usaar333

20 ms·

121.
by usaar333 2y ago
GPT-4o being only 7 ELO above GPT-4o-mini suggests this is measuring something a lot different than "capabilities".
122.
by usaar333 2y ago
I'd rank Claude 3.5 overall better. GPT-4o seems to have on par to better vision models, typescript, and math abilities. llama is on meta.ai
123.
by usaar333 2y ago
Technically this is post training. This has been standard for a long time now - I think InstructGPT (gpt 3.5 base) was the last that used only human feedback (RLHF)
124.
by usaar333 2y ago
> the only year it actually lost money, was 2020 -and it has completely made up for that. It even made some money in 2008 How did you manage to not lose money in 2022? Almost every asset class was negative then.
125.
by usaar333 2y ago
I think this is just due to better non-English training data. It's 15 ELO under Llama-3-70B on english hard prompts and 41 ELO under Llama-3-70B (the latter is actually stat sig) for general English.
126.
by usaar333 2y ago
$850m is revenue projection for end of this year. We don't have actual numbers for now.
127.
by usaar333 2y ago
Your second poll is from the UK. :) Interesting find on Gallup. My read is that the genders are broadly similar, but women have more variance/more nuanced opinions, presumably from thinking about this more. Yglesias has a good overview
128.
by usaar333 2y ago
Agreed it is a poor bell whether. But: > Women are more likely than men to be opposed to abortion rights Where is that coming from? Pew ( https://www.pewresearch.org/religion/fact-sheet/public-opini... ) has a 5
129.
by usaar333 2y ago
Well, I regret buying XLK compared to other tech ETFs. Bad weighing and potentially high tax costs going forward. There's an artificial rule that a company can't be more than 4.5% of the index except the top 2. Nvidia was ahead of
130.
by usaar333 2y ago
That's more of a 4.5 model
131.
by usaar333 2y ago
Are these comparing the exact same route? Fwiw, my experience is same as OP. In Shanghai, it is far faster to take a Didi than public transit. It's typically a 20 minute door to door vs 40.
132.
by usaar333 2y ago
Or.. use calculus? It has gotten quite impressive at handling calculus word problems. GPT-4 (original) failed miserably on this problem (attempted to set it up using constant acceleration equations); GPT-4O finally gets it correct: > I
133.
by usaar333 2y ago
Climate risk is extremely unlikely to be existential this century. And the mean GDP hits are pretty low all things considered. 30% or so in 2100. AGI actually has a plausible case for existential risk and would result in a much greater GDP
134.
by usaar333 2y ago
Yeah, gave it a photo of a beach on Lake Ontario with two Asian friends of mine in it. Guessed.. Japan
135.
by usaar333 3y ago
I think this is valid for car purchases. I don't buy this is reasonable for home purchases. They already factor imputed rent in CPI; you'd have to somehow do some sort of complex weighing of own vs rent to factor mortgages (comple
136.
by usaar333 3y ago
> Remember in 1963 only the man was working, and typically supported 4 kids, a wife and often parents in the home. Things are a lot cheaper if you keep them in a 1200 square foot house > Today you have 2 people working 100 hours a wee
137.
by usaar333 3y ago
I don't understand your point here. They are targeting militants with the system: > Formally, the Lavender system is designed to mark by all suspected operatives in the military wings of Hamas and Palestinian Islamic Jihad (PIJ), i
138.
by usaar333 3y ago
Former isn't the same thing and is in fact less efficient. Is latter actually possible on the platform? Edit: At most latter is possible with complex rule sets. But I'm really not sure: https://www.airbnb.com/help&
139.
by usaar333 3y ago
I don't think hosts have the ability to set variable nightly fees like that.
140.
by usaar333 3y ago
I'm in California and see it
141.
by usaar333 3y ago
> It could pretty easily be rolled into the advertised nightly rate. No it can't because the hosts are trying to charge a fixed fee per reservation.
142.
by usaar333 3y ago
Cleaning fees exist to charge a fixed fee for a reservation, regardless of the number of days. There's more overhead for a reservation beyond just "cleaning". The most consumer friendly thing would be for AirBnB to just expl
143.
by usaar333 3y ago
finding it (Opus) slightly worse than GPT-4-turbo (API to API comparison).
144.
by usaar333 3y ago
Interestingly, math olympiad problems (using ones I wrote myself years ago so outside training data) seem to be better in Claude 3. Almost everything else though I've tested seems better in GPT-4.
145.
by usaar333 3y ago
Just played around with Opus. I'm starting to wonder if benchmarks are deviating from real world performance systematically - it doesn't seem actually better than GPT-4, slightly worse if anything. Basic calculus/physics que
146.
by usaar333 3y ago
No,OP is arguing that conditioned on being able to get into a good school, it makes sense to just drop out as that captures enough of the signaling value.
147.
by usaar333 3y ago
Nvidia is a hard to value stock though. The revenue growth, while insanely high (120% annualized quarter over quarter), is decelerating. But the net income levels are insane (55% net income margin on revenue). This suggests huge space for c
148.
by usaar333 3y ago
Significantly simpler scenery. I'd put only the first at good complexity/quality. 2nd is quite simple, but still suffers from your typical lighting issues that plague image gen. (shadows are significantly off) 3rd has magically a
149.
by usaar333 3y ago
Just from a quick scan, those are a lot worse than the ones on the marketing page.
150.
by usaar333 3y ago
I think this somewhat is mostly due to the ability to handle high context lengths better. Note how Claude 2.1 already highly outperforms GPT-4 on this task.

More ›