@matei_zaharia May have missed in the link, but is the IE bench dataset you used to benchmark publicly available?
English
Aria Haghighi
3.9K posts
Aria Haghighi
@aria42
Tinkering on new ideas
Seattle, WA เข้าร่วม Şubat 2008
413 กำลังติดตาม3.6K ผู้ติดตาม
Prompt optimization is becoming a powerful technique for improving AI that can even beat SFT! Here are some of our research results with GEPA at Databricks, in difficult Agent Bricks info extraction tasks. We can match the best models at 90x lower cost, or improve them by ~6%.
English
@Jonathan_Blow Agreed! I think it’s due to growing up during music streaming era and listening way more broadly in terms of genre and era than past distribution models.
English
It seems like a sizeable fraction of Gen Z actually have taste in music, such that there are rising Gen Z pop stars who are actual good musicians again?
Whitepill...
English
@RoseSilicon @pfau Important is different from abstracted away. Understanding and being able to debug raw code will still be important but it also will be abstracted away and something only a small percent of engineers will be able to effectively do
English
Starting to wonder if a big reason LLMs for coding are gaining so much traction is not their quality, but their addictiveness, in the same way social media is addictive. It's so tempting to just press the "generate" button like it's a pachinko machine rather than sit and think.
English
@BlackHC @wellingmax I don’t think companies hold principles. They do things to signal and curry favor. In the past, tech companies were labor constrained and now if they remove a few calendar invites, they’re more likely to get the administration to allow them to build datacenters.
English
@wellingmax I'm concerned because it looks like everyone is backtracking without hesitation (incl on other principles as well)
English
What is going on here behind the scenes? Is BigTech pressured by the Trump administration to shed all DEI efforts? I find this trend of BigTech kneeling before the new emperor very concerning.
search.app/LLSDenGsmaQHqw…
English
@ChrisGPotts In fairness, Twitter barely captured a fraction of the value they created either. Agreed OpenAI will probably hit profitability, but I'm still skeptical you'll see meaningful margins given competitive landscape.
English
I am confident OpenAI will become profitable. They are smart, creative, highly incentivized, and well-funded. On the other hand, any app/company that depends on capturing most of the value from OpenAI's models has an uncertain future, like the Twitter apps of old.
English
@DominusClarus @amasad Quoting from the link you shared "[UHC] dismissed about one in every three claims in 2023 — the most of any major insurer. That’s twice the industry average of 16 percent"
English
@aria42 @amasad Adjusting this 8.7% number (seen below) from 2022 with the reported 31% increase from 2022 puts the claim rejection rate at 11.5%. Roughly in line with the results of this survey.
bostonglobe.com/2024/12/05/dat…
English
The typical healthcare insurance company is making 3-5% profit margin, barely for-profit companies. They suck because they’re pseudo governmental corporations with regulatory capture.
Secular Talk ([email protected])@KyleKulinski
English
@DominusClarus @amasad The specific claim is about UHC, which is what the original picture is commenting on. That still has a substantially higher denial rate than the categories in your image.
English
@DominusClarus @amasad Can you provide a link to justify the claim? Asking perplexity and their is a 2x diff between UHC vs ACA plans. perplexity.ai/search/uhc-cla…
English
@alex_gurevich @IsraelHayomHeb The Pentagon has denied she’s a suspect on the record. What part of the Hayom piece has actual evidence? I’ve read the English translation and might’ve missed something, but that’s also reporting allegations.
jpost.com/breaking-news/…
English
I saw the Sky News source as well as Israel Hayom (@IsraelHayomHeb) which is Israel’s most widely distributed newspaper. But you’re right that it is not yet confirmed by the U.S. There was also a lot of speculation it was her before these reports came out. So evidence seems to be building.
English
Individual campaign contribution limit is $3.3k
He means Trump’s JFC which is legally prohibited from coordinating with the campaign. In reality, everyone knows it’s a way for the donor class to make unlimited contributions. Political opinions aside, this is bad for democracy
Shaun Maguire@shaunmmaguire
I just donated $300k to President Trump The timing isn't a coincidence
English
@AndrewLazarus4 What data are you looking at? All the charts I've seen look roughly like this. To be clear, I think a lot of what is going is "seasonal" like the quote retweet says, but people aren't wrong this has gotten worse fred.stlouisfed.org/series/CUUR000…
English
@aria42 Buying power (real wages) has increased 12 months in a row.
To paraphrase DuBois, a lot of this is about the psychological wages of MAGA.
English
Not surprising people don't know formal definitions of recession or inflation. Perception of the economy is about your buying power and that has gotten worse for most. "Vibecession" discourse feels like sticking your head in the sand and ignoring this.
theguardian.com/us-news/articl…
English
@SijunHe @RaniNelken Was this one material in the case over purchasing?
English
@jayvanbavel The chart you're showing are mostly not stat-sig differences and paper reports exact p-values for this figure confirming. The study has lots of other issues, but why highlight these differences?
English
Banning #smartphones in over 400 schools led to:
-decreased psychological symptoms among girls by 29%
-decreased bullying by boys and girls by 43%
-increased GPA among girls by .08 SDs
Effects were larger for girls from low SES families openaccess.nhh.no/nhh-xmlui/bits…
English
@irl_danB What’s incorrect in the description? There is blurring the line between what’s inherent to the model architecture (the recipe stuff is broadly similar to FFN key-value “memories”) and how we create demonstrations (the user-agreeable tone).
English
without any disrespect at all I’m not sure Yann knows how LLMs work
which feels amazing and improbable and bewildering and makes me question my own sanity because how is that even possible
am I the crazy one chat?
Yann LeCun@ylecun
How to be as "smart" as Auto-Regressive LLMs: - memorize lots of problem statements together with recipes on how to solve them. - to solve a new problem, retrieve the recipe whose problem statement superficially matches the new problem. - apply the recipe blindly and declare victory. - do not use basic logic. - do not use common sense to check your solution. - do not use a mental model of the situation as a sanity check. - do not simulate the scenario in your mind using your world model. - when someone tells you your solution is wrong, reply "I'm sorry, you are right" and apply another irrelevant recipe. Knowledge accumulation is not a substitute for actual understanding.
English
Just joined airchat.
First thought: a fresh new take on social media. But have my doubts.
Made by @naval.
It’s Invite-only. So comment and I’ll send an invite
English