jwzMar 14

This is some boss-level forensic accounting demonstrating that Facebook secretly wrote and is shepherding the various "age verification" bills.

Age Verification Lobbying: Dark Money, Model Legislation & Institutional Capture: How corporate lobbying, think tank infrastructure, competing model legislation, and obscured funding networks are shaping age verification policy across 45 states and Congress...
https://jwz.org/b/yk4c

Show threadvalMar 14
@jwz It's not "boss-level forensic accounting", it was hallucinated by a LLM (cf disclosure at the end of https://tboteproject.com/git/hekate/attestation-findings).
The "analysis" contains nonsense like: https://tboteproject.com/git/hekate/attestation-findings/src/branch/main/output/reports/anomaly_report.md
attestation-findings

Age attestation investigation

Forgejo: Beyond coding. We Forge.
Show threadJuno JoveMar 14

@val @jwz
Just because a model "found" it looking through a host of data doesn't necessarily mean it's a hallucination... :/

But yes, Boss-Level-Investigation, it is not.

Show threadTael 🔜 AC26
@jupiter @val @jwz Using an LLM for this purpose introduces an untold number of embarrassing failure points which could poison or misdirect the entire investigation. Literally any piece of data, big or small, could be entirely fabricated and every single one needs to be traced back to its source to validate it. The OP used an LLM for this task presumably because they didn't feel capable of ingesting that amount of information on their own, therefore they certainly were not capable of exhaustively checking its work. As a result, I'm not interested in reading, let alone supporting or spreading, their conclusion which I cannot trust to be well-researched and accurate.
Mar 15 at 4:16pmWeb
Show threadJuno JoveMar 15

@tael @val @jwz
I don't disagree with this.
I think LLMs are tools, so if they just slapped it all into SPSS instead and found a correlation, and jumped to conclusions, I'd be as wary.

LLMs are by now far better at reading comprehension than humans, especially when it's about finding needles in haystacks.

They do suck when asked to check financial records, which is why I whole-heartedly agree that this is a flawed methodology to jump to conclusions from.

Show threadTael 🔜 AC26Mar 15
@jupiter @val @jwz LLMs are terrible at reading comprehension. They can't actually *comprehend* anything. They are bad tools because they don't deliver predictable results. Whether you receive an accurate result or even one relevant to your prompt is a matter of statistical probability. SPSS is deterministic software; I'd be far more willing to listen to any conclusion formed based on that technology. LLM use is an automatic disqualification for a discussion of this seriousness. Did you read this page linked by @val?
https://tboteproject.com/git/hekate/attestation-findings/src/branch/main/output/reports/anomaly_report.md
It's pure "AI LARP" and amounts to an LLM Gish gallop. This analysis is worthless.
attestation-findings/output/reports/anomaly_report.md at main

attestation-findings

Forgejo: Beyond coding. We Forge.