Hacker News Reader

Comment by minimaxir | Hacker News Reader

Over the past few years, I had helped Ed with some difficult nuances around the obscure technical aspects of serving LLMs (e.g. benchmarks and model caching); he has also shouted myself and Simon Willison out positively multiple times. I stopped assisting him because he repeatedly misused said advice to the most cynical interpretation ("how can this be interpreted to make AI boosters sound crazy?) and often made it misleading at best. Nowadays I suspect he views me as one of those crazy AI boosters.

Not challenging him was a mistake in hindsight, and I own it. Going forward I will no longer be helping writers/journalists with a clear anti-AI bias.

View on HN

He made a blog post, literally last month, and it said that AIs have no use outside of coding.

It just shows he's done zero research on the things he talks about all day. Radiologists are using them, ad firms, artists, translators, law firms, auditors... It's hard to think of a white collar firm not using them.

One thing I've noticed is that since LLMs came out, scientific papers with poor English have essentially disappeared. Just one of the many ways that AI is changing the world. People are using AI in all kinds of fields.

Oh c’mon. You must be trolling. Great, now we have have scientific papers with formally correct English, but data and sources of which are fabricated and most of the paragraphs are utter nonsense. And publications are drowning on these. Oh thanks for changing the world!

AI has been great for scientific research. It writes code, it can help solve difficult mathematical problems, you can bounce ideas off of it, you can learn background info on a scientific question quickly, etc.

I haven't noticed any uptick of trash papers in quality journals, or even on the arXiv. I know submissions are way up to all the journals, but peer review appears to still be working to filter out obvious junk.

And yes, papers being written in understandable English is important. Most scientists around the world are not native speakers of English, and now they can all suddenly write in perfect English.

I am not saying LLM-based software is not useful. For instance, you can use it to verify claims made on HN:

  What we find is sobering. Submission volume has risen by 42% since November 2022. At the same time, submission writing quality also began to decline at the end of 2022, with Flesch Reading Ease (a standard measure of writing quality) 1.28 standard deviations (SD) lower in January 2026 relative to January 2021. Submissions that are heavily AI-generated account for nearly all of these trends. The quality of reviews has also dropped sharply since November 2022, driven by an increase in heavily AI-generated reviews. These reviews, in addition to being of worse quality, are also narrower in their emphasis, focusing more on theory and less on data. In short, more research is being submitted, with more AI writing that is lower in quality—not better.[1]

AI researchers themselves are saying that gaming the peer-review system with barrage of AI is extremely frustrating and pointless:

  Peer review is unpaid work that we do (often on nights and weekends) because peer review on our own work is so valuable. Spending hours going through a submission and then realizing that there are hallucinated citations is infuriating as it is a waste of our time! If you haven’t spent enough time with your work to even get the references correct, then a.) why should we spend time reviewing it for you, and b.) what are you hoping to accomplish with the submission in the first place? Learning from reviews requires reflecting on your work, and you need to spend time with your work in order to do this.[2]

There’s also serious concern whether the peer-review system can survive the slop-era [3]. So excuse me if I found your comment on ”better English” and ”great for scientific research” rather amusing. It’s like shooting yourself in the foot and celebrating how much you will save on shoes.

[1] https://pubsonline.informs.org/doi/10.1287/orsc.2026.ed.v37....

[2] https://geospatialml.com/posts/reviewing-ai-slop/

[3] https://arstechnica.com/science/2026/08/peer-review-is-overw...

You seem to think papers being legible is not important. Fine. Who needs to actually understand what the authors wrote?

But AI is extremely powerful for scientific research, and is massively increasing scientific output. It has side-effects that will have to be dealt with, like an increase in junk submitted to the journals, but there are ways to deal with that (the simplest one would be to use AI to screen out obvious junk).

So no, science is not "shooting itself in the foot." AI is a massive boost for science, but like every new powerful tool, it has side effects. The biggest side-effect in the long term may be that AI does all the work for us, but we haven't crossed that bridge yet.

I do think legible writing is important. You know how you can achieve that? By learning how to write. If you want to be a software developer you need to be able to write code, if you want to be a runner you have to run, if you want to be a researcher you must be able to write clearly and concisely in English. There’s multiple tools to help in that none of which are LLM’s. That is one of the dumber uses for LLM’s, complete waste of computing.

In terms of your claims, you notice how I just backed up my arguments with sources from actual scientists, who say that this is a serious problem which those ai-based ai-slop-spotters won’t fix? Whereas your claims are based with… Nothing. You just really want to believe.

Before AI, there were tons of very bright researchers around the world who couldn't write English well. It was a major problem. It isn't anymore. Writing well is actually very difficult, even if English is your native language.

> There’s multiple tools to help in that none of which are LLM’s.

The first good tool I ever saw was an LLM. The tools before then just weren't very good.

> That is one of the dumber uses for LLM’s, complete waste of computing.

