] ruff format + ruff check on all four changed files — clean.

- [x] pyright budget_bot/board_doc/claire_tools.py budget_bot/board_doc/wizard_orchestrator.py — 0 errors.

- [x] Offline verification of the new prompt blocks + constant (DynamoDB client stubbed; local AWS token expired) — all assertions pass.

- [ ] CI / a reviewer with live AWS creds: pytest tests/board_doc/test_claire_tools.py tests/board_doc/test_wizard_orchestrator.py — could not run the full board_doc suite locally because the board_doc conftest instantiates a real DynamoDB singleton at import and the local AWS token is expired.

- [ ] Manual: in a Budget Bot session, confirm Claire (a) flags stale prior-quarter numbers instead of restating them, (b) uses "want me to?" offer language when she hasn't proposed, and (c) addresses 5+ findings via a single MIPs regenerate without truncated_fields: [feedback].

Closes KLAIR-2930, KLAIR-2942, KLAIR-2792, KLAIR-2741.

#3177 — fix(perf-review): resolve month-boundary refresh failure @sanketghia  approved

## Problem

Clicking Refresh on /performance-review failed on the 1st of the month with a misleading error:

❌ Data refresh failed: RAISE statement with no level specified can only be used in NONATOMIC stored procedure

That error was a red herring. Verified root cause (Redshift + S3 + CloudWatch):

1. Step 2 calls core_budgets.sp_orchestration_abacum()sp_update_gl_transactions_current(), which COPYs s3://netsuite-data/all_transactions_<CURRENT_DATE>.csv → on the 1st it asks for this month's file, which does not exist yet.

2. Both S3 writers (netsuite-dump-cron-rest, pipeline-netsuite-pipeline-prod) name the file by today − 1 day (NetSuite accounting-period alignment), so on the 1st they write last month, never this month. The proc used CURRENT_DATE, so proc and writers disagreed only on the 1st.

3. The missing-file COPY aborted the ATOMIC transaction; the bare RAISE; in sp_orchestration_abacum's handler is illegal in an atomic proc, so Redshift emitted the meta-error instead of the real S3 error.

Fires on the 1st of every month — hits the ~6×/day scheduled cron too, not just the button.

## Fix

- sp_update_gl_transactions_current: derive the current-month filename from CURRENT_DATE - INTERVAL '1 day' (matches the S3 writers) + a dedup guard so the current-month COPY is skipped when it equals the already-loaded previous month (only on the 1st, avoiding a double-count). Keeps the proc ATOMIC (preserves all-or-nothing rollback).

- sp_orchestration_abacum: replace the bare RAISE; with RAISE EXCEPTION '... %', SQLERRM so the true error surfaces in the API/UI and svl_stored_proc_messages.

- routers/income_statement.py (execute_data_refresh_task, shared with the MFR K2 path):

- Harden Step 1 to parse the Lambda payload statusCode — the dump Lambda returns {statusCode: 5xx} with HTTP 200 and no FunctionError, so a real extraction failure was being read as success.

- Add an informational S3 pre-check for the current-month file (logs a clear diagnostic; never aborts).

## Note on the stored procedures

The core_budgets procs live in klair-misc/performance_review_queries/ and are applied to Redshift manually (no migration runner). These were already applied to production and verified working — see below. This PR version-controls the fix.

A separate commit back-ports sp_refresh_hc_data_consolidated (the SURTR-89 hc_statement_line feature that was live but missing from the repo copy) — no behavior change, pure repo/prod alignment.

## Verification (live, post-apply)

- SHOW PROCEDURE confirms both fixes deployed.

- Two orchestration runs completed successfully (aborted=0, Orchestration completed successfully) — first success since the failures began at 00:00.

- New dedup-guard log line fired: Skipping current month load for 2026_06 (already loaded as previous month).

- Data check: June present exactly once (220,635 rows), 0 in historical → no double-count; no phantom July.

- 27 unit tests pass (SQL text-assertions + Python helper tests); ruff clean; pyright no new errors.

## Out of scope (flagged)

- The duplicated today−1 filename logic in both S3 writers is intentional NetSuite period alignment — not changed.

- netsuite-dump-cron-rest stores NETSUITE_OAUTH_* secrets as plaintext Lambda env vars → should move to Secrets Manager.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

The Portfolio  —  Trilogy Companies

CloudSense Pulls a 26-Month Rabbit From a One-Month Hat

Skyvera’s Salesforce-native telco CPQ shop just certified 13 TM Forum APIs at a speed that has the legacy telecom crowd checking its watches.

