5555watch

After clicking an entry, below, "What Hacker News says", we see comments with mentions and surrounding text. It's now squashed, maybe you can add a function to expand on click, to see the comment with newline separation if such exist. Cause very often these comments presents a list of their favorite book, and an explanation of why. When these are squashed, it's not clear which explanation is preceding or following the entry.

Also, clicking the link "On "...."" it doesn't bring to the exact comment, only that thread or topic, so you need to search for it. It could bring directly to that comment.

I've recently been trying to write a paper with an application in a field that I'm not an expert at. I gave the LLM agents to analyze the methods, the data, the outputs, plus some explanations from me what I want to say and show, and told it to write a section of the paper. It did. Then I asked like 3-4 more different frontiers to do the same.

They all do.. something[1]. They write visually great looking papers. Even good on a quick read.

But once I try porting their proposed Section ideas into the paper, I notice that for each sentence I either need to do a complete rewrite, or a deep dive into it (plus the literature), to confirm or delete that idea/sentence.

If it were my field, it would be much simpler to just write everything directly, rather than attack every sentence of LLMs. As it's not my field, it's a decent way to understand the nuances and learn, but I get increasingly disillusioned by AI writing skills.

[1]: Old and new Gemini models are too categorical, everything is a breakthrough and a marvel; Sol/Astra gets bogged down with too many details, and "it's X, not Y" garbage. For some reason, it's very defending against potential referee questions which don't add anything to the main idea; Opus 5.5 or Fable are basically the same. Maybe it's their way of anti-hallucination..

What are we looking at here?

They nerfed it heavily through the subscription, together with the UI and Pricing changes (introduced 100$ plan). Maybe it's back now, but at the time it was observable and reproducible... The API variant seems to be untouched.

What harness are you using with Argon?

Argon is gas, and Carbon is a solid? But I don't see an obvious progression afterwards

I loved Gemini 3.1 Pro and it was very sad when they nerfed it at around May this year. I think it did pretty good as a chat, and quite decent on Antigravity (tried very little). Been sitting on 5.6 / Astra and Codex since. Occasionally do bring up the 3.1 Pro through the API, because I love the writing style, but keep noticing that I'm much more productive with the frontier, despite the love to Gemini.

Tried 4 Argon, and it's very much different from 3.1, felt very much the same as any other Frontier, just without the intelligence.

Discord is very hard to understand for me. But the young people love it for some reason. Had to use it for some projects or hackatons, and it's still a mystery why it's so popular. No one found older links that were deeper in the chats, had to constantly tag people to see anything, random chats pushed the important information away. It was overall terrible for everyone yet everyone still loves it.

And weirdly, I then understood that there's no real global platform for a quick chat that everyone knows and CAN actively use. In universities we use Teams, but beyond that people are never active there...

But I also don't understand Twitter or Mastodon. It feels like there's so much better ways to both post stuff and reach people.

In my minimal attempts to use rsync, I always find examples where they always use a ton of flags alongside the locations. I'm not gonna learn those flags if cp can do it intuitively and with minimal extra commands. Maybe it's just me.

On a somewhat related note, I really hate that in scp -r and -R mean entirely different things.

Google has a small stake in Anthropic

Maybe. But developers are also private people who will also play with their private subs and projects. In my opinion, their private experience might influence some corporate level decisions.

Being grandfathered by OAI and happy is not the same as having both, and noticing "hmm maybe Claude is much better for my case, Ill suggest that to our manager"

It's not that terrible then if that's the case. 5.6 Sol was superb in my eyes, and while I've tried a million things, there hasn't been anything that 6 Astra improved or did better than 5.6 Sol, while costing a ton of time and resources. Including research topics, where it should have excelled. So even if 6.1 Sol is no worse than 5.6 but cheaper and faster, the gutted Pro200 might still make sense.

My favorite comparison was that we're basically buying a GTX 1070 Ti every month, except instead of a GPU we get some extra productivity.

The 200$ plan was appealing because you got 4x usage for 2x the price.

Now, as it's linear, it makes much more sense to downgrade to 100$ OAI and pick up a 100$ Claude sub. (without doing the numbers) the usage should remain the same, total paid the same, but having access to best of both worlds. It should be a win for the user, and a loss for OAI.

With this in mind, it sounds like a fumble by OAI.

If Siri and Apple Intelligence is disabled in the settings, does it still do the download?

GPTZero says: "We are highly confident this text was AI generated (100%)"

Am I the only one that's tired of "my random shower thought turned into full article with AI" articles?

Would be interesting to see Speedometer 3.1 tests as well. I find they translate really well to user experience, and previously the M5 Base was beating the equivalent Ryzen/Intel there as well.

Not that many years ago (maybe 2012-2014), we were taught working with R using only the base-R. Its' native script editor was a Windows Notepad or something equivalent. We used this for years.. Professors as well. For some reason we didn't know about RStudio or any other fancy IDE.

Do you have a good textbook to reference to where the theory starts from PAC-Bayes?

Another useful link [0], page 91. Effectively, the more complex the solution, the heavier is the upper bound on true risk. It doesn't mean a simpler model is necessarily better. But the complexity brings its own larger support for mistakes to live in. The _likelihood_ of being _more correct_ is probably related to larger sample required to learn a more complex model.

[0]: https://www.cs.huji.ac.il/~shais/UnderstandingMachineLearnin...

Recalling their recent misuse report [0], I kind of agree with them. And performance/intelligence wise I think we are already at quite a nice local optima.

[0]: https://www.anthropic.com/threat-intelligence-report-septemb...

This reminded me of anecdotes of people discussing with friends about buying a random specific item, and then suddenly seeing it advertised everywhere before even googling about it.

Next step, discussing your Navier Stokes solutions with friends might require leaving your phone in another room.

At least in my experience, the issue with ArXiv is that the expectation is that the draft should be already in a good enough state. And polishing plus writing the meat around the main result can take a lot of time

Does it make sense to start privately and then open the repo after publication? Will the dates be retained? Also, isn't commit history easy to spoof?

Is there an alternative link without certificate issues?

I also think that it's quite a bad PR for them, is it really worth the Millenium prize? Is it not enough that top mathematicians are already actively using these tools? In the long term this would lead to potentially profitable collaborations with universities? Why throw it away so early? Unless they really believe they're gonna solve all math problems now and reputation doesn't matter.

The ultimate drive for some researches is the pursuit of knowledge. If I'm stuck at some block which prevents me from continuing in some direction that I want, of course I would like some help. I believe we already have nonzero collaborative proofs on math.SE, I can't recall good examples, but I have definitely seen citations to mathSE before.

So for me it sounds quite natural to also share this with AI especially under the privacy assumption. Also there's the assumption of scale -- maybe your problem is not large enough for anyone to care to scoop; and just for blind retraining, how do they know that the proof is even correct to include it into training? I have definitely received a ton of incorrect proofs before. So the SNR of such private chats is also not clear. I'm imagining millions of masters/phd students also trying to solve various random things with various capabilities, but how much real signal is there?

Or, if you're using, you need to push a pre-print relatively soon after AI help.

Maybe a dumb observation, but if a chain of people were working on the problem for a long time, it's not difficult to imagine that someone accidentally prompted a model with their personal or some other account without the privacy set correctly.

Then again, maybe this is my internal cope, hoping that they're not secretly training on private chats.