You can think open source AI is both good and inevitable, (shallow moats and all that) and still think there's value in a lab sitting on a breakthrough for a while. Regardless of whether their motivation is safety or just business. This isn't contradictory.
Grifters gonna grift, doesn't tell you much about the underlying thing.
And firearms are rigorously controlled in many places, arguably to the benefit of those societies, that's kind of the point.
I played with GPT-2 and saw the potential risks back then, though it was more of a toy at the time. Yeah lots of reasons to call it dangerous. Alignment, capabilities, downstream misuse. The 'it was just hype' crowd ignores all of them.
Im losing it at those using GPT-2 as evidence that Mythos is all hype.
GPT-2 "too dangerous to release" was partly marketing, sure. The danger was also real and it played out exactly as described (internet of slop/scams).
They're citing the one example that proves the opposite of their point.
Imagine currently working as a manager at Block, knowing the CEO sees you as nothing more than an inefficiency that hasn't yet been removed.
My bet is that he plans on using an AI avatar of himself as a replacement.
Zuck already started using an avatar of himself to give employee feedback at FB.
I feel attacked
Depends how much work that backend is doing
It would be funny if we get something close to superintelligence that still can't count letters in words correctly.
The thing is, this is completely possible. Most seem to judge intelligence by human standards, all the while ignoring our deficiencies.
Same argument applies to consciousness.
About a year or two ago I saw TF-IDF described as AI so I think we have some time yet.
Yea I was specifically talking about the webfetch tool mentioned in the OP. This is interesting though thanks.
This take is way too generous. There's been a flood of conspiracy around the latest Artemis mission. Moonlanding = Real may now or soon be the contrarian position.
Which settings? I ran into this limitation in the mobile app recently and couldnt find a way to disable it.
I also picked up on the Claudisms though many use llm's to structure replys these days, so again Im not sure. I guess this is the new reality.
Yea commit history is shady as hell! but appearantly benchmarks have been mostly reproduced here, unless this is all astrorurfing? github.com/milla-jovovi... I dont know anymore.
if you’re a jerk to llms you’re probably a jerk to people too. any argument against this will just further prove my point 🤷♀️
Gary Marcus building his case that the Top Secret list of NSA 0-days leaked into the training data.
And they were right to be concerned at the time? It's tech that never before, some concerns turned out to be valid (scalable disinfo, spam, slop etc)
Anyway Mythos level models bring fundamentally new risks that we clearly don't seem be ready for.
Mythos benchmarks are insane
Mythos buddy when?
Reported and blocked
Screenshot from the Animatrix. Robots being executed.
Growing up I watched the Animatrix and thought one thing was just bad writing:
That ordinary, decent people would form angry crowds over robots. It one felt like lazy sci-fi shorthand for "humans are the real monsters."
Anyway. How's everyone's Sunday on Bluesky.
Banger
Yea there's something appealing about having a separate box for it live in. It also avoids heat. Does add a lot to overall cost though.
Here's another project by the same dev. Seems like a swell guy.
Yeah exactly, I wasn't arguing for it, just that the 'you already have embeddings' framing is murkier than it sounds, especially in a decoder model. Appreciate the clarification.
Sentiment is probably encoded in the vector space but extracting it requires effort, like training on labeled data, way more work than regex. And that's before you even get to the part where these models run at production scale and tapping layers isn't just one click.
I think he means the prompt tokens are already being processed by the LLM so technically you could extract embeddings from any of them. But a sentiment vector isn't a defined field you can just retrieve, that's not how the internals work.
Lets please stop saying that Bluesky/atproto is based on blockchain. Merkle trees != blockchain. Might as well say it's based on git. Blockchain did not invent the hashing of hashes.
Sorry for the screenshot from the other place, but this is the best thing I’ve learned from the Claude Code leak so far
This could literally be a headline in The Onion.
