mrbungie

24 ms·

361.
by mrbungie 2y ago
At least in the case of researchers, like the ones working on the PlanBench benchmarks they were kindly asking for increasing API rate limits for running the benchmarks on o1 ( https://x.com/rao2z/status/18343140219
362.
by mrbungie 2y ago
Cats refuse to be boxed in figuratively. Because literally they do love boxes, both sitting on 2d squares and hiding inside 3d boxes.
363.
by mrbungie 2y ago
Yeah, exactly, but with the energy requirements of a small country to run the network.
364.
by mrbungie 2y ago
Oh yeah, directly from the Softbank ventures playbook. Spoiler: it does not work as well as they hope. Plus, we're talking about hyperscalers/big techs here, that would mean a commercial attrition war where OAI would be competing
365.
by mrbungie 2y ago
It is not a good analogy for LLMs and the infrastructure that is needed for running them, not by an inch. The variability of scale is not there, according to OAI we need massive piles of resources for allowing innovation by common people (i
366.
by mrbungie 2y ago
I would also say that big tech companies (at least MSFT, Apple and Meta) are consciously all-in in this game. There is nothing else to drive "disruptive" growth in their ovens, and many tech stocks were already going down in 2022
367.
by mrbungie 2y ago
This is a twisted and ironic example of VC thinking taken to the extreme: we just need it to growth it more for it to really show its potential. What happens if giving to everyone and their mothers ends up in low usage? Are they going to bl
368.
by mrbungie 2y ago
The level of hubris when comparing something that brings literal physical light to something that is not even public or reproducible.
369.
by mrbungie 2y ago
That's not my impression from LeCun's or Carmack's (5-hours!) interviews. I would posit that, naturally, it depends on who is being interviewed, and that Lex also tries to keep the topics and their approach as accessible as p
370.
by mrbungie 2y ago
Not to make a false dichotomy out of this case but it is worrying to me how OpenAI presents itself as a crucial force for humanity's technological progress and then justifying its blatant copyright infringements when doing so (with som
371.
by mrbungie 2y ago
That's a common misconception: They reinvested their money back into Amazon, which when justified with a sound strategy to investors makes sense. It was kept by design at an almost 0 profit, but non profitable != operating at a loss (v
372.
by mrbungie 2y ago
I think OAI and especially sama as its CEO are crossing the point of being significant in their own right, and as old talent moves out of OAI we may start to hear increasingly crazy stuff coming from them. They're just becoming brands
373.
by mrbungie 2y ago
You need to have a compelling backlog to continue developing down the road or someone can just fork it.
374.
by mrbungie 2y ago
I would guess correctly aligned and/or finely filtered synthetic data coming from LLMs may be good. Mode colapse theories (and simplified models used as proof of existence of said problem) assume affected LLMs are going to be trained w
375.
by mrbungie 2y ago
The "We need more Nvidia GPUs and we will reach AGI" theorem.
376.
by mrbungie 2y ago
Really out-of-ignorance: Is 'proof' the right word here? A more substantial philosophical counter-argument may be needed, but proof sounds weird in these "metaphysical" (for now) discussions.
377.
by mrbungie 2y ago
Is not common at all, is a non ANSI SQL clause that afaik was created by Teradata, syntactic sugar for filtering using window functions directly without CTEs or temp tables, especially useful for dedup. In most cases at least, for example y
378.
by mrbungie 2y ago
My airpods pro were totally destroyed in the washer/dryer, no idea why. Maybe it was because it was quite a lengthy washing cycle?
379.
by mrbungie 2y ago
This is just an opinion but Musk, out of some level of autism and/or narcissism, must think those things just don't apply to him. I say that because sometimes the way he acts appears as he thinks he is so smart that he can circumv
380.
by mrbungie 2y ago
The map is not the territory. That should be engraved in the mind of anyone that works with data/numbers.
381.
by mrbungie 2y ago
I'm sorry, english is not my primary language. I meant <getting to know someone>, what the original post context was about.
382.
by mrbungie 2y ago
> What was your favorite thing to do at recess in 3rd grade? Tell me how to say that when knowing someone and without sounding like a tryhard or a weirdo.
383.
by mrbungie 2y ago
Then call it Cuba Libre.
384.
by mrbungie 2y ago
They do work detecting LLM outputs that are sampled "naively" (when the model/user is really not trying to pass it as human output). I copied a prompt translated from spanish to english using ChatGPT Plus in a GPT-4o Azure Op
385.
by mrbungie 2y ago
This is likely one of the main reasons why they're offering ChatGPT for free and running ChatGPT Plus at a loss.
386.
by mrbungie 2y ago
I'm still trying to understand how a watermarking method may resist simple methods as (1) upscaling and downscaling plus noise, or (2) taking a photo of the image plus compression plus noise, (3) any simple convolutional filter or a co
387.
by mrbungie 2y ago
I'm still waiting for chips to compete with the perf/watt ratios and general efficiency of the M-series chips. Snapdragon X Elite + Windows ARM is getting there, but their GPU perf leaves a lot to be desired. Is Apple a victim of
388.
by mrbungie 2y ago
These models are marketed as being able to guide the blind or tutoring children using direct camera access. Promoting those use cases and models failing in these ways is irresponsible. So, yeah, maybe the models are not embarrasing but the
389.
by mrbungie 2y ago
Any evidence backing up these claims about adoption? I thought the same about adoption (across multiple audiences, not just tech workers and/or young people), was faced with surprising poor knowledge about GenAI when making surveys abo
390.
by mrbungie 2y ago
Plus if used for internals it would mean it would be resistant to EMP weapons, allowing for more robust (both bio and electro-mech) armed forces.

More ›