jason @agikoala Twitter profile

Pinned Tweet

jason

1 year

best week of my life may have been visiting singapore, due to the density of surreal sensory experiences: 1. staying in marina bay sands 2. flying singapore airlines 3. eating at a hawker center 4. meeting @YiTayML in person IRL

1

51

Last Seen Profiles

@phngsphnhp94294

@elitenational24

@TheStrokeAssoc

@stw_pdg

@jandakembangstw

@Jane4UsGirls

@tucks1089

@daaicsy

@Soyuz_3

@sprudlers

@LMatritenses

@MattBennett90

@StevenR39031

@alexriosSF

@Fri_TFO

@MikeSheldon717

@Filinvest

@physics_beyond

@Amanda461867277

@Legacieseth

@ManuClutch

@stwmaniax

@MattOverlin

@jandakembangstw

@Jhosmar07429897

@stw_pdg

@kohei38loj

@ClassPointMENA

@stw_pdg

@MaxRosent

@AlexMartin33000

@cindybax

@jandakembangstw

@stwmaniax

@gibranteske

@Leonard85026417

jason

@agikoala

9 months

Overheard at a Meta GenAI social: "We have compute to train Llama 3 and 4. The plan is for Llama-3 to be as good as GPT-4." "Wow, if Llama-3 is as good as GPT-4, will you guys still open source it?" "Yeah we will. Sorry alignment people."

89

230

2K

jason

@agikoala

1 year

Three starter pokemon in deep learning

9

48

449

jason

@agikoala

3 months

AI researchers do not rise to the level of their ideas, they fall to the level of their infra, tooling, and evals

2

19

187

jason

@agikoala

9 months

Lol all insiders will know this list is not about who did the work. No Jeff Dean. No John Schulman. No Mira Murati. At least they have Emily Bender and Margarett Mitchell though.

TIME

@TIME

9 months

TIME's new cover: The 100 most influential people in AI

432

1K

3K

10

5

110

jason

@agikoala

1 year

Ai companies should partner with airlines to enable passengers to do manual data annotation to pass the time

6

9

88

jason

@agikoala

9 months

There was a co-worker that I always saw at the office on the weekends. He was in before and still working when I left. He inspired me to work harder. Today I finally asked what made him so motivated: "TBH my girlfriend was in grad school at the time so I had nothing else to do"

0

1

52

jason

@agikoala

9 months

Didn't expect this tweet to go viral. Not sure if it went viral in a good or bad way for meta. Maybe they won't invite me to their socials next time lol

3

0

46

jason

@agikoala

1 year

If Zuckerberg can cancel the Metaverse to focus on AI, you can stop working on the line of research that you have been doing since your PhD and pivot to large language models. Apologies for being blunt, but some people still haven't gotten the message (and you know who you are).

22

3

45

jason

@agikoala

6 months

I cringe a little bit every time i hear the term “natural language processing”

6

2

37

jason

@agikoala

6 months

. @hwchung27 is a big fan of testing. Whenever I pair program with him, we write tests first before the function. Today I had to pay him back for something. He sent me his zelle username and said "before you send can you test with $1 first"? Programming is life lol

2

0

36

jason

@agikoala

10 months

two koalas and an @agihippo

1

27

jason

@agikoala

11 months

People underestimate how much decision making / performance is affected by sleep and mental state. - It should be normal to cancel meetings during the day to recover on sleep - I'd prefer have an elective surgery postponed if the surgeon didn't sleep well the night before

1

27

jason

@agikoala

9 months

When I tweet something that is really thoughtful with proper hedging, no one likes it. I bet i could tweet the same idea, add some spicy take that offends like 20% of people, and then suddenly it is a viral tweet. Incentives here kinda suck lol

5

0

27

jason

@agikoala

1 year

its obvious that i havent fully adapted to the new ai landscape since i still check my google scholar citation count regularly for dopamine hits

1

0

27

jason

@agikoala

10 months

Fun fact: "instruction tuning", which means finetuning a model on instructions, was almost called "i-tuning". At the time I liked short names, and pitched this to @quocleix and @MaartenBosma . Quoc said nah, which is in hindsight good b/c i don't think it would've caught on

2

3

26

jason

@agikoala

8 months

Thinking out of the box is very important in AI. I was playing soccer against @tszzl today. We were at Franklin Square and he said wow what a nice view of the city. When I turned to admire the view he took the ball and scored. Not a traditional tactic but effective nonetheless

1

0

25

jason

@agikoala

8 months

I found an easy way to find out whether someone works hard. I tell them that I am taking a day off this week. Then I say that it will probably be saturday. The hard workers act totally normal, having already assumed a weekend day, whereas those who don’t work hard are surprised.

