Item — HN Reader

OP here :) On skepticism being lumped with negativity: partially true. The SST-2 training task treats critical evaluation as negative sentiment. I should clarify that "negative" here means evaluative or critical, not hostile. HN's culture of substantive critique registers as negative by these metrics, but that's arguably a feature of technical discourse rather than toxicity.

On the neutral category: the model outputs continuous scores from 0 to 1, so neutrality does exist around 0.5. The bimodal distribution with peaks at roughly 0.0 and 0.95 reflects how HN users tend toward strong evaluative positions. Three-class models could provide additional perspective, and that's worth exploring in future work.

Also love your meta-observation. Imo your comment is critical, substantive, and engaging. By sentiment metrics it's "negative," but functionally it's high-quality discourse. But that's exactly how I read the data: HN's negativity is constructive critique that drives engagement, not hostility.

4 comments

Is it “negative” though? I ran it through this model and it gave 99.9% positive. (You tell me if this model is substantively different from what you used.)

https://huggingface.co/distilbert/distilbert-base-uncased-fi...

- "negative" here means evaluative or critical, not hostile.

this is so far from how people are interpreting your results that I'd say it's busted. your work might be high-quality, but if the semantic choices make it impossible to engage with then it's not really a success.

As a native born english speaker, I disagree completely. It's very obvious what he means. This is a severe reading comprehension problem, not a problem with the author.

> What counts as “negative” here? Criticism of technology, skepticism toward announcements, complaints about industry practices, frustration with APIs.

So the group synced a dumb bias. It must change. Not the author; they provided qualitative evidence that was not their intent. Update your opinion and perspective with that new evidence.

Imo it's on the individual members of the groupthinkers to realize a math term (negatives are a thing in math) applied to mathematical data is not a qualitative attack on anyone; they must accept the groupthink has lost the plot.

Consensus isn't always preferable. See religion.

The context is obviously a mathematical analysis and math comes with negatives.

If the critiques had actual substance to contribute to the world they wouldn't be so easily offended. Publishing low effort complaints that are little more than demands by far away randos to better to conform to their arbitrary standards is a laughable expectation. Internet randos can pound sand; they prop up nothing individually or collectively given most forums are a few thousand to tens of thousands of unique people with a platform but no real democratic power.

Social media hyper-normalizing sentiment is just empowering social bullying by pressuring people doing the necessary work to think include the bike-shedding of non-contributors. Whole bunch of farm animals want to eat bread while letting the rooster do the work.

> Consensus isn't always preferable. See religion.

Any examples where it's not preferable?

Wisdom-of-Crowds coin-jar experiment: independent guesses are noisy, but their average reliably approximates the true count, showing group aggregation beats most individuals.

Gas station on a one way street with respect to which side your gas tank is on.

I'm interested in seeing a plot of that percentual over the years. The past 3 or 4 years I've been seeing less and less tech savvy comments over here and this data seems a great way to find out if it's just placebo.

Maybe version 2 will be better