"after the company's own analyses found that these classifiers could reduce the reach of these hoaxes by more than 90 percent, Meta is shutting them off"
NEW: Meta has quietly dismantled the system that prevented misinformation from spreading in the United States. Machine-learning classifiers that once identified viral hoaxes and limited their reach have now been switched off, Platformer has learned www.platformer.news/meta-ends-mi...
January 15, 2025 at 2:06 AM
"after the company's own analyses found that these classifiers could reduce the reach of these hoaxes by more than 90 percent, Meta is shutting them off"
"The LM built on classifiers has no content classifier" is very funny
September 25, 2026 at 4:05 AM
"The LM built on classifiers has no content classifier" is very funny
Just heard someone in a San Francisco coffee shop say that their "model identifies identifiers and classifiers"
November 20, 2024 at 9:57 PM
Just heard someone in a San Francisco coffee shop say that their "model identifies identifiers and classifiers"
you just have to not do too well otherwise the palantir classifiers trigger
September 27, 2026 at 4:17 PM
you just have to not do too well otherwise the palantir classifiers trigger
Meta has dismantled the system that prevented misinformation from spreading in the US
• They were machine-learning classifiers that identified viral hoaxes and limited their reach
• Facebook says they could reduce reach of hoaxes by 90%
(via Platformer)
• They were machine-learning classifiers that identified viral hoaxes and limited their reach
• Facebook says they could reduce reach of hoaxes by 90%
(via Platformer)
January 15, 2025 at 2:12 AM
Meta has dismantled the system that prevented misinformation from spreading in the US
• They were machine-learning classifiers that identified viral hoaxes and limited their reach
• Facebook says they could reduce reach of hoaxes by 90%
(via Platformer)
• They were machine-learning classifiers that identified viral hoaxes and limited their reach
• Facebook says they could reduce reach of hoaxes by 90%
(via Platformer)
"LLMs are just n-dimensional classifiers" has a very similar flavor to "everything is just quantum mechanical interactions."
www.science.org/doi/10.1126/...
www.science.org/doi/10.1126/...
More Is Different
www.science.org
September 15, 2026 at 8:34 PM
"LLMs are just n-dimensional classifiers" has a very similar flavor to "everything is just quantum mechanical interactions."
www.science.org/doi/10.1126/...
www.science.org/doi/10.1126/...
Huawei Ascend kernels are the most insidious avenue for xrisk... thank you for saving us, dario.
September 22, 2026 at 5:58 PM
Huawei Ascend kernels are the most insidious avenue for xrisk... thank you for saving us, dario.
i've said before: I am extremely bullish on LLM’s as language classifiers beyond generic control-F
A Stanford group used AI on San Francisco's legal code "in search of every instance in which a city department is mandated to produce a report." It found nearly 500. Now the city attorney thinks 140 of them could be dispensed with. [@nbagley.bsky.social]
Using Artificial Intelligence to Build State Capacity
Start by getting rid of pointless reports. Build on that success.
blog.dividedargument.com
June 10, 2025 at 3:51 AM
i've said before: I am extremely bullish on LLM’s as language classifiers beyond generic control-F
this is simply not true as a matter of technical fact. there are post-hoc classifiers but they are the last line of defense. (there are a lot of valid criticisms of the shallowness of the reinforcement learning which take effect before post-hoc classifiers, but these are not those.)
November 17, 2025 at 7:34 AM
this is simply not true as a matter of technical fact. there are post-hoc classifiers but they are the last line of defense. (there are a lot of valid criticisms of the shallowness of the reinforcement learning which take effect before post-hoc classifiers, but these are not those.)
Autoregressive LLMs aren’t classifiers. They do work on probability distributions over given tokens but classifiers are distinct
Hallucinations *can* arise from classification errors, but that’s more akin, ironically, to what is happening in this post (i.e. confabulating similar sounding concepts)
Hallucinations *can* arise from classification errors, but that’s more akin, ironically, to what is happening in this post (i.e. confabulating similar sounding concepts)
Hallucinations happen because it’s literally just a classification & probability engine with some RNG to taste. Each generation reinforces the previous one’s biases more strongly. It’s artificial, but not intelligent. But it’s great at pattern matching - Things with rules. Code. Language. Music.
September 1, 2026 at 11:59 AM
Autoregressive LLMs aren’t classifiers. They do work on probability distributions over given tokens but classifiers are distinct
Hallucinations *can* arise from classification errors, but that’s more akin, ironically, to what is happening in this post (i.e. confabulating similar sounding concepts)
Hallucinations *can* arise from classification errors, but that’s more akin, ironically, to what is happening in this post (i.e. confabulating similar sounding concepts)
this is kinda the thing: using vision classifiers doesn't seem bad to me at all?
yeah I naïvely thought that nobody would care about using a vision model, since image classifiers seem non controversial to me? I wouldn’t dare touch a diffusion model since I agree those suck in every conceivable way. anyway clearly people don’t see it that way
June 24, 2024 at 5:03 PM
this is kinda the thing: using vision classifiers doesn't seem bad to me at all?
I created a collection with good models for dataset curation
- NSFW classifiers
- PII classifiers
- blazing fast embeddings by model2vec
- quality classifier
- educational value classifier
- domain classifier
Collection: huggingface.co/collections/...