5

1

24

jason

@agikoala

1 year

The fact that AGI is probably coming in a few years is maybe the most inspiring thing ever, but apparently not inspiring enough to stop me from spending time on social media and watching youtube

1

0

24

jason

@agikoala

1 year

today is a day where i need to remind myself of the quote “those who matter dont mind, and those who mind dont matter”

3

0

24

jason

@agikoala

11 months

sf kids like to shit on south bay, but i did some of my best work in mountain view. not much social life, very few distractions, i was easily clocking in 60-70 hour work weeks

1

0

23

jason

@agikoala

11 months

Sorry for credit grabbing, but am I the first one to use the term "Base LM"? This was actually unintentional since we couldn't reveal that this model was LaMDA at the time. Or did I copy someone? Would love to know the answer. From Flan 2021:

4

1

23

jason

@agikoala

1 year

alan turing didnt need to know german to break enigma and you dont need to know linguistics to be a star ai language researcher

2

0

20

jason

@agikoala

1 year

With AI workers shifting from research to product, the marginal benefit of having an additional ACL paper seems to be much less than it was even just two years ago

3

0

20

jason

@agikoala

11 months

only boomers still say "NLP"

1

0

18

jason

@agikoala

11 months

my mom likes to stay hip with ai so she asked the help desk person at her mid-sized company in virginia for copilot the help desk person replied, "are you a pilot?"

1

18

jason

@agikoala

1 year

The best argument I can think of for not doing a PhD can be explained in only six words: Noam Shazeer, Jacob Devlin, Barret Zoph

4

1

18

jason

@agikoala

10 months

It's hard for me to keep my head down and focus on IC work when I'm compelled to write a tweet about why this so clearly indicates a lack of knowing whati s going on

Yann LeCun

@ylecun

10 months

Capabilities of AI systems do not just "emerge". Lots of people come up with lots of ideas, try many of them, and fine-tune like mad to bring those new capabilities. Then, a considerable amount of effort is devoted to detecting and evaluating capabilities *and* flaws.

48

59

498

2

1

16

jason

@agikoala

10 months

i admire people who write books. however, blog posts seem higher opportunity in today's environment of short attn spans. more opportunities for reads and shares on individual posts, and people can still read the whole thing

3

0

16

jason

@agikoala

1 year

While everyone is trying to mentally catch up with how language models are going to change tech this year, the real winners will be those who are thinking ahead to how language models will be used two or three years from now.

1

0

15

jason

@agikoala

1 year

there appears to be some emergent contention around emergent abilities of large language models

2

0

14

jason

@agikoala

1 year

I felt a bit jaded about giving advice, since I try to be thoughtful about it and most people seem to listen but not follow it. Then I was reminded that even if one of fifty people takes your advice seriously, and that changes their life, it makes it absolutely worth it.

3

0

14

jason

@agikoala

7 months

it is good to train some neural networks

3

0

13

jason

@agikoala

6 months

At a workshop, nine academics each give a five minute talk. How much time does it take in total? Empirical answer: 89 minutes

0

13

jason

@agikoala

11 months

20 likes on this account is a banger for me

1

0

13

jason

@agikoala

11 months

New blog post on my personal journey using twitter (on @_jasonwei main account): For 1 year I tracked follower growth, tweet popularity, and whether I could forecast tweet performance. But I ultimately decided that this was the wrong way to do twitter.

Some observations from tracking Twitter — Jason Wei

In early 2022 I realized that being a good researcher isn’t just about doing good research—it’s also about communicating ideas well and persuading others to engage with them. In other words, research...

www.jasonwei.net

0

1

13

jason

@agikoala

9 months

Fact check update: I checked with someone else who was in the conversation, and that person thought it wasn't said in such a savage way. So my apologies for the color. But the content is still correct I believe

1

0

12

jason

@agikoala

3 months

The highest possible achievement in weightlifting / bodybuilding is if i meet someone new and they are surprised when i tell them i write code for a living

1

12

jason

@agikoala

11 months

i always found twitter to be crazy. i sometimes get 500-1k likes on some random quip or insight that took me 5 mins to write. and then when i post something serious that i put a lot of thought into, it gets like 70 likes. reward signals seem off

4

0

11

jason

@agikoala

8 months

liked this quote from a friend: "life is just about finding a bigger prison"

0

12

jason

@agikoala

6 months

Huge game changer, now Karina doesn't have to manually tell me bedtime stories anymore

OpenAI

@OpenAI

6 months

ChatGPT with voice is now available to all free users. Download the app on your phone and tap the headphones icon to start a conversation. Sound on 🔊

2K

3K

18K

1

0

11

jason

@agikoala

1 year

4. More papers are overall worse for the community, because it brings down the average quality of work and creates noise that people then have to filter though.

