bigyikes

6 ms·

31.
by bigyikes 2y ago
Is there evidence that Apple is training a model large enough to require a huge amount of compute?
32.
by bigyikes 2y ago
Interestingly, your description of remembering what something looks like matches my experience exactly, but in my dreams I see vividly. Aphantasic except in my dreams?
33.
by bigyikes 2y ago
Since there are no people on board…… I hope it blows up
34.
by bigyikes 2y ago
Turn up the temperature? Markov chains have a cruder understanding of language. Turn up the temperature (the “randomness”) of an LLM and you can achieve a similarly crude approximation. Further, author uses ChatGPT-3.5. ChatGPT has been rlh
35.
by bigyikes 2y ago
Show me the statistics. If there is data which shows that autonomous vehicles are significantly safer than than the 90th percentile driver, then I absolutely will put my children in it. In fact, it might be irresponsible for me not to. Scho
36.
by bigyikes 2y ago
In a hypothetical future where the reliability of LLMs improves, I can imagine the model being able to craft optimizations that a traditional compiler cannot. Like there are already cases where hand-rolling assembly can eke out performance
37.
by bigyikes 2y ago
I’m begging for trains in the Texas triangle. It’s such a perfect use case. It’s ~3 hours between any of San Antonio-Austin, Houston, and Dallas-Fort Worth. Long enough to be a horrible drive, short enough to be a waste of a flight. There’s
38.
by bigyikes 2y ago
There is no inconsistency. The first quote is meant to highlight that commercial planes have autopilot while e.g. Cessnas do not. The second quote emphasizes the importance of training despite autopilot. There is no inconsistency in pointin
39.
by bigyikes 2y ago
I mean… a little bit, yeah. These giant trucks and SUVs really do increase danger, noise, road wear, spot size, etc. Edit: Less than an hour later I got ran out of my lane by a guy in a truck :’) That’s what I get for driving a tiny car, I
40.
by bigyikes 2y ago
>It just moves the homeless around, satisfying nobody permanently. I bet it satisfies the people who live near these encampments. It’s also not obvious that it just moves people around. Tolerating encampments incentivizes that sort of li
41.
by bigyikes 2y ago
I tried to sign up but never got a confirmation email.
42.
by bigyikes 2y ago
Would this be an issue if we called a spade a spade? These aren’t boneless wings, they’re chickie nuggies :) Is it legal for my McNuggets to contain bones?
43.
by bigyikes 2y ago
Once self driving is solved, the custom hardware and software is largely a fixed cost which can pay for itself over long time horizons. Meat-based drivers must be continually paid and don’t get any cheaper over time.
44.
by bigyikes 2y ago
Vehicle depreciation doesn’t matter unless you’re selling the thing. Wear and tear will be reduced with an all-electric drivetrain, which is uncommon for Ubers outside of California. Costs will be further reduced by bulk maintenance and sol
45.
by bigyikes 2y ago
Are they foolish otherwise? They’re making 50,000+ rides per week already, with a very limited rollout. They only have to match the cost and convenience of Uber in order to utterly dominate the market. Once they fully solve self driving, wh
46.
by bigyikes 2y ago
Sad? Or maybe an indication that your opinion isn’t the universal truth…
47.
by bigyikes 2y ago
Do you have any indication whether this “universal” space might be shared between models, or is it unique to the architecture and training set? Maybe it’s crazy, but is there any possibility that, say, Llama and Mistral use the same represe
48.
by bigyikes 2y ago
Very informative paper. They find: * Inner layers of transformers share a representation space * Some middle layers can be dropped without total failure (though it results in reduced performance) * Middle layers are not interchangeable, the
49.
by bigyikes 2y ago
Starter pricing was 0.5%. Scale pricing was 0.8%. Both plans were consolidated into a single plan which is 0.7%. This is good news if you were a Scale customer and bad news if you were a Starter customer.
50.
by bigyikes 2y ago
It’s a 60% increase in fees. It’s a 0.3 percentage point increase in fees. The former is important because it reflects the cost increase customers will face. Their costs don’t increase by 0.3%, they increase by 60% (Except I don’t think any
51.
by bigyikes 2y ago
Agreed. There is no world in which React is faster than vanilla JS, except maybe certain contrived examples. Selling React as “fast” has got to be one of the greatest bamboozles of all time. I know I thought it was fast when I first started
52.
by bigyikes 2y ago
I have this app open right now, actually. It's honestly... okay. I wish I had something better installed. It gets the job done, but the UI leaves something to be desired. It's free though, so I can't complain. I use it to man
53.
by bigyikes 2y ago
The economy is predicated on future growth, so if growth slows due to lack of workforce, bad things may happen. Social programs become insolvent. It becomes difficult to take care of an aging population. (Not sure what makes you say “cheap
54.
by bigyikes 2y ago
“Ultimately, it comes down to taste.” - Steve Jobs
55.
by bigyikes 2y ago
Give me B! Fasteners are fun. Have you been to mcmaster.com? It’s a master class in online shopping UX. Their fastener section is a joy to browse :)
56.
by bigyikes 2y ago
If the law doesn’t take into account human and corporate behavior, it’s not a good law. It may ultimately be the corporation’s “fault” but that doesn’t change the fact that the law created perverse incentives. I hear the same argument for t
57.
by bigyikes 2y ago
What is the alternative to the access problem? Would you prefer apps be able to access that stuff without permission? I get it’s annoying but at least it’s safe?
58.
by bigyikes 2y ago
It could be, but I have a feeling we’ll find out sooner rather than later. Whatever GPT-5 turns out to be will set the tone. If there is a clear increase in intelligence over GPT-4, I don’t think it’s a bubble. If it turns out to be a modes
59.
by bigyikes 2y ago
Do you also find the two term limit for presidents weird then?
60.
by bigyikes 2y ago
>We use synthetic data heavily to create Nemotron-4-340B-Instruct: over 98% of our training data has been synthetically generated throughout our alignment process. Very interesting to see synthetic data used so heavily during alignment.

More ›