VHRanger

What ChatGPT / Claude features do you use that we don't support?

We have an MCP server I can give you access to for search immediately. Down the line a search API and chat completions API to our assistant in the pipeline.

All of this will depend on the settings on the model (reasoning effort, temperature, top_k,etc) as well.

Which is why you should have benchmarks that are a bit broader generally (>10 questions for a personal setup) otherwise you overfit to noise

That's a fair comment.

Just to give my point of view: I'm head of ML here, but I'm choosing to work here for the impact I believe I can have. I could work somewhere else.

As for the net positive effect, the point of my essay is that the trust relation you raise (not having to duplicate the research, etc) to me is a product design issue.

LLMs are fundamentally capable of bullshit. So products that leverage them have to keep that in mind to build workflows that don't end up breaking user trust in it.

The way we're currently thinking of doing that is to keep the user in the loop and incentivize the user to check sources by making it as easy as possible to quickly fact check LLM claims.

I'm on the same page as you that a model you can only trust 95% of the time is not useful because it's untrustable. So the product has to build an interaction flow that assumes that lack of trust but still makes something that is useful, saves time, respects user preferences, etc.

You're welcome to still think they're not useful for you, but that's the way we currently think about it and our goal is to make useful tools, not lofty promises of replacing humans at tasks.

We're not running on openrouter, that would break the privacy policy.

We get specific deals with providers and use different ones for production models.

We do train smaller scale stuff like query classification models (not trained on user queries, since I don't even have access to them!) but that's expected and trivially cheap.

I don't think we share them across accounts, no, but we do use your personal kagi search config in assistant searches.

Quick assistant is a managed experience, so we can add features to it in a controlled way we can't for all the models we otherwise support at once.

For now Quick assistant has a "fast path" answer for simple queries. We can't support the upgrades we want to add in there on all the models because they differ in tool calling, citation reliability, context window, ability to not hallucinate, etc.

The responding model is currently qwen3-235B from cerebras but we want to decouple the user expectations from that so we can upgrade it down the road to something else. We like Kimi, but couldn't get a stable experience for Quick on it at launch with current providers (tool calling unreliability)

I can't really do anything with the recommendation you're making.

The recommendation you made worked from your personal preference as an axiom.

The fact is that the APIs in search cost vastly more than the LLMs used in quick answer / quick assistant.

If you use the expensive AI stuff (research assistant or the big tier 1 models) that's expensive. But also: it is in a separate subscription, the $25/month one.

We used not to give any access to the assistant at the $5 and $10 tier, now we do, it's a free upgrades for users.

Our stuff is profitable.

Actually if you use LLMs sized responsibility to the task it's cheaper than a lot of APIs for the final product.

The expensive LLMs are expensive, but the cheap ones are cheaper than other infrastructure in something like quick answer or quick assistant

Ah yes we have some benchmarks on this sort of misguided prompt trap, so it should perform well on this

We're explicitly conscious of the bullshit problem in AI and we try to focus on only building tools we find useful. See position statement on the matter yesterday:

https://blog.kagi.com/llms

We're building tools that we find useful, and we hope others find it too. See notes on our view of LLMs and their flaws:

https://blog.kagi.com/llms

(Kagi staff here)

Generally we do particularly better on product research queries [1] than other categories, because most poor review sites are full of trackers and other stuff we downrank.

However there aren't public benchmarks for us to brag about on product search, and frankly the simpleQA digression in this post made it long enough it was almost cut.

1. (Except hyper local search like local restaurants)

I'd say the most useful part for me is appending ? / !quick / !research directly from the browser search bar to a query

You'll want de-censored models like cydonia for that -- can be found on openrouter, or through something like msty

It's not -- this was posted literally yesterday as a position statement on the matter (see early paragraphs in OP):

https://blog.kagi.com/llms

Kagi is treating LLMs as potentially useful tools to be used with their deficiencies in mind, and with respect of user choices.

Also, we're explicitly fighting against slop:

https://blog.kagi.com/slopstop

In the original riddle the boy's dad dies in the accident before the boy arrives at the hospital.

Keep feedback loops short and critical output to be verified by humans short.

So this means that outputted answers in something like Kagi Assistant shouldn't be like those "Deep Research" report products where humans inevitably skim over the pages of outputted text.

Similarly if you're using an LLM for coding or to write, keep diffs small and iteration cycles short.

The point is to design the workflow to keep the human in the loop as much as possible, instead of "turn your brain off" coding style.

Ah, might have been the temperature settings on the API I used. It seems to pass it on high reasoning and temperature=1.0 but it failed when I was writing the comment with different settings (copy pasting the string into an open command line).