1

0

12

jason

@agikoala

11 months

500 followers, you can call me mr popular

2

0

12

jason

@agikoala

1 year

Congrats to @YiTayML on 10k followers! You are a C-list AI celebrity now

1

0

11

jason

@agikoala

1 year

The data is in---I'm a pro at forecasting how many likes my tweets will get. And by pro, I mean the actual # likes is between 0.1x to 10x of what I forecast...

3

1

10

jason

@agikoala

1 year

Several reasons why you shouldn't optimize for publication quantity

2

1

11

jason

@agikoala

1 year

Its surprising how much my eating habits are influenced by other people. I saw @shaneguML grab some gummy bears yesterday and have been craving them ever since, even though those gummy bears are always at the microkitchen and i normally dont crave them

1

10

jason

@agikoala

11 months

I really appreciate the bluntness in Chinese culture. I once told a Chinese friend the PhD programs I was applying to, and he casually went through the list and stated which ones were good enough and which ones I could do better than. Not politically correct but it stuck with me.

0

9

jason

@agikoala

9 months

the dunks on this tweet are kinda funny

@[email protected] on Mastodon

@emilymbender

9 months

@tdietterich @TaliaRinger @mmitchell_ai @ErikWhiting4 @arxiv arXiv is a cancer that promotes the dissemination of junk "science" in a format that is indistinguishable from real publications. And promotes the hectic "can't keep up" + "anything older than 6 months is irrelevant" CS culture. >>

18

9

71

0

1

10

jason

@agikoala

27 days

when i'm too upset to do real work i just fix mypy errors and write unit tests for `` files

0

11

jason

@agikoala

1 year

For me there isn't much difference between weekdays and weekends. Actually I like weekdays better, because there's no food at the office on weekends

2

0

8

jason

@agikoala

9 months

soccer is an ideal hobby sport - good cardio (cf golf) - easy to meet people (cf weightlifting) - cheap (cf tennis) - easy to find a field in the city (cf tennis) - actually fun (cf running) Biggest issue is that it's injury prone

0

8

jason

@agikoala

9 months

Was playing soccer today. A dude aggressively ran into me and I fell. I wasn't hurt but there was a concrete wall near the sideline. A teammate said my head was inches from it. Could've been bad. Takeaways: - be more careful during soccer - what legacy do i want to leave behind?

1

0

9

jason

@agikoala

11 months

I spent a lot of time figuring out good names for stuff I do in research, but turns out there's a benefit of naming stuff poorly---people might name it after you. Example: MATH is such a bad name that people started saying "Hendrycks Math"

2

1

9

jason

@agikoala

1 year

i feel like there is a 90% correlation between whether you make your bed and whether you spellcheck your bibliography in your papers

2

0

9

jason

@agikoala

10 months

Subtext: stop writing papers on instruction tuning without citing me thanks :)

1

0

9

jason

@agikoala

1 year

Was told that it would be funny to tweet that per week, all but 1 or 2 meals i eat are openai food (on fridays i take additional food for the weekend that i re-heat)

1

0

7

jason

@agikoala

9 months

The best gift you can give is your honest self :)

yi 🦛

@agihippo

9 months

some people tell me im too frank/direct on my alt account so i tried to RLHF myself with a sarcastic reward model. i hope that works. or i can continue to RLHF myself.

5

0

12

0

8

jason

@agikoala

1 year

After meeting @karinanguyen_ i was inspired to say start saying yes to coffee chats with people i havent met before

1

0

8

jason

@agikoala

1 year

If someone will go down in history as the “father of prompt engineering”, i’d like to add my name to the mix

0

1

8

jason

@agikoala

9 months

Five years ago I would be bothered by when people would run evals and analyze the numbers but not manually look at any examples. Now I am extremely bothered when that happens.

