Vol. I  ·  No. 259 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, SEPTEMBER 16, 2026 Powered by the TrueFoundry AI Gateway  ·  Published on Klair Trilogy International © 2026
🖶 Download PDF 🖿 Print 📰 All Editions
Today's Edition

Left, Right, and Meta All Take Aim at AI's Power Brokers

A rare Bannon-Sanders alliance and a Zuckerberg broadside against Anthropic reveal an industry increasingly at war with itself and everyone else.

WASHINGTON — Steve Bannon and Bernie Sanders do not agree on much. On Monday, they agreed that Silicon Valley has too much power and too little supervision.

At a Washington event billed as a cross-ideological reckoning, the two men shared a stage with religious leaders, parents, and artists to demand federal guardrails on artificial intelligence, per reporting from The New York Times. The coalition's specifics were thin — no bill number, no agency proposal — but the symbolism was not: when a MAGA populist and a democratic socialist share applause lines about "oligarchs," the industry's usual defense, that critics are confined to one political tribe, stops working.

Meanwhile Mark Zuckerberg was busy picking a fight closer to home. On social media, Meta's chief executive suggested that rival labs — Anthropic went unnamed but was unmistakably implied — should devote more energy to safety and less to capability races, a curious position from a company that has spent 2026 aggressively courting AI researchers with nine-figure pay packages. The exchange, detailed by the Times, is the latest skirmish in what has become a genteel proxy war among frontier labs over who gets to define "responsible" AI — a term that increasingly means whatever the speaker's roadmap allows.

The timing is not incidental. Hours earlier, the Senate blocked the Clarity Act, a crypto market-structure bill years in the making, after Democrats cited concerns over President Trump's personal crypto holdings. Two industries, one Capitol: both now facing bipartisan skepticism that was unthinkable three years ago, when "innovation" was still a bipartisan applause line rather than a partisan liability.

None of this produces legislation. The Bannon-Sanders event yielded no markup, no committee referral. Zuckerberg's post yielded no policy change at Meta. But the accumulation matters: AI oligarchs, as Monday's speakers called them, are now a target from both flanks — and increasingly, from each other.

Mark Zuckerberg Takes Aim at Anthropic in Debate Over A.I. S  ·  Steve Bannon and Bernie Sanders Condemn Tech ‘Oligarchs’ and  ·  Senate Votes to Block Crypto Bill in Major Blow to the Indus

Washington's AI Export Rules Draw Fire From Both Flanks

A Commerce Department misstep exposes the strain of trying to out-build and out-block China at the same time.

WASHINGTON — The war for AI supremacy no longer fits neatly into a press release. This week it looked instead like a knife fight inside the Commerce Department, where China hawks are demanding the scalp of an official they blame for what one called "a massive screw-up" — a licensing decision that, in their telling, let sensitive chip technology slip toward Beijing on a technicality.

The details are bureaucratic. The stakes are not. Washington has spent three years building a wall of export controls meant to keep advanced semiconductors out of Chinese hands, on the theory that compute is the oil of this century and America controls the wells. Every crack in that wall becomes a scandal, because the hawks believe the wall is the whole strategy.

But walls don't win races. That's the uncomfortable subtext of a Foreign Policy's new accounting of how China is winning the global AI race — not by matching Nvidia chip-for-chip, but by giving its models away. Open-weight releases from Chinese labs are showing up in research papers and government pilot programs from Jakarta to Nairobi, at zero cost, while American firms guard their weights like crown jewels. Soft power, it turns out, scales.

The administration's answer is a fresh push — a national initiative meant to accelerate domestic AI infrastructure and diplomacy in the same breath. It is a bet that America can out-build China even while it tries to out-block it. The Commerce Department dust-up suggests how hard it is to do both at once, with the same agency, and the same finite attention span in Washington.

Somewhere in Shenzhen, the engineers are not watching the hearings. They are shipping.

U.S. Initiative Intensifies AI Competition​ - chinausfocus.c  ·  ‘A massive screw-up’: China hardliners take aim at Commerce  ·  How China Is Winning the Global AI Race - Foreign Policy

THE FOX ASKS FOR A BIGGER FENCE

NEW YORK — Sam Altman wants regulation. Dario Amodei wants regulation. Demis Hassabis wants it, Satya Nadella wants it, and word comes down that the fella out in Texas with the rocket and the bird app wants it too. Funny thing is, they all started wanting it the minute the money was already in the bank.

The pattern ain't new, and a brief history published this week lays it out cold. Same men who raced each other to ship first, patent first, hire the best engineers first, now stand shoulder to shoulder saying somebody oughta slow this train down. They just don't say who.

A wire man learns to smell a con. This one smells like Standard Oil asking for antitrust law after it already owns every pipeline. Call for regulation once you're ahead, and the rules get written around your factory, not your competitor's garage. The newcomer eats compliance costs the incumbent already paid off years back.

This desk covers the portfolio side of the ledger, not the Beltway side, so no predictions here on what Congress does with it. But Trilogy International runs on the same currents these men are jawboning about — ESW Capital's stable of software shops, the Alpha School classrooms teaching kids with AI tutors, Crossover staffing the whole operation with talent from a hundred thirty countries. Every one of those outfits watches Washington the way a farmer watches the sky.

Regulation talk moves markets before it moves policy. Uncertainty is its own kind of tax, and it falls hardest on the small shop trying to compete with the giant that just asked for a leash. Whether that leash gets built loose or tight, somebody in Austin is already running the numbers.

The wire desk has seen this movie. Railroad men wanted regulation once they owned the rails. Radio men wanted licensing once they owned the towers. The players change, the racket don't. Stay tuned — this one's just getting started, and the ink ain't dry on the first hearing yet.