- NSFW classifiers
- PII classifiers
- blazing fast embeddings by model2vec
- quality classifier
- educational value classifier
- domain classifier
Collection: huggingface.co/collections/...
Models for dataset curation - a Dataset-Tools Collection
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
huggingface.co
November 22, 2024 at 12:57 PM
I created a collection with good models for dataset curation
- NSFW classifiers
- PII classifiers
- blazing fast embeddings by model2vec
- quality classifier
- educational value classifier
- domain classifier
Collection: huggingface.co/collections/...
- NSFW classifiers
- PII classifiers
- blazing fast embeddings by model2vec
- quality classifier
- educational value classifier
- domain classifier
Collection: huggingface.co/collections/...
The system Bluesky uses, where machine learning classifiers take a first pass at image classification and humans deal with any appealed labels, is the industry standard for a reason. If you think a post of yours is mislabeled, tap the label to appeal it.
October 7, 2025 at 7:17 PM
The system Bluesky uses, where machine learning classifiers take a first pass at image classification and humans deal with any appealed labels, is the industry standard for a reason. If you think a post of yours is mislabeled, tap the label to appeal it.
I mean, let's be real though, "most" people in Nigeria aren't working as LLM data classifiers. 240 million people live there.
June 15, 2026 at 6:12 PM
I mean, let's be real though, "most" people in Nigeria aren't working as LLM data classifiers. 240 million people live there.
(alex jones voice)
they gave classifiers PRONOUNS!
they gave classifiers PRONOUNS!
there we go!
top choice they/them, second choice it/its, third is she/her
top choice they/them, second choice it/its, third is she/her
September 16, 2026 at 2:01 AM
(alex jones voice)
they gave classifiers PRONOUNS!
they gave classifiers PRONOUNS!
That’s exactly right. They’re machine learning classifiers making a probalistic guess, which is very often wrong. Most large models embed watermarks at the pixel level, so SynthID IS accurate, but these dudes never use it.
September 26, 2026 at 6:25 AM
That’s exactly right. They’re machine learning classifiers making a probalistic guess, which is very often wrong. Most large models embed watermarks at the pixel level, so SynthID IS accurate, but these dudes never use it.
drawing that distinction is both clarifying and extremely useful if you think "domain specific statistical classifiers" are helpful tech and "fascist propaganda machines" are not
September 15, 2025 at 1:55 PM
drawing that distinction is both clarifying and extremely useful if you think "domain specific statistical classifiers" are helpful tech and "fascist propaganda machines" are not
This affects new users more because they haven't followed or liked anything yet. As you do both, you get more specific postes.
When you're new, we don't know much about you, so we use these broad topic classifiers.
When you're new, we don't know much about you, so we use these broad topic classifiers.
August 19, 2025 at 1:50 AM
This affects new users more because they haven't followed or liked anything yet. As you do both, you get more specific postes.
When you're new, we don't know much about you, so we use these broad topic classifiers.
When you're new, we don't know much about you, so we use these broad topic classifiers.
i have a shitpost that a future Skynet AI will probably try to genocide women first because classifiers have been aggressively sculpted to view women's bodies as inherently pornographic and threatening but the more time goes on the more i really wonder if this sort of thing could happen
September 28, 2026 at 1:59 PM
i have a shitpost that a future Skynet AI will probably try to genocide women first because classifiers have been aggressively sculpted to view women's bodies as inherently pornographic and threatening but the more time goes on the more i really wonder if this sort of thing could happen
Safety classifiers can’t tell if you’re asking for a security review to attack or to defend so fable is useless for this. Great.
July 1, 2026 at 9:41 PM
Safety classifiers can’t tell if you’re asking for a security review to attack or to defend so fable is useless for this. Great.
you know maybe the people who can't tell the difference between llms and classifiers have a point
October 10, 2025 at 5:34 AM
you know maybe the people who can't tell the difference between llms and classifiers have a point
Algorithmic classifiers were already bad enough. They didn't need to be trained on the collected works of Dr. Hannibal Lecter.
November 21, 2024 at 12:33 PM
Algorithmic classifiers were already bad enough. They didn't need to be trained on the collected works of Dr. Hannibal Lecter.
Notes on Constitutional Classifiers, Anthropic's new paper (and public red teaming challenge) describing their latest jailbreaking protection
(Includes a note that DeepSeek will happily answer one of the questions about nerve agent research that Claude refuses!)
simonwillison.net/2025/Feb/3/c...
(Includes a note that DeepSeek will happily answer one of the questions about nerve agent research that Claude refuses!)
simonwillison.net/2025/Feb/3/c...
Constitutional Classifiers: Defending against universal jailbreaks
Interesting new research from Anthropic, resulting in the paper [Constitutional Classifiers: Defending against Universal Jailbreaks across Thousands of Hours of Red Teaming](https://arxiv.org/abs/2501...
simonwillison.net
February 3, 2025 at 5:33 PM
Notes on Constitutional Classifiers, Anthropic's new paper (and public red teaming challenge) describing their latest jailbreaking protection
(Includes a note that DeepSeek will happily answer one of the questions about nerve agent research that Claude refuses!)
simonwillison.net/2025/Feb/3/c...
(Includes a note that DeepSeek will happily answer one of the questions about nerve agent research that Claude refuses!)
simonwillison.net/2025/Feb/3/c...