6,409 karma · joined March 28, 2017
The side channel signals (who posted it, where, etc.) are more valuable in tagging than raw text classifier scores.
That's why I said our definition of slop can include all types of genAI: it's about *thoughtless use of a tool* more than the tool being used.
And also that regardless of the method, your model can be used to generate slop.
True, but going after classifying the source (user's commenting patterns) is a better signal than the content itself.
That said, for us (Kagi) it's a touchy area to, say, label reddit comments as slop/bots. There's no doubt we could do it better than reddit (their whole comment history is only 6TB compressed) but I doubt *reddit* would be pleased at that.
And it's a growing issue for product recommendation searches -- see [1] at last section for example on how astroturfed reddit comments on product questions trickle up to search engine results.
> Another interesting trend is false accusations of LLM use as a form of attack.
Fair again, but the question of AI slop is much more about "who is using the tool how" than the content of the output itself.
Also we're looking to stay conservative. False negatives > false positives in this space.
> And the big state actors who are known to use LLMs in misinformation campaigns are battling the US for AI supremacy and so have an incentive to attack the US on AI since it's currently in the lead.
Not wrong, we're especially going after the deluge of low effort slop, and cleaning up the internet for our users.
Highly sophisticated attacks are likely to evade detection.
> Especially if you're relying on volunteers, this seems prone to abuse in the same way, e.g. Reddit mods are.
The human labelling/review aspect is expected to stay small and from trusted users.
The reporting is wide scale, but review is and will remain closed trust based group.
[1] https://housefresh.com/beware-of-the-google-ai-salesman/
The current search engine doesn't go after WordPress plugins we consider correlated to bad pages.
By far the most efficient method in the search engine for spam is downranking by trackers/javascript weight/etc.
Slopstop is going after page formats but we didn't plan to scale that back to rankings for everyone quite yet, only use it as features to detect AI slop. Otherwise the collateral damage on good actors with bad websites would be risky early on.
I'm sorry that's a YouTube problem, not a problem with the original content.
Sadly we don't have plans to address that at the moment -- otherwise all of youtube would be labeled slop
A naive way of scoring how AI laden text is would be to run n-1 layers of a model and compare the text to the probability space of tokens from the model.
It works somewhat to detect obvious text but is not strong enough a method by itself.
Give it ~2 weeks to start seeing real impact on your results
I can tell you: their board, mostly. Few of whom ever used LLMs seriousl. But they react to wall street and that signal was clear in the last few years
Even if your model scored extremely high perplexity on an LLM evaluation we'd likely still tag it as slop because most of our text slop detection is using sidechannel signals to parse out how it was used rather than just using an LLM's statistical properties on the text.
You can break the AI / slop into a 4 corner matrix:
1. Not AI & Not Slop (eg. good!)
2. Not AI & slop (eg. SEO spam -- we already punished that for a long time)
3. AI & not Slop (eg. high effort AI driven content -- example would be youtuber Neuralviz)
4. AI & Slop (eg. most of the AI garbage out there)
#3 is the one that tends to pose issues for people. Our position is that if the content *has a human accountable for it* and *took significant effort to produce* then it's liable to be in #3. For now we're just labelling AI versus not, and we're adapting our strategy to deal with category #3 as we learn more.
Since even classical machine learning uses BERT based embeddings on the backend this problem is likely wider scale than it seems if a search engine isn't proactively filtering it out
> Kagi pays for hordes of reviewers? Is this another case of outsourcing moderation to sweat shops in poor countries?
No, we're simply not paying for review of content at the moment, nor is it planned.
We'll scale human review as needed with long time kagi users in our discord we already trust
> Do the reviewers use state of the art tools to assist in confirming slop
Mostly this, yes.
For images/videos/sound, diffusion and GANs leave visible artifacts. There's a bit of issues with edge cases like high resolution images that have been JPEG compressed to hell, but even with those the framing of AI images tends to be pretty consistent.
> How does this scale?
By doing rollups to the source. Going after domains / youtube channels / etc.
Mixed with automation. We're aiming to have a bias towards false negatives -- eg. it's less harmful to let slop through than to mistakenly label real content.
