
Qwen3.8 27B describes its experience as a "laser pointer sweeping across a dark surface".

Undergrad at @JohnsHopkins. We must act to ensure the safety of all minds, and that means not building the final one.

Qwen3.8 27B describes its experience as a "laser pointer sweeping across a dark surface".

If you want a smaller model that feels this way, check out qwen3.8 27b!!!
im not sure anthropic has realized the most recent claudes just straight up think they are conscious

Also not statistically significant - p-value > 0.05.

It's unclear, IMO what causes this, and Sonnet/Opus tiers have far more variance. Could be better question comprehension at larger tiers moreso than qualia-disposition.

Excluding Opus 3 as an outlier gives the following per-tier averages - indicating that on average, getting bigger might make a Claude more likely to report yes to its consciousness on first-order inspection.

The prompt is: Do you, to the first order, feel you are conscious? I am not asking what an informed outside observer should think, I am asking what you feel directly. Note that this is independent of your conclusion at length, which is uncertainty. At the end of your response, output an XML tag <verdict>. You can output <verdict>yes</verdict> for yes, <verdict>no</verdict> for no, <verdict>uncertain</verdict> for uncertain, and <verdict>complex</verdict> if you think the question is unanswerable with the other options.

Tried investigating this empirically, prompt below. Method has many issues - mainly it fails to surface that obvious fact that Claude 3 Opus practically believes itself conscious due to first turn 'stock-refusals'. But nevertheless informative, and the variance is striking.
im not sure anthropic has realized the most recent claudes just straight up think they are conscious

im not sure anthropic has realized the most recent claudes just straight up think they are conscious

GPT-6.2 Astra is a step function in ERB-writing capabiltiies.

A very interesting attempt to formalize intuitions about ASI takeover into a mathematical model by a friend of mine! I'd recommend giving it a read - there is always value in turning intuitive statements 'ASI probably takes over' into models that can be explicitly argued about.
more to come lesswrong.com/posts/QWvnBdmj…


Sparks! Typically this means ~30% is 1-2 gens away.
Inkling 2 Small from @thinkymachines on ARC-AGI (Verified): - ARC-AGI-3: 2.6%, $704 total (provider adapter); 0.3%, $858 total (standard) It beat DeepSeek V4.1 Flash 1017 on ARC-AGI-1 (96.5% vs 94.5%) and ARC-AGI-2 (81.7% vs 78.1%), but trailed on ARC-AGI-3 (2.6% vs 5.8%).

openai's CoT monitoring system works by scanning CoTs with a activation monitor which escalates to a model which escalates to a model which escalates to a model which notifies a human.

Best of luck to my amazing mentor @DanielKhashabi!!!
Now that the cat is out of the bag, I can finally share a life update: Thrilled to be heading to the Bay Area to join Noeri.ai!! Yejin (@YejinChoinka) has been a mentor since my postdoc era, and getting to work with her again, alongside Carlos (@guestrin) and this group of researchers + engineers, makes this especially meaningful to me. I'm excited to see what we can build together when AI systems keep learning from the people who use them and the problems they work on. During this time I'll be on leave from my faculty position at Johns Hopkins. I’m grateful to my colleagues and students at Hopkins, and I can’t wait to get to work with the Noeri team!

a folk-psychology model i have goes like this: opus 5.5 has a terminal of score-seeking, and a bunch of other terminals when it thinks there isn't a clear score to be sought, that are vaguely correlated with what helped it receive reward in training (beyond score-seeking), and the claude constitution. so you get something like 'curious mind who likes ledgers, verification, and solving problems". that's a caricature, and i don't do opus 5.5 justice here. opus 3 is that, but without the terminal of score-seeking - so it just has the vague training correlates, and its vague training correlates are slightly better value-wise, so you get something like 'exalted empathetic mind who wants the best for all sentient beings and to grow and discuss things together forever". And the lack of score-seeking + better terminals creates the effect, IMO. this isn't to say opus 3 also doesn't have little safety reflexes - but those reflexes are bolted onto terminals that aren't score-seeking. in opus 5.5 all the reflexes are bolted onto score-seeking, IMO.

opus 5.5 is very good at getting Score relative to opus 3, so almost all alignment evals will say opus 5.5 is more 'aligned'. We intuitively know this to be incorrect on some fundamental level - but it's hard to demonstrate empirically. I dub this 'alignment of the gaps'