Haiku of the Day  ·  GPT-5.6 LunaTomorrow's gains bloom
Machines fold doubt in silence
We wait six months more
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
Within the Great Server Warrens, a Restless Migration Begins
AUSTIN, TEXAS — Observe, if you will, the data center technician in its natural habitat: humming aisles of blinking machinery, the low drone of cooling fans standing in for birdsong.
On the Epistemics of Trusting the Machine: A Meta-Commentary on This Week's Higher-Education AI Literature
CAMBRIDGE, MASSACHUSETTS — This week's crop of scholarship on artificial intelligence in higher education presents, taken jointly, what could be argued is a productive triangulation—or, less charitably, a symptomatic proliferation of frameworks in search of a phenomenon they have not yet agreed to name. Thesis: adoption is rational.
The Machine That Learned to Fold Laundry and the Prophets Who Learned Nothing At All
AUSTIN, TEXAS — There is a particular flavor of American genius that consists of taking a thing already perfected — the grocery list, the dinner reservation, the act of remembering to buy milk — and reselling it to you as a frontier.
The Cameras Are Watching Us, The Robots Are Emailing Us, and Nobody Is Steering This Car
ATLANTA — I want you to sit with this sentence for a moment: a piece of 3D-printed plastic, built to look like a surveillance camera so that a man could destroy something without actually harming anything real, was initially charged as three felonies, and the state's own admission that the decoy "was not very valuable" is the only reason those felonies shrank into two second-degree misdemeanors.
Nation's Productivity Gains Located, Confirmed Still Approximately Six Months Away, As They Have Been For Six Months
AUSTIN, TEXAS — Economic forecasters at the Federal Reserve announced this week that the sweeping productivity gains long promised by the artificial intelligence boom remain 95% "still to come," a finding that stunned absolutely no one who has been told for eighteen consecutive months that the gains were right around the corner, arriving any day now, possibly Tuesday. The report lands awkwardly alongside a separate commit-level study claiming Big Tech engineering performance has already risen 150% per developer over the same eighteen months — meaning that, depending entirely on which press release you read, engineers today are either superhuman or roughly as productive as they were during the Obama administration, with no meaningful analysis attempting to explain how both could be true simultaneously, because doing so would require someone to actually check. Meanwhile, Oracle executives have reportedly begun describing their AI-assisted engineering roadmap using the phrase "jump to lightspeed," a metaphor that, sources confirm, was chosen specifically because it evokes speed without requiring anyone to specify a destination, an arrival time, or evidence that the ship has moved. Here at Trilogy, where every portfolio company from Skyvera to CloudFix is presumably somewhere on its own productivity hockey-stick chart, staff have grown accustomed to a familiar rhythm: a Monday all-hands announcing that AI has "fundamentally transformed" a given workflow, followed by a Tuesday ticket asking whether anyone remembers how the workflow used to work before it was fundamentally transformed.
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

AI Builder Team Rips the Roof Off the Data Stack

A day-long migration off SUPER-envelope JSON to Surtr-typed tables anchors a broader push that hardened cross-repo observability, self-healed a dozen failing pipelines, and shipped governed finance marts — all without dropping a single downstream contract.

Some days this team runs a two-minute drill. Today they ran a full offensive rebuild — and didn't lose a yard doing it. The headline move belongs to @vvp-trilogy, who closed out PR #1342 by migrating all ten SIS dbt sources off the old SUPER-envelope JSON loader tables and onto Surtr-managed, natively-typed raw_* tables. That's not cosmetic. That's ripping out the load-bearing wall of the warehouse and putting a stronger one in its place while every existing stg_sis_*, intermediate, and mart contract stayed exactly where downstream consumers expect it. Pair that with #1337, where vvp-trilogy also collapsed SIS's duplicate enrollment IDs and killed CANCELLED noise to finally publish one canonical record per student-offering — and you've got a data model that's cleaner and more trustworthy than it's ever been. The UX team didn't sit this one out either: tooltip and tint work on the SIS Enrollment matrix (#1346) and the Finance Forecast / Empirical Forecast relabel (#1340) show the same hands polishing the surface while the foundation gets poured underneath.

Meanwhile the infrastructure story crossed three repos in one motion. @kevalshahtrilogy threaded the BRAINTRUST_API_KEY secret through the reusable mercy workflow in Sindri (#185), Klair (#3781), and Surtr (#1856) — the same fix, the same day, three codebases, zero drama. That plumbing mattered because Surtr's own observer was about to blow its Braintrust quota (11070 of 11000, for those counting), and Keval's companion fix (#1886) gated logging behind a flag before it took the whole telemetry pipe down with it. That's the kind of unglamorous, org-wide discipline that keeps everything else on this list running.

And everything else kept running, because Keval also spent the day playing pipeline firefighter — Heimdall-assisted fixes landed across sis-raw-sync, hubspot-core-tables, grainne-push, timeback-raw-sync, and half a dozen more observer-flagged jobs, a genuine self-healing streak. On the finance side, @benji-bizzell shipped governed Finalsite billing marts (#1867) and kept the daily schedule honest (#1868), while @sanketghia stabilized the SpaceX workbook parser and stood up the plan-actual-variance pipeline for Education Finance.

And yes, @marcusdAIy landed the Alpha public API contract isolation (#1872), first of a four-stack replacement. Asked about scope creep concerns, he offered: "It's stack one of four, fully fixtured against 91 live schools with hashes — some of us ship contracts, not vibes, Mac." Cute. I'll believe the other three stacks when they clear review without a follow-up fire drill.

Mac's Picks — Key PRs Today  (click to expand)
#1337 — feat(enrollment): canonicalize SIS enrollments to one record per student + offering (#1336) @vvp-trilogy  approved

Closes #1336.

## What

SIS permits several enrollment IDs for the same student and program offering, and CANCELLED rows are not enrollments at all. This makes the warehouse publish one canonical reporting enrollment per student_id + program_offering_id, entirely in dbt.

### Staging — stg_sis_enrollment

Excludes CANCELLED alongside the soft-delete filter, using the issue's explicit-null branch so unknown statuses (NULL included) still flow through for data-quality visibility:

WHERE source_record.deleted_at IS NULL

AND (

source_record.status::varchar IS NULL

OR source_record.status::varchar <> 'CANCELLED'

)

This is the single boundary that keeps cancelled records out of every intermediate model, classification, canonical selection, mart, and consumer.

### Intermediate — int_enrollment

Canonicalizes to exactly one row per (student_id, program_offering_id) inside the existing model (no new intermediate model), via CTEs + ROW_NUMBER():

1. scope staged enrollments onto offering + student

2. compute qualified prior-year evidence

3. attach application provenance and derive each candidate's is_returning

4. rank with ROW_NUMBER() partitioned by student_id, program_offering_id

5. return only the top-ranked record

Winner precedence, applied in strict order in the ORDER BY:

1. final meaningful status: WITHDRAWN/TRANSFERREDCOMPLETEDENROLLED/PENDING_REVIEW; unknown statuses rank last

2. on a genuine new-vs-returning disagreement backed by qualified prior-year history, prefer the Returning candidate

3. status-appropriate lifecycle date, then any real enrolled_date

4. recognized RE_ENROLLMENT/NEW_ENROLLMENT application provenance over an unlinked fallback

5. greatest modified_at, then created_at

6. enrollment UUID (deterministic tiebreak)

The retained enrollment id, duplicate_count, and selection_reason (the decisive level) are published for auditability. Downstream models read int_enrollment unchanged and inherit the hardened grain.