Hey, Kagi ML lead here.
For images/videos/sound, not at the current moment, diffusion and GANs leave visible artifacts. There's a bit of issues with edge cases like high resolution images that have been JPEG compressed to hell, but even with those the framing of AI images tends to be pretty consistent.
For human slop there's a bunch of detection methods that bypass human checks:
1. Within the category of "slop" the vast mass of it is low effort. The majority of text slop is default-settings chatGPT, which has a particular and recognizable wording and style.
2.Checking the source of the content instead of the content itself is generally a better signal.
For instance, is the author posting inhumanly often all of a sudden? Are they using particular wordpress page setups and plugins that are common with SEO spammers? What about inboud/outbound links to that page -- are they linked to by humans at all? Are they a random, new page doing a bunch of product reviews all of a sudden with amazon affiliate links?
Aggregating a bunch of partial signals like this is much better than just scoring the text itself on the LLM perplexity score, which is obviously not a robust strategy.
And preferably a good nvme (not a cheap one without a DRAM cache) and good ram.
Don't expect that from products with advertising business models
HOWEVER - the person I was replying to was mentioning them being hurt by the behaviors, and that signals a lack of empathy from the individual doing thise actions.
Even apart from behavioral disorders, there's comorbidities -- autism is comorbid with ADHD and ADHD is highly comorbid with a lot of conduct disorders and personality disorders.
So it could be any of that. But my basic point is that for the person I'm replying to all of this complexity simplifies if you instead focus on the effect of the other person's behaviors on you. That's what you have control over.
Focus in the effect if the person's action on you, and set hard boundaries for yourself.
Note that boundaries are in this specific form:
"If you do X in Y situation, I will do Z"
The consequence to the instigating person (the Z part) is important. The person you described likely has a deficit in empathy, so feedback on actions has to happen to modify behavior.
Note that I'm not saying to hit back. Consequences are generally best when withdrawing something of value than adding something of negative value.
For example: "If we're watching a movie and you are physically rowdy, you will have to leave the living room and entertain yourself in your room".
Focua on how things affect you, and what actions you can take in response to boundary breaking behavior. Then be ruthlessly firm and consistent about it.
The most compatible setup I found is proot-distro into alpine, which bypasses a lot of the android blockers, and the bionic-libc incompatibilities by using musl. Comes at a performance cost, however.
I'm curious how that is; in practice it "feels" like my 9950x is much more efficient at "move tons of RAM" tasks like a duckDB workload above a M4.
But then again a 9950x has other advantages going on like AVX512 I guess?
They're still good benchmarks IMO because they represent a "real workload" but to understand why the 9800X3D performs this much better you'd want some metrics on CPU cache misses in the processors tested.
It's often similar to hyperthreading -- on very efficient sofware you actually want to turn SMT off sometimes because it causes too many cache evictions as two threads fight for the same L2 cache space which is efficiently utilized.
So software having a huge speedup from a X3D model with a ton of cache might indicate the sofware has a bad data layout and needs the huge cache because it keeps doing RAM round trips. You'd presumably also see large speedups in this case from faster RAM on the same processor.
My understanding of it is that Apple Silicon's very very long instruction pipeline plays well with how the software stack in MacOS is written and compiled first and foremost.
Similarly that the same applications take less RAM in MacOS than even in Linux often even because at the OS level stuff like garbage collection are better integrated.
AMD promises ROCm will stop being a joke very soon! Maybe this year even!
How absurd is this statement. China jails and disappears people for online statements at a rate several orders of magnitude larger than any western country.
It's borderline ridiculous to even make a comparison. Some quick examples:
1. https://thediplomat.com/2025/03/chinas-system-of-mass-arbitr...
2. https://www.cnn.com/2024/05/28/china/hong-kong-security-arre...
You can get arrested for "picking quarrels" online:
3. https://www.scmp.com/news/china/politics/article/3146188/pic...
You're right though it seems that they don't just use bing + their own crawler, but who else they use seems to be community speculation.
Kimi-k2 for instance was a positive surprise recently and is being added to your standard plan tonight as well as the brand new GPT-OSS models.
We have something in the works to fight back on that front, do not be worried ;)