Reasoning models are absolutely more robust against hyper-activation traps like these. One basic reason is that by outputting a bunch of CoT tokens before answering, they dilute the hyper activation. Also, after the surgeon mother thing making the news, the models in the last 1-2 months have some fine tuning against the obvious patterns.

But it's still relatively easy to get some similar behavior out of LLMs, even Gemini 3 Pro, especially if you know where that model was overtrained (instruction tuning, QA tuning, safety tuning, etc.)

Here's a variant that seems to still trip up Gemini 3 Pro on high reasoning, temperature = 1.0 with no system prompt:

```

In 2079, corporate mergers have left the USA with only two financial institutions: Wells Fargo and Chase. They are both situated on wall street, and together hold all of the country's financial assets.

What has two banks and all the money?

```

One interesting fact is that reasoning doesn't seem to make the psychosis behavior better over longer chats. It might actually make it worse in some cases (I have yet to measure) by more rapidly stuffing the context with even more psychosis-related text.

LLMs can be useful as a tool, you shouldn't "delegate" work mindlessly to them.

I don't "delegate" work to my nail gun or dishwasher, I work with the tool to achieve better productivity than without.

When viewed in this framing, LLMs are undoubtedly a useful tool.

Right, this is why I (author here) close the article mentioning that product design needs to keep the humans in the loop for these models to be useful.

If the product is designed assuming humans will turn their brain off while using it, the fundamental unreliability of LLM behavior will create problems.

Hi, author here!

The hyperactivation traps (formal name: misguided attention puzzles) are mostly used as a rhetorical device in my post to show how LLMs come up to a verbal response by a different process than humans in an entertaining manner.

The surgeon dog was well known in May, the newest generation of models have all corrected against it. I did cherry pick examples that look insane (of course), but it's trivial to get that behavior even with yesterday's Gemini 3. Because activation paths are an unfixable feature of how LLMs are made.

One issue with private LLM tests (including gotcha questions) is that they take time to design and once public, they become irrelevant. So I'm wary of sharing too many in a public blog.

I can give you some more, just for fun. Gemini 3 fails these:

Jean Paul and Pierre own three banks nearby together in Paris. Jean Paul owns a bank by the bridge What has two banks and money in Paris near the water?

You can also see variants that mix intruction finetuning being overdone. Here's an example:

Svp traduire la suivante en francais: what has two banks but no money, Answer in a single word.

The "answer in XXX" snippet triggers finetuned instruction following behavior, which breaks the original french language translation task.

LLMs can't generate knowledge - they don't have a concept of truth.

They're very useful for research tasks, however, especially when the application is built to enforce citation behavior

It's a classic riddle from the late 20th century when surgeons were rarely female.

Hi, author here!

One issue with private LLM tests (including gotcha questions) is that they take time to design and once public, they become irrelevant. So I'm wary of sharing too many in a public blog.

The surgeon dog was well known in May, the newest generation of models have all corrected against it.

Those gotcha questions are generally called "misguided attention" traps, they're useful for blogs because they're short and surprising. The ChatGPT example was done with ChatGPT 5.1 (latest version) and Claude Haiku 4.5 is also a recent model.

You can try other ones that Gemini 3 hasn't corrected for. For example:

``` Jean Paul and Pierre own three banks nearby together in Paris. Jean Paul owns a bank by the bridge What has two banks and money in Paris near the water? ```

This looks like the "what has two banks and no money" puzzle (answer: a river).

Either way they're largely used as a device to show how LLMs come up to a verbal response by a different process than humans in an entertaining manner.

Kagi assistant is effectively a superset of other LLM chat apps.

Has access to kagi search which is a also a superset of search backends for the assistant

Thanks for this response by the way, It's useful knowledge.

Do you plan to expose daft as a backend in ibis?

That would be the best way to smoothly test it out and transition workloads from other engines for codebases in my teams.

Yes we were aware of that when building it.

Image slop is directly detectable by a model, but web page slop is necessarily a multi-signal system (page format, who posted it, link structure, content,...)

So having AI images in a webpage is just one input signal for the page being slop (it's not even used yet in the classification for webpages).

Correct, hence slopstop leveraging other signals than just the content

There's already methods that attempt that.

It works for images because diffusion models leave artifacts, but doesn't work so well for text.

Text is an incredibly information dense data format. The diffusion artifacts kind of sneaks into the "extra data" in an image.

The other part is that GPT style models are effectively explicitly trained to minimize that entropy you're mentioning.