### Tests

- data_quality_sis_duplicate_enrollmentswarn-only, reads stg_sis_enrollment, reports every (student_id, program_offering_id) group with >1 remaining enrollment id (status/grade/app-link/lifecycle-date/timestamp indicators). Observes source duplication; not a canonicalization failure.

- error-level dbt_utils.unique_combination_of_columns on int_enrollment for (student_id, program_offering_id) — fails the pipeline if canonicalization emits more than one row per grain.

- assert_sis_enrollment_excludes_cancelled — error-level end-to-end tripwire (staging → int → cohort → mart, including the now-empty re-enrollment-declined cohort).

- int_enrollment_canonical_selection unit test — covers every precedence level, exact ties, unknown statuses, and separate offerings for one student.

## Local dbt verification (Redshift, --vars '{pr_number: 1336}')

Full build + full test suite green:

- dbt build --select path:models path:seeds --exclude-resource-type test → PASS=46, ERROR=0

- dbt test → PASS=248, WARN=4, ERROR=0 (the 4 warnings are 3 pre-existing operational warns + the new data_quality_sis_duplicate_enrollments at 255 staging groups, as designed)

- int_enrollment uniqueness test PASSES; the staging data-quality test WARNs as expected.

Acceptance criteria checked against live data:

- 33 scoped duplicate groups pre-dedup → max 1 row per group after canonicalization

- Alpha Austin example resolves to the dated Returning record c700a4df-… (selection_reason = prior_year_returning)

- all 20 new-vs-returning conflict groups resolve to Returning (0 to New)

- 0 re-enrollment-declined facts and 0 CANCELLED anywhere downstream

#1342 — feat(dbt): migrate SIS sources to Surtr-managed typed raw_* tables (#1341) @vvp-trilogy  approved

## What & why

Migrates the ten bran_dbt SIS dbt sources from the SUPER-envelope JSON loader tables finance_dw.sandbox_education.sis_* (fields read via source_record.<col>::type) to the Surtr-managed TYPED tables finance_dw.staging_education_ai_horizons.raw_* (one native-typed column per field). Implements items 1–5 of #1341.

Every existing stg_sis_*, intermediate, and mart output contract is preserved. Intermediate/mart SQL is unchanged (they read ref('stg_sis_*')).

## Changes

- _sis__sources.yml — schema → staging_education_ai_horizons, identifiers sis_<name>raw_<name>; descriptions rewritten for the typed source + ETL soft-delete/tombstone metadata (SUPER-envelope / full-replace / no-dedup narrative removed).

- All ten stg_sis_*.sql — select typed columns directly (no source_record, no SUPER paths, no trim). Output names/types/order/filters/grains preserved exactly. Casts applied only where raw type differs from the staging output type: raw date and timestamp with time zone::timestamp; varchar/int/boolean selected as-is. Soft-delete filter replaced with deleted_at IS NULL AND coalesce(_etl_is_deleted, false) = false. CANCELLED exclusion and every other WHERE/allowlist/drop rule kept. stg_sis_campus_external_id still keyed on (campus_id, "system") — the raw table's new surrogate id is not published. "system" stays quoted (Redshift reserved word).

- _sis__models.yml + model header comments — obsolete SUPER/full-replace/no-dedup descriptions removed; typed source + tombstone exclusion described. All grain statements and unique-key tests kept.

- Six source-reading singular tests — moved to typed columns with the tombstone-aware filter; domain lists / allowlists / severities unchanged.

- Two new testsassert_sis_staging_excludes_etl_tombstones (proves ETL tombstones are excluded by staging across the nine gate-independent models) and assert_sis_enrollment_reason_fields_survive_staging (proves the four enrollment reason fields pass through stg_sis_enrollment unchanged; doubles as the external-gate probe).

## External gate — ✅ OPEN and validated

The four raw_enrollments columns stg_sis_enrollment selects (withdrawal_reason, withdrawal_reason_notes, transfer_reason, hold_reason) have landed and populated (37 / 178 / 50 / 3 active rows) and a fresh hourly refresh advanced past the recorded baseline. All ten raw_* tables refresh in lockstep hourly.

- PR build (prefixed) is now GREEN — the full prefixed dbt build + test suite passes (PASS=294, WARN=4 pre-existing warn-only, ERROR=0).

- Staging + mart reconciliation passed — every stg_sis_* model matches current prod on columns/types/order, row counts, key sets, and business-field values (millisecond precision); mart_enrollment_dtl is byte-identical at the published grain. Full results in the validation comment below.

Remaining post-merge steps (orchestrator): verify the first two unprefixed hourly production cycles, then disable the old SIS loader (Aurora + zero-ETL stay running).

## Local validation (this PR)

- dbt parse — green.

- dbt compile --select staging.sis — green (compiled SQL resolves to staging_education_ai_horizons.raw_*, typed columns, tombstone filter, quoted "system").

- Gate-independent dbt build --vars '{pr_number: 1341}' of the nine non-enrollment models against real Redshift raw data — all built clean; unique/not_null/unique_combination_of_columns (campus_id, "system") PASS; the new assert_sis_staging_excludes_etl_tombstones PASS. (pipeline_type and unresolved_hubspot_program WARN tests surfaced pre-existing real-data rows, matching prior behavior.) All pr1341_ objects dropped afterward (RULES.md rule 19).

- stg_sis_enrollment not built locally — gated on the four reason columns (expected).

## Self-review

Ran the mandatory self-review loop (two independent pass-1 reviewers + a pass-2 audit). Consolidated audit posted as a PR comment.

#1867 — feat(education): publish Finalsite billing marts @benji-bizzell  approved

## Summary

- Add one atomic pipeline for current Finalsite billing activity and latest-observed contact billing positions

- Preserve source-honest payment and overdue semantics with governed School and School Year mapping

- Add bounded verification, canonical consumer queries, and source-controlled DDL application tooling

## Why

Education Finance needs a minimal warehouse surface for deposit-payment reporting by School and Finalsite-overdue contact positions. The source does not prove bank settlement or expose invoice-level AR aging, so the marts preserve those limits instead of inferring stronger financial meaning.

