Hacker News Reader

Comment by levocardia | Hacker News Reader

with web search, is knowledge cutoff really relevant anymore? Or is this more of a comment on how long it took them to do post-training?

View on HN

In my experience, web search often tanks the quality of the output.

I don't know if it's because of context clogging or that the model can't tell what's a high quality source from garbage.

I've defaulted to web search off and turn it on via the tools menu as needed.

I feel the same. LLMs using web search ironically seem to have less thoughtful output. Part of the reason for using LLMs is to explore somewhat novel ideas. I think with web search it aligns too strongly to the results rather than the overall request making it a slow search-engine.

That makes sense. They're doing their interpretation on the fly for one thing. For another just because they now have data that is 10 months more recent than their cutoff they don't have any of the intervening information. That's gotta make it tough.

I also find that it gets way more snarky. The internet brings that bad taint.

Web search often tanks the quality of MY output these days too. Context clogging seems a reasonable description of what I experience when I try to use the normal web.

THIS. I do my best work after a long vigorous walk and contemplation, while listening to Bach sipping espresso. (Not exaggerating much.) If I go on HN or slack or ClickUp or work email, context is slammed and I cannot do /clear so fast. Even looking up something quick on the web or an LLM causes a dirtying.

I haven't tried ChatGPT web search, but my experience with Claude web search is very good. It's actually what sold me and made me start using LLMs as part of my day to day. The citations they leave (I assume ChatGPT does the same) are killer for making sure I'm not being BSd on certain points.

That’s interesting. I use the API and there are zero citations with Claude, charGPT and Gemini. Only Kagi assistant gives me some, which is why I prefer it when researching facts.

What software to you use? The native Claude app? What subscription do you have?

Claude directly (web and mobile) with the Pro ($20) subscriptions.

I found it very similar to Kagi Assistant (which I also use).

How often you actually check the citations? They seems to confidentally cite things but then they also say different things what source has.

It depends on the question. I was having a casual chat with my dad and we wondered how Apple's revenue was split amongst products, and it was just to chat about so I didn't check.

On the other hand, I got an overview of Postgres RLS and I checked the majority of those citations since those answers were going to be critical.

Web search is super important for frameworks that are not (sufficiently?) in the training data. o3 often pulls info from Swift forums to find and fix obscure Swift concurrency issues for me.

In my experience none of the frontier models I tried (o3, Opus 4, Gemini 2.5 Pro) was able to solve Swift concurrency issues, with or without web search. At least not sufficiently for Swift 6 language mode. They don’t seem to have a mental model of the whole concept and how things (actors, isolation, Tasks) need to play together.

> They don’t seem to have a mental model of the whole concept and how things (actors, isolation, Tasks) need to play together.

to be fair, does anyone ¯\_(ツ)_/¯

This. It’s a bunch of rules you need to juggle in your head.

Completely opposite experience here (with Claude). Most of my googling is now done through Claude- it can find and digest a d compile information much quicker and better than I'd do myself. Without web search you're basically asking an LLM to pull facts out of its ass- good luck with trusting the results.

Kagi really helps with this. They built a good search engine first, then wired it up to AI stuff.

It still is, not all queries trigger web search, and it takes more tokens and time to do research. ChatGPT will confidently give me outdated information, and unless I know it’s wrong and ask it to research, it wouldn’t know it is wrong. Having a more recent knowledge base can be very useful (for example, knowing who the president is without looking it up, making references to newer node versions instead of old ones)

The problem, perhaps illusory that it's easy to fix, is that the model will choose solutions that are a year old, e.g. thinking database/logger versions from December '24 are new and usable in a greenfield project despite newer quarterly LTS releases superseding them. I try to avoid humanizing these models, but could it be that in training/posttraining one could make it so the timestamp is fed in via the system prompt and actually respected? I've begged models to choose "new" dependencies after $DATE but they all still snap back to 2024

The biggest issue I can think of is code recommendations with out of date versions of packages. Maybe the quality of code has deteriorated in the past year and scraping github is not as useful to them anymore?

Knowledge cutoff isn’t a big deal for current events. Anything truly recent will have to be fed into the context anyway.

Where it does matter is for code generation. It’s error-prone and inefficient to try teaching a model how to use a new framework version via context alone, especially if the model was trained on an older API surface.

I wonder if it would even be helpful because they avoid the increasing AI content

This is what I was thinking. Eventually most new material could be AI produced (including a lot of slop).

Still relevant, as it means that a coding agent is more likely to get things right without searching. That saves time, money, and improves accuracy of results.

It absolutely is, for example, even in coding where new design patterns or language features aren't easy to leverage.

Web search enables targeted info to be "updated" at query time. But it doesn't get used for every query and you're practically limited in how much you can query.

Isn’t this an issue with eg Cloudflare removing a portion of the web? I’m all for it from the perspective of people not having their content repackaged by an LLM, but it means that web search can’t check all sources.

Web pages become prompt, so you still need the model to analyze

I've been having a lot of issues with chatgpt's knowledge of DuckDb being out of date. It doesn't think DuckDb enforces foreign keys, for instance.

Yes, totally. The model will not know about new versions of libraries, features recently deprecated, etc..

Question: do web search results that GPT kick back get "read" and backpropagated into the model?

Right now nothing affects the underlying model weights. They are computed once during pretraining at enormous expense, adjusted incrementally during training, and then left untouched until the next frontier model is built.

Being able to adjust the weights will be the next big leap IMO, maybe the last one. It won't happen in real time but periodically, during intervals which I imagine we'll refer to as "sleep." At that point the model will do everything we do, at least potentially.

Falling back to web search is a crutch, its slower and often bloats context resulting in worse output.

Yes, because it may not know that it needs to do a web search for the most relevant information.

Reply on news.ycombinator.com