Allowing people who speak different languages to communicate is a very good use of compute, in my opinion.

> notice how I just backed up my arguments with sources from actual scientists

What do you think I am?

The good news is that, while the number of fabricated papers have accelerated, so too is our ability to detect fabrications. Also, fabricated papers were a massive problem well before AI was useful in research. I think it's fair to characterise AI as our saviour here instead of the villain.

The broader issue pertains to the [Replication Crisis.](https://en.wikipedia.org/wiki/Replication_crisis) For decades, journals relied mostly on the honour system. Many people think peer review requires validating the data. It does not. It's basically a cursory check that the claimed methods are valid. The data is usually not validated. In fact, the data is usually not even produced for the journal or peer reviewers at all. This has become such a large issue that research began to sample studies to see if they could replicate the results given the claimed methods. A shocking number of papers could not be replicated. Scientists are loathe to claim this proves fraud, but when one has carefully written 50 pages of dense research methodology to arrive at statistically significant findings, it's clear that they would know if their experiments were replicable.

There is also another major issue with academic research as it exists: findings are (almost) only ever published when they confirm the hypothesis. Researchers begin from an often very biased perspective about the way they believe the world works (sociology is one of the worst for this), and work backwards from their conclusion. They might conduct dozens or hundreds of experiments, all while fail and are never published. Then they land on an experiment which can be massaged to reflect the message, and that is published. The fact is that all those "failures" are equally insightful research, and it should all be published for a variety of reasons.

There are also major issues re how research is gated and funded, and how many hobby researchers are never given the opportunity to publish at all - even if they could afford to pay the journal bribe.

Long story short, this entire industry was on the verge of collapse prior to AI. I cannot imagine it could get any worse, and I'm hopefully that AI will enable us to easily screen fraudulent and non-replicable research in the future. I expect 98% of papers to be rejected.

>I will no longer be helping writers/journalists with a clear anti-AI bias.

Take out the 'anti-AI' qualification, and you've got a decent general principle.

This reads slightly weird and more like a pledge of alignment than a statement from the heart.

But apart from that, I guess that's always the learning? Journalists (or people labeling themselves as such) often have their own story they want to tell, and usually do so by building it out of little blocks of reality stacked together to form the desired picture.

I would predict a similarly frustrating experience being equally probable even without the "clear anti-AI bias" attribute set.

That's an unfair characterization of tech journalists in general. When BuzzFeed News was around, I worked closely with them on technical aspects and projects to ensure that everything was correct and accurate, and there were very receptive to it; it was not a "directionally correct" thing, the journalists wanted to avoid any unintentional. That was also one of the reasons I was open-minded to helping Ed even though we ideologically disagree: the truth is what's important.

BuzzFeed News incidentally was how I first came across Ed on Twitter from his tech PR work about 9 years ago, back when he was writing guest editorials on Gizmodo about Person of Interest (which is very very funny in hindsight). It's from the heart that I'm a bit bummed out that things turned out this way.

Not to dunk on your friends and all, but wasn't Buzzfeed the exact outlet at the forefront of dismantling the credibility of the word "journalism"?

Like.. they surely had valid pieces, but I remember them also being the ones smashing the system and trust to pieces by being very social-media-engagement-native (for the lack of a better term).

The LLM as this beautiful straight man default human simulator tells me that BuzzFeed News was some kind of investigative journalism daughter company of Buzzfeed the destroyer of worlds, and with that these would for sure be two unconnected entities, not to be judged like that.

And, for all legal intents and purposes, I am of course sure it is completely right. I could almost bet that this exact trick is why there even was a legit journalism daughter company with the same name in the first place.

So that one could well-actually all the bad words away, by pointing at it and either saying "Hey it's a different thing!!!111" or "Hey but we're also doing good work in that branch!!!!1111". Always depending on which is more opportune. Association or disassociation.

Anyway. I do believe you that your friends were trying to do the right thing.

I just have doubts as to why the place they were in even existed in the first place.

Certainly to enable them to do great journalistic work, but maybe not because the corporate superstructure actually cared by heart about great journalistic work.

At least I do not see it around no more, which might indicate that it outlived its corporate usefulness as a moral shield.

They really were separate. BuzzFeed News won a Pulitzer in 2021 (source: Wikipedia). It's unfortunate that they shared the same name.

It's wild to watch you lecture someone who just made you aware of the topic that you are lecturing him about.

> back when he was writing guest editorials on Gizmodo about Person of Interest (which is very very funny in hindsight)

Sometimes life has a sense of humor

> That's an unfair characterization of tech journalists in general.

Not at all. Journalists in general have a very poor reputation and Gaming/Tech journalists have an even poorer reputation. I have regularly found various stories to be incorrect and/or so poorly reported that I could spend all of 5 minutes doing a web search and find the original post / video or article and at best it is often misleading.

Reply on news.ycombinator.com