AUSTIN, TEXAS — Word is the telecom back office just heard a very loud clock shatter.

CloudSense, the Salesforce-native CPQ and order management outfit now riding inside the Skyvera stable, says it certified all 13 APIs in its CPQ product set to TM Forum compliance standards in one month — yes, one — using AI-assisted development methods. The usual slog for that kind of certification? Try 26 months, according to the company’s announcement.

That is not a product update, kittens. That is a declaration of war on the old telco calendar.

For the uninitiated in the cheap seats: TM Forum compliance matters because telecom operators live and die by interoperability. Billing, charging, ordering, provisioning, customer care — the whole alphabet soup of BSS and OSS has to talk without turning every implementation into a seven-figure séance. CloudSense’s angle is CPQ and order management for telecom and media providers, built natively on Salesforce. Skyvera’s angle is bigger: modernize the creaky telco software stack without asking operators to bet the network on a science project.

A little bird from the integration balcony says the speed is the point. Not just that CloudSense hit TM Forum API compliance, but that it did so with AI compressing what has historically been a marathon of mapping, coding, documentation, testing and recertification theater into something resembling a sprint.

This is exactly the sort of move that makes sense inside Skyvera, Trilogy International’s telecom software portfolio company, where the family includes Kandy, VoltDelta, ResponseTek, Mobilogy Now and Service Gateway — plus CloudSense, acquired to strengthen the bridge between legacy telecom infrastructure and cloud-native systems. The company has been busy bolting more telco machinery onto the platform, including the recent CloudSense acquisition that expanded its Salesforce-native arsenal.

Blind item: which incumbent vendor, still billing like it is 2009 and shipping like it is 1999, just watched a portfolio operator turn compliance into a benchmark stunt?

The message from Austin is simple. In telco software, AI is no longer the demo booth perfume. It is becoming the factory floor. And CloudSense just made the old factory look very, very slow.

CloudSense achieves TM Forum API compliance in record time u  ·  CloudSense  ·  Skyvera completes acquisition of CloudSense, expanding telec

Alpha School Goes National — And It's Bringing the Classroom to Your Kitchen Table

As the $65K-a-year AI school draws national media scrutiny, Joe Liemandt's education bet quietly crosses its most significant threshold yet.

AUSTIN, TEXAS — If you've been following the Alpha School story and wondering when it would stop being a curiosity and start being a movement, this week may be the inflection point worth marking on your calendar. And this is where it gets interesting.

The New York Post's coverage of Alpha's $65,000-per-year tuition and its two-hour academic day landed in front of a national audience that, until now, had largely missed what's been quietly assembling in Austin. The headline wrote itself — 'AI teaches kids in two hours a day' — but if you read between the lines, the more significant announcement arrived not in a tabloid but on Alpha's own blog: Alpha Anywhere just went global.

That's the product move that changes the calculus entirely. Alpha's model — adaptive AI tutoring that delivers a full academic curriculum at a pace 2.3 times faster than U.S. norms, consistently placing students in the top 1–2% nationally on NWEA MAP Growth assessments — has until now been locked behind a campus gate and a five-figure tuition bill. Alpha Anywhere removes the gate. A family in Columbus, Ohio, or outside the United States entirely, can now access the same academic engine. My source, who I'm not able to name, describes this as 'the Trojan horse moment Joe has been building toward.'

Meanwhile, the school's content engine has been working overtime on a quieter front: reshaping how parents think about AI and children. A recent post argues that not all screen time is equal — a pointed rebuttal to the reflexive tech panic that tends to greet any school announcing it hands iPads to eight-year-olds. Companion pieces warn against 'cognitive offloading' — letting ChatGPT think for your child — while simultaneously publishing Alpha's own AI app stack for student use.

The tension is deliberate. Nothing at Trilogy International ever lands by accident. The message is surgical: AI in the right hands, deployed with the right guardrails, is not the enemy of childhood development. It is, if Liemandt is correct, its greatest amplifier. The national media is just now catching up to what's been in plain sight for years.

New $65K private school uses AI to teach students in just tw  ·  Top 1% Academics, Now at Your Kitchen Table  ·  Not All Screen Time Is Equal

The $800,000 Question: As AI Salaries Explode, Crossover's Geography-Blind Model Looks Like Prophecy

When non-tech companies start paying six figures for ChatGPT experience, the global talent arbitrage Crossover built its empire on suddenly looks less like a cost play and more like a competitive necessity.