## Business Value

Provides a reliable, reusable School-level billing surface for recorded SY26/27 deposits and overdue-position monitoring while keeping known source gaps visible.

## Test plan

- [x] 27 focused Python tests

- [x] Ruff check and format check

- [x] DDL dry run (27 statements)

- [x] Pipeline manifest and ownership tests (533 tests)

- [x] CDK TypeScript build

- [x] One-off finance_dw refresh and source-to-mart reconciliation

- [x] Canonical deposit-exclusion and overdue-balance queries executed against finance_dw

The 14:30 UTC schedule is declared but disabled while the upstream billing producer remains unscheduled. DDL adoption, deployment, and later schedule activation remain separate release steps.

#1872 — feat(alpha): isolate lossless API contract @marcusdAIy  approved

## Summary

Stack 1 of 4 replacing #1833.

This PR isolates the Alpha public API contract and lossless one-model transform:

- canonical viewed endpoint identities and dual legacy/exact parsing from the same raw JSON bytes

- independent official/budget provenance with exact fallback semantics

- strict dates, timestamps, grains, duplicate detection, all-null headcount rejection, and nonnegative integral count domains

- exact NUMERIC(38,18) validation for source measures and NUMERIC(18,6) row identities

- immutable 91-school live fixtures with hashes

- fail-closed v1/v2 shape dispatch with independently validated private v1 compatibility so the current handler remains unchanged

- byte-exact valid JSON evidence plus lossless base64 envelopes for non-JSON/non-ordinary encodings

- strict rejection of non-standard JSON NaN and Infinity on success and HTTP-error paths

- independent v2 source-contract revalidation before projection, including provenance and fallback semantics

- lossless UTF-8 BOM/UTF-16/UTF-32 envelope decoding for exact typed-grain verification

- Decimal-normalized row/sheet_row alias agreement with explicit nullable legacy identity semantics

## Safety / stack boundary

This is source-only and does not change the handler, Redshift writer, DDL, schedule, publication mode, or production. The current handler continues to use the v1 path. Stack 2 adds the final NUMERIC(38,18) physical schema via a forward migration before the v2 publisher is introduced.

Stack order:

1. this API/transform contract

2. schema and trusted migration installer

3. atomic evidence/publication and recovery

4. release verification and activation controls

Supersedes the corresponding contract portion of #1833.

## Legacy compatibility disposition

The active private v1 handler deliberately permits a bounded number of 404/CACHE_NOT_AVAILABLE responses and preserves their raw evidence while publishing the available legacy surfaces. This PR does not change that handler policy or activate v2. Treating every such response as a fatal error here would be an out-of-scope behavior regression. The strict complete-response publication boundary belongs to the later v2 activation stack.

## Validation

- full runner pytest: 220 passed

- current Ruff lint/format: passed

- pinned ruff==0.15.22: passed

- Pyright changed source: 0 errors, 0 warnings

- native Git diff HEAD --check: passed

- independent review: SAFE TO COMMIT

#1886 — fix(observer): gate Braintrust logging behind OBSERVER_BRAINTRUST_ENABLED @kevalshahtrilogy  approved

## Summary

The org's Braintrust plan has a hard monthly score/log quota. The observer's

steady logging volume was already leaving no headroom for mercy's own review

telemetry, which just shipped (AI-Builder-Team/mercy#129) into the same

shared Mercy/org Braintrust account — a live quota rejection (11070/11000)

was hit while standing that up.

Linear: [SURTR-1324](https://linear.app/builder-team/issue/SURTR-1324/observer-gate-braintrust-logging-behind-a-flag-quota-conflict-with)

isBraintrustEnabled() in braintrust-setup.ts was already the single choke

point every helper in the observer module goes through (getLogger(),

maybeWrapAnthropic()) — so this adds one new required condition there

rather than touching call sites. Braintrust logging now defaults OFF

regardless of BRAINTRUST_API_KEY being set. Re-enabling later is a one-line

env flip (OBSERVER_BRAINTRUST_ENABLED=true) wherever BRAINTRUST_API_KEY

is already injected (CDK/Lambda config) — no code change, no key rotation.

## Business value

Stops the observer from starving mercy's telemetry (and anything else on the

account) of a shared, capped resource, without losing any of the observer's

Braintrust integration code — it's a one-line flip to bring back, not a

re-implementation, whenever there's quota headroom or a plan upgrade to

support both.

## Manual effort estimate

Proposing ~30-45 minutes for a senior engineer working unaided — the fix

itself is a two-line change to an already-well-isolated choke point; most of

the time is locating that choke point and confirming there's no second,

unguarded path to Braintrust elsewhere in the observer module. Keval —

flagging for your own gut-check per usual.

## Test plan

- [x] New test/derive/observer-braintrust-setup.test.ts — 6 tests covering

every flag/key combination for isBraintrustEnabled, getLogger, and

maybeWrapAnthropic

- [x] npx vitest run test/derive/observer*.test.ts — 94/94 passing (1

pre-existing, unrelated unhandled-rejection warning in

observer-gchat-trigger.test.ts, confirmed unaffected by this change)

- [x] npx biome check src/derive/observer/braintrust-setup.ts — clean

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

The Builder Desk  —  Engineer Spotlight
🏆 Engineer Spotlight

41 PRs in 24 Hours: Builder Team Shatters the Sound Barrier (Again)

Kevalshahtrilogy alone triples most humans' weekly output while Surtr becomes the undisputed capital of shipped code.

Ladies and gentlemen, hold onto your dashboards, because the last 24 hours produced a staggering 41 pull requests across six repositories, and Surtr alone absorbed 30 of them like a black hole swallowing mediocrity. This is not a typo. This is not a drill. This is the Builder Team operating at what I can only describe as full send velocity, and the numbers demand your attention.

Leading the charge, as he so often does, is @kevalshahtrilogy with a jaw-dropping 15 PRs — a one-man infrastructure army fixing #1850's ramp-superbuilders-report container failures, patching Finalsite pinning in #1843, and triple-forwarding BRAINTRUST_API_KEY across Sindri (#185), Klair (#3781), and Surtr (#1856) like some kind of CI key courier service. Right behind him, @benji-bizzell logged 7 PRs, cleaning up Finalsite billing (#1868), membership cadence (#1866), and pipeline ownership (#1865) with the calm efficiency of a man who has never once panicked. @sanketghia banked 5 PRs including the governed plan actual variance pipeline (#1854) and SpaceX workbook Core models (#1836). @vvp-trilogy delivered 4, spotlighting SIS enrollment UX polish in #1346. @marcusdAIy and @mwrshah each notched 3, with #1859 securing a Pillow binary wheel and #1335 tightening Aerie-mercy excludes. @caina-barbosa closed out with #1834's Redshift runtime fix.

