> A new and notable feature is the automated fact-checking tool, a Gemini-based “veracity engine for real-time analysis” (“Vera” for short), which runs on every tweet.
My apologies, please allow me to clarify. This is not for posts submitted. I would want the ability to send the HN comment contents and the prompt to any public LLM model; enough requests and models and you’ll converge somewhere near Truth. One model biased? Certainly possible. All models biased? Unlikely.
This would be a public HN comment evaluation and verification engine in the browser. One could go so far as to build a website based on the concept, consuming the HN comment firehose and scoring comments as they’re created, an append only transaction log of sorts. Comments are permanent after two hours, so you only have to run each comment through each model once if you wait until that expiration window passes.
“Computah, validate truthiness of all comments in this thread.”
(Bonus points for scoring HN users by how factually accurate they are over time from their comment history)
enough requests and models and you’ll converge somewhere near Truth
I think this kind of centrism is false. Consider two kids who have to divide a cookie between them. One kid demands 100%, the other only 50%. Sounds like your hypothetical model converges on a 75-25 split, which is unfair. Falsehoods are false, and mostly so are partial falsehoods.
How would you design the pipeline and user interface to reflect the temporal nature and potential theta decay of evaluations? Reprocess comments on a cadence?
For example, however you represent a comment, it could then support expansion of a tree to each evaluation snapshot point in time, like expanding a directory listing its files. My thought on this is somewhat similar to Wayback Machine’s interface per URL; the URL remains static over time, with a calendar and indicators for date of snapshot, and drill down into each snapshot.
In this use case, the comment is the URL, and each LLM evaluation of the comment is a snapshot. Append only.
I suspect that would backfire immediately, as it would mean all stories are indirectly (via many people using this) filtered by what Gemini finds plausible.
Purely as an example, when Trump was making the threats against Greenland, I tried asking ChatGPT what the economic implications would be and how to hedge against them, and it had a great deal of difficulty with even the rhetoric being "current events" rather than "a fictional scenario that will never happen".
> A new and notable feature is the automated fact-checking tool, a Gemini-based “veracity engine for real-time analysis” (“Vera” for short), which runs on every tweet.
All my kingdom for this in an HN extension.
What about then the AI itself is trained on manipulated information?
https://www.theguardian.com/world/2026/aug/26/fake-thinktank...
Anything that trains on public, unvetted data is a liability for accuracy
My apologies, please allow me to clarify. This is not for posts submitted. I would want the ability to send the HN comment contents and the prompt to any public LLM model; enough requests and models and you’ll converge somewhere near Truth. One model biased? Certainly possible. All models biased? Unlikely.
This would be a public HN comment evaluation and verification engine in the browser. One could go so far as to build a website based on the concept, consuming the HN comment firehose and scoring comments as they’re created, an append only transaction log of sorts. Comments are permanent after two hours, so you only have to run each comment through each model once if you wait until that expiration window passes.
“Computah, validate truthiness of all comments in this thread.”
(Bonus points for scoring HN users by how factually accurate they are over time from their comment history)
enough requests and models and you’ll converge somewhere near Truth
I think this kind of centrism is false. Consider two kids who have to divide a cookie between them. One kid demands 100%, the other only 50%. Sounds like your hypothetical model converges on a 75-25 split, which is unfair. Falsehoods are false, and mostly so are partial falsehoods.
This is not very curious of you :) we should try and see! The experiment is the journey.
The hard part is that it locks you in to whatever people believe at the moment, which might be proven wrong in the future
How would you design the pipeline and user interface to reflect the temporal nature and potential theta decay of evaluations? Reprocess comments on a cadence?
For example, however you represent a comment, it could then support expansion of a tree to each evaluation snapshot point in time, like expanding a directory listing its files. My thought on this is somewhat similar to Wayback Machine’s interface per URL; the URL remains static over time, with a calendar and indicators for date of snapshot, and drill down into each snapshot.
In this use case, the comment is the URL, and each LLM evaluation of the comment is a snapshot. Append only.
I suspect that would backfire immediately, as it would mean all stories are indirectly (via many people using this) filtered by what Gemini finds plausible.
Purely as an example, when Trump was making the threats against Greenland, I tried asking ChatGPT what the economic implications would be and how to hedge against them, and it had a great deal of difficulty with even the rhetoric being "current events" rather than "a fictional scenario that will never happen".
X as a name fails yet again.