← The AI Hype Audit — all 645 verdicts

REAL

Holds up: double-counted revenue from a bad SQL join is a real and common text-to-SQL failure, and NVIDIA's open-source Auto Ontology really does give agents certified business terms, safe joins and an 'I can't answer that' path

Facebook · Nov 2026

The claimAn AI-written query joined orders to invoices and back to orders, so every sale counted twice and investors saw $4M instead of $2M. NVIDIA's Auto-Ontology fixes the guessing: it reads your database and builds a map of business terms, reviewed formulas and safe joins; the agent looks terms up before writing SQL, the SQL is checked against the database, and it says so when data cannot answer. Free, open source (Apache-2.0), bring any OpenAI-compatible model. Comment AITOOLS for setup steps.

The SQL lesson is right. Joining through a one-to-many table and back again multiplies rows, and summing after that inflates totals; it is one of the most common ways an AI or a human gets revenue wrong. The tool is real and described fairly. NVIDIA/auto-ontology on GitHub is Apache-2.0, launched around November 1, 2026, and its README describes business terms mapped to columns, reviewed and certified SQL attributes (formulas), join paths taken from the ontology's foreign keys 'with no model involved', static checks plus a model judge on generated SQL, and a rule that the agent reports when a question cannot be answered from the data. It exposes nine read-only MCP tools, supports ten databases including Postgres, Snowflake, Databricks and SQLite, and uses NVIDIA NIM by default but accepts any OpenAI-compatible endpoint. Caveats: it is new, NVIDIA's 90 percent time saving is self-reported, a reviewer notes the eval repo ships a harness without published results, the project is not accepting outside contributions, and 'certified' formulas still need a person to certify them. The skit's own punchline proves that limit: no ontology catches test orders unless you define them out. 'Top of GitHub trending' was not verified. What is being sold: a comment-for-DM hook ('Comment AITOOLS'), watermarked aitools.shikharjaiswal.in.

What holds up

  • Read the NVIDIA/auto-ontology README: Apache-2.0, business terms, certified SQL attributes, joins from semantic foreign keys, SQL check and repair, refuses unanswerable questions.
  • README confirms NIM by default plus any OpenAI-compatible endpoint, nine read-only MCP tools, ten database engines.
  • Independent coverage (MarkTechPost, Nov 1, 2026; reviews) dates the launch and flags self-reported speed claims and no published eval results.
  • Contact sheet shows the double join (orders to invoices to orders) and the test-account row at 41 percent.

What doesn’t

  • Comment-for-DM hook for a public GitHub repo.
  • New project; accuracy benefits are not independently benchmarked.
  • Human review is still required to certify terms and exclude test data.
  • 'Top of GitHub trending' not verified.

The catch

Real tool, real failure mode. It cuts guessing, but it only knows what 'revenue' means after a person writes it down and certifies it.

How to actually do it

  • Before trusting any AI revenue number, compare its total against your payment processor or accounting report for the same period.
  • Try NVIDIA/auto-ontology on a read-only copy of your database, point it at a model you already use, and certify the top five terms (revenue, customer, order) yourself.
  • Flag and exclude test accounts and internal orders at the data level so every tool, AI or not, skips them.

A skit that teaches a genuine SQL trap and points to a genuine, free NVIDIA project. The fine print is that the tool is weeks old and depends on human-certified definitions.

Confidence
Medium
Posted by
an animated AI-dev skit page (about 6,000 plays); reel sent to Buddy Mon Nov 9, 2026, 6:58 AM

See the original claim →

We test hype for free. We build the real thing for a living.

Thirty minutes, no pitch — and you'll leave with something useful either way.

Book a call with Todd or start with the free Business Checkup →

The Verdict Weekly

Three verdicts every Friday. Free forever, unsubscribe anytime, no spam — that would be ironic.

© Schreier Group · schreiergroup.com · See a wild AI claim? Drop it here and we'll test it.