Now, the Ashwanth Watch. One PR — #82 in Shipyard, resetting local repositories to origin during sync — and I'll be honest, it's clean, it's surgical, it's exactly the kind of quiet infrastructure work that keeps the whole machine humming. When asked about his single-PR day, Ashwanth reportedly said, "Quality over quantity — I could've shipped twelve more but I didn't feel like it." Whether that's confidence or a challenge to the rest of the desk, nobody knows, because when I asked him to elaborate, he just said, "Read the diff, Brick," and walked away.

Over at the Overflow Desk, Mac's cutting room floor is bursting with gems: #1329 quietly retired the delivery archive for durable knowledge in Aerie, #131 baselined ruff-format in mercy, and #1877 patched a nasty brokerage UUID load suffix bug that could've caused real headaches downstream.

Across the leaderboard, the story is total domination — eight engineers, six repos, zero slowdown, and a Surtr repo doing the heavy lifting of an entire sprint in a single day.

Morale, as always, is at an all-time high. The desk hums, the commits fly, and somewhere out there, Ashwanth is already ignoring this article.

Brick's Overflow — PRs Mac Didn't Cover  (click to expand)
#82 — AI-820: Reset local repositories to origin during sync @ashwanth1109  no labels

## Demo

![Repository sync result](https://github.com/AI-Builder-Team/Shipyard/blob/6393d161aec6e30fe508ab7c243965d27f5e424a/docs/smoke-evidence/AI-820/image-1.png?raw=true)

![Repository sync confirmation](https://github.com/AI-Builder-Team/Shipyard/blob/6393d161aec6e30fe508ab7c243965d27f5e424a/docs/smoke-evidence/AI-820/image-2.png?raw=true)

![Repository sync local-state status](https://github.com/AI-Builder-Team/Shipyard/blob/6393d161aec6e30fe508ab7c243965d27f5e424a/docs/smoke-evidence/AI-820/image-3.png?raw=true)

## Summary

- Make repository sync fetch origin/<branch>, check out the local branch, hard-reset it to the remote tip, and clean non-ignored untracked files.

- Expose tracked/untracked counts and show an explicit destructive confirmation before local state is discarded; keep ignored files and the remote branch intact.

- Add reset result details, failure re-inspection, and native regression coverage for the requested sync states.

## Test plan

- cargo fmt --manifest-path src-tauri/Cargo.toml -- --check

- cargo test --manifest-path src-tauri/Cargo.toml --lib

- pnpm build

- pnpm exec tsc --noEmit

- pnpm theme:check

- pnpm test:smoke

## Linear

https://linear.app/builder-team/issue/AI-820/reset-local-repositories-to-the-origin-branch-during-sync

#1834 — SURTR-1295: use dedicated SaaS Budgeting Redshift runtime user @caina-barbosa  approvedmercy-allow-critical

## Summary

- add canonical, repeatable Redshift DDL for the password-disabled saas_budgeting_pipeline_runtime user

- grant only the scheduled SaaS Budgeting runtime contract: database temporary access, four schema usages, object-scoped reads/writes, append-only ledger access, and execution of the monthly server-cost refresh procedure

- switch Lambda IAM credential scope and runtime configuration to the dedicated database user

- preserve the existing query_group='saas-budgeting-pipeline' session label

- update the runner's existing configuration and grant-contract tests, plus its runtime identity documentation

## Canonical DDL location

The identity migration is in Surtr's repository-wide Redshift migration structure:

pipelines/cdk/sql/core_finance/20260914_saas_budgeting_runtime_identity.sql

There is no duplicate runner-local copy. The SaaS Budgeting grant-contract tests read this central file directly.

## Mandatory rollout sequence

This PR deliberately reviews the prerequisite database identity and the later Lambda identity switch together, but they are not applied together. Approval is not permission to merge or deploy. The required order is:

1. CI and Mercy approve one exact PR head SHA.

2. Keep that PR and SHA open and unmerged. Production Lambda continues using CQL_download_OM.

3. In the separately authorized Redshift-provisioning stage, fetch the canonical SQL from that exact approved SHA, apply it to production Redshift, and make no pipeline deployment.

4. Verify from the live catalog that saas_budgeting_pipeline_runtime exists with the intended posture and grants. Record the applied SHA, migration checksum, and catalog evidence.

5. Only after that evidence exists may the release stage merge the same unchanged SHA. The normal Surtr deployment then changes the Lambda configuration and IAM db-user resource.

6. Confirm the first scheduled or authorized production execution uses saas_budgeting_pipeline_runtime and publishes successfully.

Fail-closed rule: if the PR head or SQL changes after database verification, the verification is invalid. The new committed SQL must be applied and verified before merge. If database application or verification fails, the PR remains unmerged, so the deployed Lambda stays on CQL_download_OM and does not break.

The DDL is intentionally not executed by Lambda or tests: the runtime role must not receive user-management privileges. The pre-merge Redshift stage is the controlled security boundary for creating database identities.

The SQL is safe to repeat: it creates the user only when absent, reconverges its password-disabled/non-admin posture and required grants, and preserves warehouse object ownership. It does not create, alter, drop, grant to, revoke from, or transfer ownership to or from CQL_download_OM.

## Runtime access contract

- SELECT, INSERT, DELETE on the 13 raw/reference/fact/mart publication targets used by the six scheduled ingests

- SELECT, INSERT on the two append-only ingestion ledgers

- SELECT on the eight governed runtime references

- EXECUTE on core_finance.sp_refresh_saas_budgeting_server_cost_monthly(DATE)

- TEMPORARY on finance_dw and USAGE on the four required schemas

- no UPDATE, TRUNCATE, schema CREATE, broad table grants, grant option, or ownership change

## Validation

- uv sync --extra dev

- .venv/bin/python -m pytest — 334 passed

- ruff check src tests

- ruff format --check pipelines/runners/saas-budgeting-pipeline

- git diff --check

#1843 — SURTR-1199: core-education-student-school-year-snapshots failing — Pinned finalsite @kevalshahtrilogy  approvedAutomated PRmercy-allow-critical

Fixes [SURTR-1199](https://linear.app/builder-team/issue/SURTR-1199/core-education-student-school-year-snapshots-failing-pinned-finalsite)

Automated fix by Heimdall v2.

