kromem

8 ms·

61.
by kromem 2y ago
Their manipulation of the vectors and the effects produced would suggest that it isn't that the SAE is just finding phantom representations that aren't really there.
62.
by kromem 2y ago
The Othello-GPT and Chess-GPT lines of work. Was the first research work that clued me into what Anthropic's work today ended up demonstrating.
63.
by kromem 2y ago
At least for right now this approach would in most cases still be like using a shotgun instead of a scalpel. Over the next year or so I'm sure it will refine enough to be able to be more like a vector multiplier on activation, but simp
64.
by kromem 2y ago
I don't think that's exactly accurate. What's likely different is that GPT-4o can output the tonality instructions for text to speech now. It's probably the same voice, but different instructions for generations. One was
65.
by kromem 2y ago
Most impressive was the incredulity to the 'okay' during the counting demo after the n th interruption. Was quickly apparent that text only is a poor medium for the variety and scope of signals that could be communicated by these
66.
by kromem 2y ago
Ah, so Thanos was just conducting a census.
67.
by kromem 2y ago
The argument of "we can't say because of all we don't know" needs constant updating. Early on, you have Elihu in Job arguing that we can't understand creation because why it rains and where snow comes from is bey
68.
by kromem 2y ago
I dunno man, maybe I'm just old but waiting "a few weeks" between announcement and free product offering doesn't feel at all like any historic rate of release elsewhere in tech, and certainly not DLC. Elden Ring came out
69.
by kromem 2y ago
While I largely agree with the premise of a converged multimodal world model, I'd argue it's more like Plato's images/ eikons than the ideal forms/ eidelons . The representations don't seem to be lossless, and
70.
by kromem 2y ago
Yeah - he's one of the only people I've seen talk on the topic who really seems to understand where it's going and how to get there. It's possible he's evangelized others at OAI who can carry the torch, but I'm
71.
by kromem 2y ago
We're underestimating the degree of world modeling taking place. The only models we are good enough at proving this is happening in are smaller toy models we can fully introspect. So we know it's happening to a degree. We also k
72.
by kromem 2y ago
The 5 model is probably around the corner, and will probably be Pro only. Until then, 5x higher usage limits on Pro and chat memory are the selling features.
73.
by kromem 2y ago
Around six months ago I did a private trends presentation for old clients where I mentioned that the industry was sleeping on the importance of ego modeling in LLMs for reasoning and overall performance, and that the fine tuning them into &
74.
by kromem 2y ago
It is true, but when it's not zero shot there's a possibility that you are introducing additional information with the example. As well, depending on the study there have been issues with effectively a 'halting' assistan
75.
by kromem 2y ago
Bumble's entire product value was initially built around improving conversation initiation, which is a major hurdle in the dating app market. So no, you absolutely are going to need the virtual back and forth so you can hand it off to
76.
by kromem 2y ago
I can't wait for a subscription based open world where the paths of other players caches the generations for the next ones. So your subscription pays for generations, but also fills out the persistent narrative and lore. I actually lik
77.
by kromem 2y ago
Generally 'discovery' relates to experimental results and not theory. For example, it seems like it would be more appropriate to say that the quantized nature of light was discovered in the early 20th century when confirmed by exp
78.
by kromem 2y ago
Superderminism seems the odd choice to embrace here. Also, given the Frauchiger-Renner paradox, the Occam's razor for fewest assumptions between the two would be contradictory outcomes being what needs to be embraced. Superderminism do
79.
by kromem 2y ago
The only way to really do it is to add a second layer of processing that evaluates safety while removing the task of evaluation from the base model answering. But that's around 2x the cost. Even human brains depend on the prefrontal co
80.
by kromem 2y ago
Well, we just discovered a sync error, so that might be a good edge case to start on: https://www.science.org/content/article/quantum-paradox-poin...
81.
by kromem 2y ago
Are you saying that every news outlet that reports false claims by eyewitness accounts which turn out not to be true and aren't sufficiently loud about retractions should be banned by their host countries? Because I can think of quite
82.
by kromem 2y ago
I certainly hope it's not GPT-5. This model struggles with reasoning tasks Opus does wonderfully with. A cheaper GPT-4 that's this good? Neat, I guess. But if this is stealthily OpenAI's next major release then it's clea
83.
by kromem 2y ago
It sometimes feels like I've taken crazy pills watching what was effectively a tech demo that went viral become the usecase now dictating billions of dollars of development and optimization. It's a crappy usecase. And much better
84.
by kromem 2y ago
I think people here are underappreciating just how much generative AI is going to solve the chicken and the egg problem for VR that's been plaguing it. Within the next five years, at least tens of millions of people will have interacti
85.
by kromem 2y ago
Ah, ok. My variation is it's a vegetarian wolf, a carnivorous goat, and a cabbage. There's a few different hacks that will get it to work, but one of the more interesting is switching the nouns to emojis. But almost none of the mo
86.
by kromem 2y ago
Mine is also the river crossing puzzle. What's your variation?
87.
by kromem 2y ago
We may be talking about different logic puzzles? The only model I've seen that didn't need some rather extreme adjustments to eventually solve it was Mistral large.
88.
by kromem 2y ago
I thought you were being hyperbolic and that this was maybe related to quantum neural networks or ML in at least some way, but no - you seem to be totally correct that this is about as 'AI' as snake oil health products are 'q
89.
by kromem 2y ago
LLMs can't is such an anti-pattern at this point I'm surprised that anyone still dares to stake it. The piece even has an example of a $10k bet around a can't being proven false in under a day, but somehow doesn't th
90.
by kromem 2y ago
I actually realized recently that this is probably the underlying phenomenon behind "how is my phone listening to my conversations to show me ads/articles?" The other day I was thinking about LLM aggregation and in my interna

More ›