[jw1224]

> “To stave off some obvious comments:

> yoUr'E PRoMPTiNg IT WRoNg!

> Am I though?”

Yes. You’re complaining that Gemini “shits the bed”, despite using 2.5 Flash (not Pro), without search or reasoning.

It’s a fact that some models are smarter than others. This is a task that requires reasoning so the article is hard to take seriously when the author uses a model optimised for speed (not intelligence), and doesn’t even turn reasoning on (nor suggest they’re even aware of it being a feature).

I asked the exact prompt to ChatGPT 5 Thinking and got an excellent answer with cited sources, all of which appears to be accurate.

[delusional]

I just ran the same test on Gemini 2.5 pro (I assume it enables search by default, because it added a bunch of "sources") and got the exact same result as the author. It claims ".bdi" is the ccTLD for Burundi, which is false they have .bi[1]. It claims ".time" and ".article" are TLDs.

I think the authors point stands.

EDIT: I tried it with "Deep Research" too. Here it doesn't invent either TLDs or HTML Element, but the resulting list is incomplete.

[1]: https://en.wikipedia.org/wiki/.bi

[guyomes]

I wonder if it works better if we ask the LLM to produce a script that extract the resulting list, and then we run the script on the two input lists.

There is also the question of the two input lists: it's not clear if it is better to ask the LLM to extract the two input lists directly, or again to ask the LLM to write a script that extract the two input lists from the raw text data.

[1718627440]

> It claims ".time" and ".article" are TLDs.

Maybe they will be in a time frame when the LLM model is still in use.

[softwaredoug]

In my experience reasoning and search come with their own set of tradeoffs. It works great when it works. But the variance can be wider than just an LLM.

Search and reasoning use up more context, leading to context rot, and subtler harder to detect hallucinations. Reasoning doesn’t always focus on evaluating the quality of evidence, just “problem solving” from some root set of axioms found in search.

I’ve had this happen in Claude code for example where it hallucinated a few details about a library based on what badly written forum post.

[dgfitz]

> … all of which appears to be accurate.

Isn’t that the whole goddamn rub? You don’t _know_ if they’re accurate.

[edent]

OP here. I literally opened up Gemini and used the defaults. If the defaults are shit, maybe don't offer them as the default?

Or, if LLMs are so smart, why doesn't it say "Hmmm, would you like to use a different model for this?"

Either way, disappointing.

[magicalhippo]

> Or, if LLMs are so smart, why doesn't it say "Hmmm, would you like to use a different model for this?"

That is indeed an area where LLMs don't shine.

That is, not only are they trained to always respond with an answer, they have no ability to accurately tell how confident they are in that answer. So you can't just filter out low confidence answers.

[mathewsanders]

Something I think would be interesting for model APIs and consumer apps to exposed would be the probability of each individual token generated.

I’m presuming that one class of junk/low quality output is when the model doesn’t have high probability next tokens and works with whatever poor options it has.

Maybe low probability tokens that cross some threshold could have a visual treatment to give feedback the same way word processors give feedback in a spelling or grammatical error.

But maybe I’m making a mistake thinking that token probability is related to the accuracy of output?

[StilesCrisis]

Lots of research has been done here. e.g. https://aclanthology.org/2024.findings-acl.558.pdf

[never_inline]

> Something I think would be interesting for model APIs and consumer apps to exposed would be the probability of each individual token generated.

Isn't that what logprobs is?

[hobofan]

Then criticize the providers on their defaults instead of claiming that they can't solve the problem?

> Or, if LLMs are so smart, why doesn't it say "Hmmm, would you like to use a different model for this?"

That's literally what ChatGPT did for me[0], which is consistent from what they shared at the last keynote (quick-low reasoning answer per default first, with reasoning/search only if explicitly prompted or as a follow-up). It did miss one match tough, as it somehow didn't parse the `<search>` element from the MDN docs.

[0]: https://chatgpt.com/share/68cffb5c-fd14-8005-b175-ab77d1bf58...

[pwnOrbitals]

You are pointing out a maturity issue, not a capability problem. It's clear to everyone that LLM products are immature, but saying they are incapable is misleading

[delusional]

In you mind, is there anything an LLM is _incapable_ of doing?

[maddmann]

“Defaults are shit” — is that really true though?! Just because it shits the bed on some tasks does not mean it is shit. For people integrating llms into any workflow that requires a modicum of precision or determinism, one must always evaluate output closely/have benchmarks. You must treat the llm as an incompetent but overconfident intern, and thus have fast mechanisms for measuring output and giving feedback.