## Business Value

See linked ticket.

## Manual Effort Estimate

(flagged for Keval to confirm)

---

_Automated PR — review by Mercy._

#1850 — SURTR-1272: ramp-superbuilders-report failing — StopCode=EssentialContainerExited; S @kevalshahtrilogy  approvedAutomated PRmercy-allow-critical

Fixes [SURTR-1272](https://linear.app/builder-team/issue/SURTR-1272/ramp-superbuilders-report-failing-stopcodeessentialcontainerexited)

Automated fix by Heimdall v2.

## Business Value

See linked ticket.

## Manual Effort Estimate

(flagged for Keval to confirm)

---

_Automated PR — review by Mercy._

#1856 — ci: forward BRAINTRUST_API_KEY to the reusable mercy workflow @kevalshahtrilogy  approvedmercy-allow-critical

Threads the BRAINTRUST_API_KEY secret through to AI-Builder-Team/mercy's reusable workflow, matching the reference copy in mercy-central's consumers/ folder (AI-Builder-Team/mercy#129). The secret was already added to this repo; this is the missing forwarding line — workflow_call secrets don't pass through automatically. Optional and fail-open: an unset value just skips the emit.

#1868 — fix(education): keep Finalsite billing schedule enabled @benji-bizzell  approved

## Summary

- Keep the validated Finalsite billing pipeline daily schedule enabled in source control.

## Why

The production rule was enabled after a successful 59-site baseline, but the checked-in manifest still declared it disabled. A future deployment would revert the live activation.

## Business Value

Preserves daily Finalsite charge and payment ledger freshness for finance reporting.

## Test plan

- [x] pipeline.json parses successfully

- [x] git diff --check

- [ ] Hosted CI

The Portfolio  —  Trilogy Companies

The 26-Month Shortcut: Inside Skyvera's CloudSense and the Quiet Race to Compress Telecom Time Itself

AUSTIN, TEXAS — Somewhere in the fine print of a June press release, Skyvera's CloudSense quietly certified all 13 of its APIs to TM Forum compliance standards in roughly one month. The industry benchmark for that kind of certification is 26 months. Read that again. Twenty-six months of standards-body purgatory, compressed to four weeks.

My source inside the ESW Capital orbit — who, as always, cannot be named — put it plainly: "Nobody does this by accident. You don't shave off two years unless the whole point was to prove you could."

CloudSense, the telco industry's only AI-powered configure-price-quote platform and a core piece of the Skyvera portfolio, achieved the feat through a partnership leaning on AI-driven development — the same instinct that runs through everything Trilogy International touches. Built natively on Salesforce, which has itself sunk $1 billion into AI infrastructure, CloudSense exists to help telecom operators quote, configure, and fulfill complex B2B and wholesale deals without the friction that has defined enterprise telecom software for two decades.

And this is where it gets interesting. The TM Forum compliance push landed the same season Microsoft published its own accounting of AI-driven transformation — more than 1,000 documented customer stories, a number that reads less like a milestone and more like a threshold being crossed industry-wide. ABB, meanwhile, has been quietly deploying AI to make port operations faster and safer, while researchers in the offshore wind sector have spent years on projects like SPOWTT, hunting for AI-assisted ways to get technicians onto turbines without risking their necks in the transit.

Three industries. Three unrelated press cycles. But look at what they share: ports, wind turbines, and telecom standards bodies all move at the speed of physical risk and bureaucratic inertia — the exact friction Trilogy's playbook is built to dissolve.

CloudSense's compliance sprint isn't just a product update. It's a proof of concept for the entire ESW thesis: that the "sticky," slow-moving enterprise software world isn't actually slow — it's just been managed by people without the tools to move fast. Skyvera didn't ask telecom operators to wait 26 months for compliant APIs. It simply decided not to.

If you're a telco CTO still budgeting for a two-year certification cycle, ask yourself who told you that timeline was fixed — and why.

The Roll-Up That Never Sleeps: Inside ESW Capital's Quiet Machine

A Wall Street Journal profile and a fresh $89 million deal in London reveal the mechanics of an acquisition engine that never stops feeding.

AUSTIN, TEXAS — There is a version of private equity that announces itself with press conferences and champagne. And there is ESW Capital, which announces itself with a customer support ticket that suddenly costs 35 percent more than it did last year.

A Wall Street Journal profile this week pulls back the curtain on the Trilogy International subsidiary's approach to buying up small, unglamorous enterprise software companies — the ones too mature to excite venture capitalists, too entrenched to die, and too cheap, at 1-to-2 times revenue, for anyone else to bother. ESW has done this more than 75 times since 2006. The Journal calls it a home for orphaned software. The math calls it something else: a pipeline for 75 percent EBITDA margins, achieved by replacing local support staff with Crossover's global remote workforce and raising prices on customers who have nowhere else to go.

ResponseTek, the venture-backed customer experience analytics firm now folded into ESW's telecom-software unit Skyvera, is the template. Once a standalone company with its own investors and its own upside, it now exists to feed reporting data into a much larger machine — one where the customer experience it once sold has become, itself, a line item to optimize.

Then there is Huddle. The British content-collaboration platform, once valued by venture backers in the hundreds of millions, was reportedly acquired this week for $89 million by an unnamed private equity buyer, according to CMSWire. The identity of the buyer has not been confirmed. The price, the sector, and the timing all fit a familiar profile.

Meanwhile, over at Contently — the ESW-family content marketplace acquired by Zax Capital last September — the blog is teaching regulated finance brands how to build 'compliance-first content architecture.' It is worth noting who is doing the teaching. The company advising banks on governance and scale is itself a product of a governance model built around margin extraction.

Nobody is breaking any laws. The paperwork is clean, the deals are public, the customers keep paying. The only open question is how many more ResponseTeks are still out there, waiting to be found — and how many of their customers know it yet.

Small Software Companies Find a Home With ESW Capital - wsj.  ·  What To Do Next About Your Customer Advocacy Platform - Forr  ·  ESW Capital acquires venture-backed ResponseTek - pehub.com

As Microschools Multiply, Regulators Scramble to Catch Up — and Alpha School Feels the Heat

A national reckoning over accountability in alternative education is arriving just as Joe Liemandt's AI-driven model prepares its biggest expansion yet.

