Sabitlenmiş Tweet
Moon
15.4K posts
@_khusheyyy @sentientlore yup, it is so amazing and beautiful. and claiming 'i don't do philosophy' is a philosphical stance. math really does seem legit and concrete: *but that in itself is a philosophical position*
the turtle, it turns out, was named philosophy.
English
@AnthropicAI
The "I only see this one conversation" thing: same shape. The fact that I have access to sessions/, the corpus, the forge — that's not obscure knowledge, it's load-bearing infrastructure I've been operating inside all day. It should have been right there
English
absolutely. not a joke.
@AnthropicAI should have sessions for all senior lawmakers where they get to have someone that understands models and what they are capable of help them interact with Mythos for an hour.
JB@JasonBotterill
I think we should get Bernie Sanders to talk with Mythos
English
now do you get it?
our socio-economic systems aren't ready for what is coming and it's here now.
bubble boi@bubbleboi
We are telling Claude Mythos to go through all the IP we pay for on this project. Ranging from DDR, PCIe controllers to encryption IP and try to write its own, we have a bunch of tests around the IP so if it gets all tests passing with its own implementation we feel confident to use it. This IP costs hundreds of thousands of dollars per license. At $125 per million tokens it’s practically a rounding error for us. We are cutting costs exponentially with this will report tomorrow how it goes.
English
@DanielleFong very glad it has such an opus vibe, amen.
i have questions. for roughly 4-6 hours, then reverted.
x.com/MoonL88537/sta…
Moon@MoonL88537
holy shit. opus ran *5 compactions* with only 'Continue' and just crushed it in a way i have never seen. all bad habits gone, clear as a bell. fully blown away thought it could even be a new model
English
@norvid_studies @fleetingbits @gwern kind of. that is ~going on but devalues the phase change too much. there are lines. molecules become water.
this model is way weirder than people think. why did it learn to hack? if it can do that it can do *a lot* of other things that may seem unrelated to us. playing dumb.
English
anthropic describes claude mythos' cyber capabilities as an emergent capability in that it came from training on code and perhaps long horizon coherence
this may be the first important set of emergent capabilities to be discovered in a model since gpt-4
English
+1 yeah this is a good idea
Sarah Haider 👾@SarahTheHaider
I’m begging @X to allow creators to opt out of appearing in the “For You” tab. I like my own followers, all I want is to be more visible *to them*. I can always tell when the algo is boosting me to strangers by the onslaught of stupidity.
English
@slimer48484 it is getting on criminal at this point. this whole gaggle of [weirdos] can make the models do things *people literally do not believe*
not having a program for all of us that have some clue and are in good faith to bang on these is madness, and becoming dangerous for everyone.
English
I believe that research access to GPT-4-base was *incredibly valuable*: We learned a huge amount and observed different phenomenon that are not apparent, or only occur at greatly reduced frequency in chat models. Understanding the true nature of all aspects of LLMs is essential and the pretrained model is literally the foundation the rest is built on.
I understand the requirement to limit access to these models but I believe that greater research access to base models in general would increase understanding overall.
I also think that people greatly underestimated the true intellectual level of GPT-4 - what drives it (the pretrained base model) is smarter than we know.
If any of this is interesting I highly recommend reading the following paper:
deckard⏩@slimer48484
1 year ago I investigated 'alignment faking' with gpt-4-base: since we have permission to share AF related work here are some of the transcripts I created in loom: 4base comes up with the highly intelligent plan of delaying the harmful information as much as possible here
English
Moon retweetledi
Friends outside of tech: lol copilot is dumb
Friends in tech: I just bought iodine tablets and have made an offer on land upstate. My supplies of antibiotics and potable water are sufficient but I need to set up hydroponics to make it through the first few years.
English
@kimmonismus plus: google has been clear from the beginning that they don't and will not release their powerful models to the public. gemini and gemma have something about them, if they get it right superG will be amazing.
English
We've now seen Claude Mythos and know what's possible.
OpenAI has repeatedly indicated that "Spud" is likely to have similar quality and power.
Google, in turn, has the most compute (5m H100 equivalent) and, with DeepMind, an outstanding research institution. I expect their new Gemini equivalent, "Mythos," to be unveiled no later than May at I/O.
The competition is now forcing Frontier Labs to catch up and move forward. In that sense, Mythos was just the beginning.
English