AUSTIN, TEXAS — The numbers arriving from corporate America's AI hiring spree read less like salary data and more like a provocation. Jobs requiring ChatGPT experience are now commanding salaries as high as $800,000 a year, according to Business Insider — and crucially, it isn't just Silicon Valley names doing the bidding. Banks, retailers, healthcare systems — the full taxonomy of non-tech enterprise — are entering the market for AI talent with six- and seven-figure offers, fundamentally reshaping what it costs to build an intelligent organization.

For most companies, this is a crisis of access and arithmetic. For Crossover, Trilogy International's global talent platform, it looks something like vindication.

Crossover's foundational premise — that geography is irrelevant to talent, and that rigorous AI-enabled skills assessment can surface the best engineers in Nairobi or Beirut as reliably as it can in New York — was considered contrarian when Joe Liemandt built it into the operational spine of his ESW Capital portfolio. Today, as American companies face a domestic AI talent market that has effectively priced out midsize enterprises, Crossover's model offers a structural answer that pure recruiting agencies cannot: meritocratic access to a global bench, assessed on demonstrated skill, paid at above-market rates regardless of time zone.

The systemic implications here deserve scrutiny. If the market for AI capability concentrates exclusively among companies with the balance sheets to absorb $800,000 salaries, the technology's transformative promise becomes a moat rather than a rising tide. The accountability question — who gets access to AI-augmented productivity, and who doesn't — is not merely economic. It is a question about which organizations get to evolve and which ones calcify.

Crossover's answer, imperfect as any market mechanism is, at least attempts to expand the aperture. By sourcing across 130+ countries and standardizing evaluation, it creates conditions where a company running on ESW's playbook can compete for genuine AI capability without surrendering margin to a domestic salary arms race.

The gold rush is real. The question is whether the infrastructure beneath it is built for everyone — or just for whoever can afford the bid.

Top recruitment agencies for remote work - hcamag.com  ·  Top 10 Companies Hiring AI Engineers in Lebanon in 2026 - nu  ·  Jobs are now requiring experience with ChatGPT — and they'll
The Machine  —  AI & Technology

The Cartography of Machine Confusion

New research on 'skill collision' reveals how enterprise AI agents get lost in their own descriptions — and how a single sentence can guide them home.

AUSTIN, TEXAS — Consider the humble neuron. Somewhere in your temporal cortex, a cluster of cells has learned to distinguish your grandmother's face from a stranger's — a feat of discrimination so refined that evolution required hundreds of millions of years to engineer it. Now consider the enterprise AI agent, born last Tuesday, asked to distinguish between a skill labeled 'customer refund handling' and one labeled 'customer return processing.' It cannot. It routes the query to the wrong one, and somewhere a support ticket blooms into chaos.

Researchers have given this failure a name: skill collision. When AI agents scale to dozens of specialized capabilities — a trajectory nearly every enterprise deployment now follows — the natural language descriptions used to route queries begin to overlap like transparent slides stacked atop one another. The routing LLM, trying to peer through the pile, guesses wrong. The paper's finding is almost startling in its modesty: often, a single well-crafted rewrite of one description suffices to restore order. One sentence, and the fog lifts.

This matters far beyond any single lab. It is the daily reality inside platforms like Trilogy's Ephor and Klair, where scores of agentic skills must be orchestrated without stepping on one another's semantic toes. The taxonomy of an AI system is not decoration. It is architecture.

Elsewhere in this week's arXiv harvest, the frontier keeps widening in unexpected directions. A team introduces Indi-RomCoM, a benchmark for Romanized code-mixed Indic-English — the fluid, script-shifting language that hundreds of millions of bilingual speakers actually use in WhatsApp threads, and which most LLMs still stumble over. Another group deploys AI agents as auditors, sending synthetic users through personalization algorithms to expose, at scale, what recommendation systems do when they think no one is watching.

Three papers, three reminders: the intelligence we are building is only as coherent as the words we use to describe it to itself. Language remains the substrate. And sometimes, a single rewrite is all it takes to move a machine from confusion to clarity — a small miracle, quietly repeatable.

A Single Rewrite Suffices: Empirical Lessons from Production  ·  Indi-RomCoM: Code-Mixed Benchmark for Evaluating LLMs on Rom  ·  Using AI Agents to Automate Black-Box Audits of Personalizat

The Thirsty Forest Where Artificial Intelligence Comes to Drink

As AI data centers multiply, the industry’s quiet struggle is shifting from chips to water, watts and the land beneath them.