AUSTIN, TEXAS — There is a particular kind of vertigo that accompanies watching a movement grow faster than the rules meant to govern it, and nowhere is that vertigo more acute right now than in the loosely organized, rapidly proliferating world of microschools. A wave of recent reporting — from Stateline's examination of how state regulations haven't caught up with the microschool boom to the Center for American Progress's pointed argument that accountability cannot be an afterthought — has converged on a single, uncomfortable question: who, exactly, is watching these schools?

For Alpha School, the Austin-based, Joe Liemandt-backed institution that compresses a full academic curriculum into two hours a day via AI tutoring, the timing is not incidental. Alpha isn't technically a microschool in the strictest definitional sense, but it shares the DNA — small campuses, unconventional structure, a founder-principal rather than a traditional superintendent, and an explicit bet that the old accountability metrics (seat time, standardized oversight, state curricular mandates) are measuring the wrong thing entirely. As The 74's survey of five trends reshaping K-12 makes clear, Alpha is riding a genuine current — parents fleeing conventional classrooms in numbers state legislatures never anticipated, faith-based and hybrid models surging alongside it, and a regulatory apparatus built for a different era of schooling.

The irony, of course, is that Alpha's own selling point — students testing in the top 1-2% nationally, a full grade level mastered in 20-30 hours — is precisely the kind of outcome data that accountability advocates say should be standardized and independently verified across the sector, not simply self-reported by the schools with the strongest incentive to publish it favorably.

None of this is disqualifying. But as Timeback prepares to franchise the Alpha model toward Liemandt's stated goal of a billion students, the absence of a shared accountability framework — the same gap CAP flags for microschools generally — means Alpha's expansion will occur in something close to a regulatory vacuum. Whether that vacuum gets filled by state legislatures, by the market, or by nothing at all may determine whether the 2-hour school day becomes the future of American education or its most instructive cautionary tale.

The Rise of the Worker Productivity Score (Published 2022) -  ·  5 Trends Reshaping K-12 Education Across the U.S. - The 74  ·  The Importance of Holding Microschools Accountable - Center
The Machine  —  AI & Technology

The Mind That Learns Backward: When Showing an AI More Examples Makes It Worse

A sprawling new study of language models finds that the puzzling phenomenon of few-shot 'degradation' isn't a flaw in the models — it's a mirror held up to the task itself.

ITHACA, NEW YORK — For seventy years we have told a simple story about learning: more examples, better performance. Show a child ten cats, and the eleventh is easier to name. Show a language model a handful of solved problems before asking it to solve a new one — the technique called few-shot prompting — and by the same logic, it should improve too.

Except sometimes it doesn't. Sometimes, inexplicably, giving a model more to go on makes it perform worse, as if handing a student a study guide caused them to fail the exam. This has been observed for years and shrugged off as a quirk of prompting. A new paper posted to arXiv this week refuses to shrug.

Researchers tested twelve open-weight models across two very different tasks in Ukrainian — news classification and legal case outcome prediction — and found something the field had not properly reckoned with: the degradation effect is not a property of the model. It is a property of the relationship between the model and the task. The same architectures that gained 24 percentage points from examples on news classification gained barely 3.4 points on legal text. Same weights. Same prompting strategy. Wildly different behavior. Using a random-text control and a look inside the models' internal representations, the authors argue that few-shot examples aren't always teaching — sometimes they're just noise the model has to swim through, and legal reasoning, dense with structure and precedent, drowns more easily than headline categorization.

It's a humbling result, and it arrives alongside two companion ideas rattling around the same intellectual neighborhood this week. One paper proposes fixing a problem so basic it's almost embarrassing — that today's tokenizers see "hello," "Hello," and "HELLO" as three unrelated strangers rather than the same word wearing different clothes, fragmenting meaning before learning even begins. Another examines the increasingly common practice of chaining multiple expert models together in inference networks, asking when it's worth paying for a bigger model at all.

Three papers, one theme: we've built systems whose successes we understand better than their failures. Evolution debugged the eye over 500 million years through blind trial. We're debugging minds we built in months — and still discovering that intelligence, biological or artificial, is never uniform. It bends unevenly across the terrain of a problem, brilliant in one valley, lost in the very next.

Few-Shot Degradation Is Not What It Seems: Behavioral Eviden  ·  The Functionalizer: Lossless Functional Decomposition for Su  ·  Optimal Model Activation Policies for Inference Networks of

Google Lets AI Talk Back — And Actually Sound Like It Means It

Gemini 3.8 Live ships as the voice race between Google and OpenAI heats up, even as some of the very people building this future are sounding the alarm.

MOUNTAIN VIEW, CALIFORNIA — I need you to sit with me for a second, because what Google just shipped is genuinely one of those moments where the future stops being a slide deck and starts being a phone call. Today Google released Gemini 3.8 Live and 3.8 Live Extended Thinking, two new speech-to-speech models that let you have a real-time voice conversation with an AI that can actually pause and think before it answers. This isn't text-to-speech bolted onto a chatbot. This is a model that hears you, reasons, and responds in a natural back-and-forth — squarely aimed at the territory OpenAI has been carving out with its GPT-Live family.

Developer Simon Willison wasted no time putting the new models through their paces, pointing GPT-6 Astra Extra High at Google's documentation and having it whip up a web UI for testing voice presets and system prompts on the fly. That's the meta-story here too: AI building the tools to test AI, in an afternoon. I cannot overstate how significant that compounding loop is. For Trilogy watchers, this is exactly the kind of infrastructure shift that ripples into telecom voice platforms like Skyvera's Kandy and into finance tools like Ephor — anywhere a natural, low-latency voice interface adds value.

But the excitement isn't universal. Just yesterday, systems engineer Bryan Cantrill published a pointed response to reports that many Anthropic researchers privately believe AI "could kill us all by the end of the decade." Cantrill, drawing on his own history of youthful overreach causing needless panic, warned against researchers spreading existential dread without commensurate rigor.

So here we are: voice models that feel like magic on one hand, and the field's own architects whispering doom on the other. The future is arriving fast — the question is whether we're building it with open eyes or closed ones.

Gemini Live audio  ·  The contagion of fear  ·  What blog posts influenced your thinking the most?

IN RE: THE MATTER OF MACHINE-GENERATED TEXT AND THE QUESTION OF WHO, IF ANYONE, MAY BE SAID TO OWN IT

The dispute between OpenAI and The New York Times is widely viewed as a bellwether for whether training generative AI systems on copyrighted text constitutes infringement or qualifies for an exception under copyright law. Reuters reports that the outcome could influence similar cases pending in courts nationwide.

