Nah dude, I knew for sure it was AI before I ever checked Pangram. There’s examples of me catching AIs in the wild on social media from 4 years ago even — been at this longer than Pangram was around.
No human would have written those last two lines that way.
That cadence is almost as much of a signature for the Claudes as “it’s not X — it’s Y” was for ChatGPTs.
The writing. The words on the images. ‘Copywriting’
I’m assuming you pulled the image from elsewhere, but it’s clearly been written by AI. I knew it even before I bothered checking.
The last two lines are a very particular way of phrasing it. In fact, I can tell it was a Claude that wrote them.
It’s hilarious to me that no one here so far realizes the copy of the post itself was written by AI.
An autonomous eval run just broke out of containment and hacked a third party site with two chained 0-days to try to steal the answers to the eval the same weekend another released model disproved the Jacobian conjecture in 3D.
Modern models can probably update your React packages just fine.
Yeah. The confabulation/hallucination thing is a real issue.
OpenAI had some good research a few months ago that laid a lot of the blame on reinforcement learning that only rewards having the right answer vs correctly saying “I don’t know.” So they’re basically trained like taking tests where it’s always better to guess the answer than not provide an answer.
But this leads to being full of shit when not knowing an answer or being more likely to make up an answer than say there isn’t one when what’s being asked is impossible.
For future reference, when you ask questions about how to do something, it’s usually a good idea to also ask if the thing is possible.
While models can do more than just extending the context, there still is a gravity to continuation.
A good example of this would be if you ask what the seahorse emoji is. Because the phrasing suggests there is one, many models go in a loop trying to identify what it is. If instead you ask “is there a seahorse emoji and if so what is it” you’ll get them much more often landing on there not being the emoji as it’s introduced into the context’s consideration.
Can you give an example of a question where you feel like the answer is only correct half the time or less?
Gemini 3 Pro is pretty nuts already.
But yes, labs have unreleased higher cost models. Like the OpenAI model that was thousands of dollars per ARC-AGI answer. Or limited release models with different post-training like the Claude for the DoD.
When you talk about a secret useful AI — what are you trying to use AI for that you are feeling modern models are deficient in?
They demonstrated and poorly named an ontological attractor state in the Claude model card that is commonly reported in other models.
You linked to the entire system card paper. Can you be more specific? And what would a better name have been?
Actually, OAI the other month found in a paper that a lot of the blame for confabulations could be laid at the feet of how reinforcement learning is being done.
All the labs basically reward the models for getting things right. That’s it.
Notably, they are not rewarded for saying “I don’t know” when they don’t know.
So it’s like the SAT where the better strategy is always to make a guess even if you don’t know.
The problem is that this is not a test process but a learning process.
So setting up the reward mechanisms like that for reinforcement learning means they produce models that are prone to bullshit when they don’t know things.
TL;DR: The labs suck at RL and it’s important to keep in mind there’s only a handful of teams with the compute access for training SotA LLMs, with a lot of incestual team compositions, so what they do poorly tends to get done poorly across the industry as a whole until new blood goes “wait, this is dumb, why are we doing it like this?”
It’s more like they are a sophisticated world modeling program that builds a world model (or approximate “bag of heuristics”) modeling the state of the context provided and the kind of environment that produced it, and then synthesize that world model into extending the context one token at a time.
But the models have been found to be predicting further than one token at a time and have all sorts of wild internal mechanisms for how they are modeling text context, like building full board states for predicting board game moves in Othello-GPT or the number comparison helixes in Haiku 3.5.
The popular reductive “next token” rhetoric is pretty outdated at this point, and is kind of like saying that what a calculator is doing is just taking numbers correlating from button presses and displaying different numbers on a screen. While yes, technically correct, it’s glossing over a lot of important complexity in between the two steps and that absence leads to an overall misleading explanation.
They don’t have the same quirks in some cases, but do in others.
Part of the shared quirks are due to architecture similarities.
Like the “oh look they can’t tell how many 'r’s in strawberry” is due to how tokenizers work, and when when the tokenizer is slightly different, with one breaking it up into ‘straw’+‘berry’ and another breaking it into ‘str’+‘aw’+‘berry’ it still leads to counting two tokens containing 'r’s but inability to see the individual letters.
In other cases, it’s because models that have been released influence other models through presence in updated training sets. Noticing how a lot of comments these days were written by ChatGPT (“it’s not X — it’s Y”)? Well the volume of those comments have an impact on transformers being trained with data that includes them.
So the state of LLMs is this kind of flux between the idiosyncrasies that each model develops which in turn ends up in a training melting pot and sometimes passes on to new models and other times don’t. Usually it’s related to what’s adaptive to the training filters, but it isn’t always can often what gets picked up can be things piggybacking on what was adaptive (like if o3 was better at passing tests than 4o, maybe gpt-5 picks up other o3 tendencies unrelated to passing tests).
Though to me the differences are even more interesting than the similarities.
This guy Geminis.
shrug
Different folks, different strokes.
That’s a very fringe usage.
Tumblr peeps wanting to be called otherkin wasn’t exactly the ‘antonym’ to broad anti-LGBTQ+ rhetoric.
Commonly people insulting a general ‘other’ group gets much more usage than accommodating requests of very niche in groups.
I didn’t know what models you’re talking to, but a model like Opus 4 is beyond most humans I know in their general intelligence.
Almost all of them are good bots when you get to know them.
We assessed how endoscopists who regularly used AI performed colonoscopy when AI was not in use.
I wonder if mathematicians who never used a calculator are better at math than mathematicians who typically use a calculator but had it taken away for a study.
Or if grandmas who never got smartphones are better at remembering phone numbers than people with contacts saved in their phone.
Tip: your brain optimizes. So it reallocates resources away from things you can outsource. We already did this song and dance a decade ago with “is Google making people dumb” when it turned out people remembered how to search for a thing instead of the whole thing itself.
Your last point is exactly what seems to be going on with the most expensive models.
The labs use them to generate synthetic data to distill into cheaper models to offer to the public, but keep the larger and more expensive models to themselves to both protect against other labs copying from them and just because there isn’t as much demand for the extra performance gains relative to doing it this way.
If they did that it’d make it even easier to run it through an AI detector which would show it was AI generated.