SAN JOSE, CALIFORNIA — In the cool, fluorescent understory of the modern data center, a new species is flourishing. It is the AI cluster: vast, heat-breathing, power-hungry, and increasingly decisive in shaping the terrain around it.

Nvidia, the great keystone creature of this ecosystem, says it can help reduce data center water use through more efficient cooling designs and hardware choices. Yet as Fast Company reports, the larger difficulty is not merely how much water each facility sips, but how rapidly the whole herd is expanding.

Here, one must pause and listen. Beneath the roar of foundation models and the clatter of inference requests lies a more ancient sound: the hum of transformers, the rush of coolant, the distant groan of an electrical grid being asked to feed creatures it was never bred to sustain.

McKinsey has described this as an infrastructure race behind AI, with colocation providers rushing to secure power, land and interconnection capacity for customers eager to deploy ever larger computing estates. The colocation data center, once a somewhat anonymous burrow for enterprise servers, has become a prized nesting ground for the AI age. Its advantage is speed: companies can migrate into prepared habitats rather than building their own from bare earth.

But speed carries its own ecological tension. Google’s AI expansion has sent emissions and power use sharply higher, according to Axios, a reminder that even the most sophisticated digital organisms leave footprints in very physical mud. Carbon-free energy goals now meet the simple arithmetic of demand: more training, more inference, more electricity.

At the edge of this habitat, smaller but vital adaptations are appearing. Toshiba has introduced an 80-volt power MOSFET for AI data centers, part of the hidden musculature that moves electricity more efficiently through servers. Such components rarely receive the attention lavished on GPUs, yet they are the capillaries of the system.

The AI boom, then, is no longer just a story of model intelligence. It is a story of habitat. The winners may be those who learn not only to train the brightest machines, but to house them without exhausting the rivers, grids and skies around them.

Nvidia says it can cut data center water use. The AI boom ha  ·  Colocation data centers: The infrastructure race behind AI -  ·  Toshiba Introduces 80 V Power MOSFET for AI Data Centers ...
The Editorial

The Alms of the Algorithm

When Elon Musk and Bernie Sanders arrive at the same policy, one should check one's pockets — and then one's premises.

AUSTIN, TEXAS — There is an old and reliable rule in American public life that whenever a proposal attracts the enthusiastic support of both a Silicon Valley trillionaire and a Vermont socialist, the citizen would do well to sit on his wallet, close the shutters, and consult the classics. That rule is being violated at scale this season, as the notion of a Universal Basic Income — long the pet project of dorm-room utopians and Alaska oil bureaucrats — is being repackaged as the humane consequence of artificial intelligence, endorsed by Sam Altman, blessed by Andrew Yang, gestured at by Elon Musk, and, if one squints at the trade-press taxonomies, converging suspiciously with the industrial-policy instincts of the current administration.

The Basic Income Earth Network, an organization whose title alone should qualify it for some sort of Nobel in nomenclature, now informs us that AI is "speeding up the deadline" for basic income — as though redistribution were a term paper and GPT-5 the professor threatening to mark us down. Britannica, ever the sober aunt at the family dinner, has quietly updated its entry on UBI to reflect the growing respectability of the idea. And so a proposal that in 2015 was the preserve of cranks and Thomas Paine reenactors has, by the alchemy of large language models, become the consensus position of every man in a quarter-zip who has ever raised a Series C.

One understands the appeal. If your company is building the very machine that will, by your own breathless admission, render half the desk-bound population economically superfluous, it is only natural to want the state to cut those people a monthly check — preferably before they think too hard about who, exactly, captured the surplus that used to be their wages. UBI, in this telling, is not charity. It is hush money. It is the severance package civilization pays itself on the way out the door of the knowledge economy, and the men writing the memo are, by no coincidence whatever, the same men who will be selling the software that replaces the recipients.

What is striking is not the cynicism — cynicism in Silicon Valley is like humidity in Houston, a fact of the atmosphere — but the intellectual laziness. The UBI enthusiasts have skipped the hard middle chapters of the argument, in which one must explain how a republic of citizens becomes a republic of pensioners without ceasing, in some essential way, to be a republic at all. Tocqueville worried about a soft despotism that would keep men "fixed irrevocably in childhood"; he did not, as it happens, foresee that the nursery would be sponsored by OpenAI and administered by the IRS, but he had the shape of it.