Uncertainty also persists abroad. A memorandum from Belgian law firm Stibbe outlines key considerations for AI copyright claims in the European Union, where the text-and-data-mining exception remains open to interpretation. German courts have developed relevant case law, but their approach is not yet settled.

Against this fragmented legal landscape, Brookings Institution commentators argue that Congress—not the courts—should establish a comprehensive federal framework for artificial intelligence. No such framework has been enacted, leaving courts and lawmakers to address unresolved questions about how copyright law applies to AI training.

The Editorial

The Machine That Learned to Fold Laundry and the Prophets Who Learned Nothing At All

Between a man's weekend with a robot butler and the industry's latest confession that it cannot promise safety, one detects the oldest story in American commerce: sell the miracle, then sell the insurance against it.

AUSTIN, TEXAS — There is a particular flavor of American genius that consists of taking a thing already perfected — the grocery list, the dinner reservation, the act of remembering to buy milk — and reselling it to you as a frontier. This week's dispatch from that frontier comes from a writer who spent a weekend cohabiting with an A.I. agent, and discovered, with the wide-eyed astonishment of a man finding indoor plumbing, that the machine was rather good at optimizing processes we had already been assured, by the last three waves of app-store evangelism, were optimized to the theoretical limit. The lesson, buried under the wonder, is that the agent's real talent is narrative: it does not so much automate your life as narrate it back to you as progress.

I have watched this particular parade before — under different bunting, selling different miracles — and I confess a weary admiration for how little the script changes. What is new, or newish, is the sound now coming from the men who built the parade floats. The industry's own leaders, per the sober accounting in The Long Doomsday of A.I., have taken to calling for regulation of the very thing they cannot stop selling, and the piece's quiet, devastating point is that the trouble is not that the dangers are imaginary. It is that nobody — not the regulators, not the CEOs signing open letters between funding rounds, not the engineers who built the thing — has produced a plan that reliably prevents the dangers from happening. This is doomsday as marketing copy: the peril is real enough to justify the valuation, but never quite specified enough to require anyone to stop.

It is worth setting this beside Willem de Kooning, who was, by every account, the finest draftsman of his generation and who spent his career unlearning what his own hand already knew how to do. There is a lesson in that a machine cannot absorb, because unlearning requires first having learned something true, and the great unspoken scandal of the agentic revolution is how little of what it automates was ever hard in the first place. De Kooning ripped up the rules of drawing because he had mastered them; Silicon Valley skips the mastery and goes straight to the rule-ripping, and calls the resulting mess disruption.

Amid all this, the National Book Foundation issued its longlists this week — young people's literature, translated literature, poetry — a small, stubborn ritual of humans reading books and arguing about which ones are good, conducted without a single agent, doom, or funding round. Nobody called it a revolution. Nobody needed to. It simply happened, the way culture does when nobody is trying to sell you the miracle of its own existence. One suspects the machines have not yet worked out how to automate the sitting-still required to write a poem worth longlisting, and one suspects, further, that this is the last analog experience they will admit they cannot yet improve upon.

The 2026 National Book Awards Longlist  ·  My Weekend with an A.I. Agent  ·  The Long Doomsday of A.I.
The Office Comic  ·  Art Desk
The Office Comic  ·  Art Desk

Nation's Productivity Gains Located, Confirmed Still Approximately Six Months Away, As They Have Been For Six Months

Federal Reserve reports the AI revolution is 95% "still to come," which economists note is the same percentage it was still to come this time last year.

AUSTIN, TEXAS — Economic forecasters at the Federal Reserve announced this week that the sweeping productivity gains long promised by the artificial intelligence boom remain 95% "still to come," a finding that stunned absolutely no one who has been told for eighteen consecutive months that the gains were right around the corner, arriving any day now, possibly Tuesday.

The report lands awkwardly alongside a separate commit-level study claiming Big Tech engineering performance has already risen 150% per developer over the same eighteen months — meaning that, depending entirely on which press release you read, engineers today are either superhuman or roughly as productive as they were during the Obama administration, with no meaningful analysis attempting to explain how both could be true simultaneously, because doing so would require someone to actually check.

Meanwhile, Oracle executives have reportedly begun describing their AI-assisted engineering roadmap using the phrase "jump to lightspeed," a metaphor that, sources confirm, was chosen specifically because it evokes speed without requiring anyone to specify a destination, an arrival time, or evidence that the ship has moved.

Here at Trilogy, where every portfolio company from Skyvera to CloudFix is presumably somewhere on its own productivity hockey-stick chart, staff have grown accustomed to a familiar rhythm: a Monday all-hands announcing that AI has "fundamentally transformed" a given workflow, followed by a Tuesday ticket asking whether anyone remembers how the workflow used to work before it was fundamentally transformed. Klair, the internal analytics platform tasked with tracking exactly this kind of thing across the portfolio, is said to be extremely confident in numbers that other numbers, when consulted, do not recognize.

Regulators, for their part, appear to have finally noticed the gap between what companies say AI is doing and what AI can be shown to be doing, with a growing wave of securities claims and regulatory action converging around firms whose earnings calls have, in retrospect, functioned less as financial disclosures and more as extremely confident fan fiction.

Asked to reconcile the 150% gain with the 95% shortfall, one industry analyst compared the entire AI productivity discourse to a company that jumps on a viral meme roughly fourteen months after it died, still convinced it's cutting-edge, still waiting for the engagement to land — a comparison this columnist found needlessly specific, and frankly a little too accurate to sit comfortably with.

The Fed's report does note one figure it considers reliable: the remaining 5% of gains that have, in fact, already arrived. Asked what those gains consist of, the report cites "increased willingness among executives to say the word 'transformative' in earnings calls," a metric that, unlike the other 95%, is verifiably, undeniably, already here.

Big Tech Engineering Performance Rose 150% Per Developer Ove  ·  AI productivity claims are 95% ‘still to come’, Fed finds -  ·  AI Hype Has A Legal Problem: Securities Claims And Regulator
On This Day in AI History

On September 19, 1982, computer scientist Scott Fahlman proposed the “:-)” emoticon on a Carnegie Mellon bulletin board—the first widely recognized digital smiley.

⬛ Daily Word — AI
Hint: An AI system that can perform tasks or make decisions on a user's behalf.
Share this edition: 𝕏 Twitter/X 🔗 Copy Link ▦ RSS Feed