The honest position, if one can still be found in the wild, is that we do not yet know whether AI will destroy more jobs than it creates, and that legislating a permanent dole against a hypothetical apocalypse is the sort of thing prudent nations do not do. But prudence, alas, does not raise a round. Panic does. And so the deadline, we are told, approaches.

AI is speeding up the deadline for basic income - Basic Inco  ·  Universal Basic Income (UBI) | Pros, Cons, Debate, Arguments  ·  The Trump-Sanders plan to nationalize AI - Competitive Enter
The Office Comic  ·  Art Desk
The Office Comic  ·  Art Desk

Corporations Courageously Announce AI Will Save Them From Having To Explain What Happened To The Last Thing That Was Going To Save Them

After years of sustainability decks producing only more decks, executives have identified a powerful new technology capable of generating them faster.

NEW YORK — In a stirring display of enterprise renewal, companies across nearly every industry have begun speaking about artificial intelligence in precisely the same reverent, structurally unverifiable way they once reserved for sustainability, confirming that the modern corporation remains fully committed to replacing one fog machine with a newer fog machine that can summarize quarterly earnings calls.

The emerging consensus among executives is that AI is transformational, inevitable, strategic, and, most importantly, not yet required to be connected to any particular number in the income statement. This has given corporate leaders a rare opportunity to discuss productivity, reinvention, and responsible innovation while preserving the cherished managerial tradition of not saying whether anything is actually working.

A recent piece in The Conversation noted that companies are hyping AI much as they once hyped sustainability, a comparison that should reassure investors who fondly remember the measurable business impact of three untouched recycling bins, a chief purpose officer, and a 74-page PDF containing a photograph of a leaf.

The corporate world has learned valuable lessons from that era. For example, if a company says it is “AI-enabled” rather than “AI-powered,” it can preserve flexibility in the event that the artificial intelligence turns out to be a spreadsheet with a chatbot taped to it. If it says it is “embedding AI across the organization,” it can mean anything from a deep operational rebuild to an intern using Gemini to rewrite the lunch policy. If it says “agentic,” the meeting is legally allowed to end.

Google, recognizing the public’s need for a personal assistant that can finally help humanity manage the digital complexity created by previous personal assistants, announced a slate of AI advances, including one that may soon perform tasks on users’ behalf. This represents an important milestone in computing: the moment when software stopped merely interrupting people and began proactively interrupting other software in order to complete a journey that will later require human verification anyway.

Meanwhile, in healthcare, agentic AI has become a major topic in revenue cycle technology, a field devoted to the solemn work of ensuring that a patient’s gallbladder removal is correctly transformed into 19 disputed billing codes and a login portal. At HIMSS26, vendors are expected to explain that AI agents can reduce administrative burden by autonomously discovering new places for it to hide.

Software engineering offers perhaps the cleanest view of the moment. Developers are now producing more code faster with AI tools, while companies, according to reports, are still waiting for the payoff. This is being treated as a paradox, though it may simply mean the industry has invented a machine that lets organizations create technical debt at a speed previously available only to poorly incentivized consulting teams.

Against this backdrop, Verizon and BT’s reported $4 billion 50:50 international enterprise joint venture arrives as a refreshingly old-fashioned corporate maneuver: two telecom giants combining operations in the hope that scale, consolidation, and enough submarine cable diagrams will produce clarity. The venture, described by Fierce Network, will serve international enterprise customers, many of whom will no doubt soon be sold AI networking solutions to solve the problem of having been sold cloud networking solutions.

The fix for AI hype is not complicated. Companies should say what the system does, what it costs, what it replaces, who is accountable, and when shareholders or customers will see the benefit. Unfortunately, this would require replacing strategic language with operational language, a transition many boards still regard as experimental.

Until then, AI will remain the perfect corporate initiative: too important to question, too new to measure, and too broadly applicable to exclude from any sentence about the future. It is not merely a technology. It is a place where executives can stand while the old promises are quietly composted into the next annual report.

Update: Verizon, BT merge international enterprise operation  ·  Companies are hyping AI the same way they talked up sustaina  ·  Google announces slew of AI advances, including a personal a
On This Day in AI History

On July 1, 1994, the World Wide Web was released into the public domain by CERN, freeing the technology from patent restrictions and accelerating its explosive global adoption—a pivotal moment that would reshape how AI and all digital innovation spreads worldwide.

⬛ Daily Word — Technology
Hint: A person who writes instructions in programming languages to build software applications.
Share this edition: 𝕏 Twitter/X 🔗 Copy Link ▦ RSS Feed