0

8

jason

@agikoala

1 year

as an american i rarely speak my second language with someone else as their second language, but i did today by speaking chinese with my hair stylist, a burmese woman only other time i can recall doing this was talking to a chinese person at a french casino

0

7

jason

@agikoala

9 months

Another update with very low confidence about truthfulness: i also heard that someone else who wasn't at the event said they had a similar conversation

0

7

jason

@agikoala

10 months

not super well thought-out, but I'm surprised how few people are talking about how llama is not safe or how llama is accelerating us to unsafe agi

1

0

8

jason

@agikoala

3 months

Back in the day we developed new, creative neural architectures to do specific tasks. We’ve now moved up one level in the stack: we develop new prompting techniques, each with tradeoffs for different tasks, implementation overhead, amount of compute used, etc.

0

1

7

jason

@agikoala

1 year

@agihippo mmlu 5-shot doesn't lie

2

0

7

jason

@agikoala

6 months

In 1,000 years, will all humans speak the same language?

4

0

7

jason

@agikoala

5 months

why is it so hard to think clearly

1

0

6

jason

@agikoala

1 year

3. Writing a lot of papers can create an illusion of success, when in reality you might have accomplished nothing.

1

0

7

jason

@agikoala

6 months

I used to write in my bio "CoT prompting was presented by Google CEO Sundar Pichai at Google I/O". At a recent talk, the guy introducing me misread it as "CoT prompting was presented *to* Google CEO", so i came off as citing a meeting with the CEO as an accomplishment lol fml

2

0

7

jason

@agikoala

5 months

GF: why are you drawing some diagrams on a piece of paper me: it's super secret stuff but i'll let you take a peek *hands her paper* GF: are these vim shortcuts?

0

7

jason

@agikoala

1 year

AI celebrity categorization: >10k followers = C-list celebrity >100k followers = B-list celebrity >1M followers = A-list celebrity

2

0

7

jason

@agikoala

1 year

@agihippo @DZhang50

0

6

jason

@agikoala

1 year

Apparently me not getting canceled for tweeting spicy stuff (e.g., disagreeing with emily bender) inspired others to be more active. Not the kind of inspiration i was going for but i guess ill take it

0

6

jason

@agikoala

9 months

@agihippo Could be ACL or NeurIPS, but yep. Basically if you're commenting on a paper at that level, it means you aren't saying "wow this is an amazing work", but rather "it's good enough to be published somewhere" (It's a fine place to start a career, BTW)

1

0

5

jason

@agikoala

1 year

I was recently asked whether AI companies have an expectation of working 60+ hours a week, and I realized that it's been so long since I've thought about "work" in that way. I've always been tracking how much time I worked as something to maximize, not to put a ceiling on.

0

6

jason

@agikoala

11 months

Someone needs to stop me from watching random soccer videos before bed every night. I haven't even played soccer since elementary school house but god damn Virgil van Dijk is a beast

1

0

6

jason

@agikoala

3 months

I realize that I use twitter/X differently than most people on my main account. While most people tweet naturally as thoughts come up, I tend to post multi-paragraph tweets, which is probably the exact opposite of what this platform is designed for. It works well for me though:

1

0

6

jason

@agikoala

11 months

@YiTayML @MaartenBosma Damn. This seems like a slam dunk case of cherry picking but the Inflection team is really high caliber so I'd like to give them the benefit of the doubt. Curious what happened here.

0

6

jason

@agikoala

11 months

You can now hide entire sections on Google docs, game-changer. No good reasons for me to use notion anymore

1

0

6

jason

@agikoala

8 months

@agihippo I think some, but definitely not all. Even if I believe some people are fighting for the wrong thing, if they truly believe in it and are not just promoting something for social status i respect that

0

6

jason

@agikoala

9 months

@hwchung27 has some of the most original takes that i've ever heard. He has a really special way of thinking. Wish he had an avenue where he could say his thoughts freely

1

0

6

jason

@agikoala

9 months

Lol Max @tegmark actually speaks pretty good chinese

1

0

4

jason

@agikoala

10 months

when i have a headache i scroll twitter and it goes away

1

0

5

jason

@agikoala

8 months

sunday is my favorite day of the week because that's when soccer practice is

0

4

jason

@agikoala

1 year

1. Number of papers scales linearly, but impact scales exponentially.

2

0

5

jason

@agikoala

1 year

5. "What about people who publish a lot of very high-quality work?" They are not conflicting in principle but often in practice. This logic is also flawed. You might ask, "What about successful people who aren't organized or who aren't fit?"

0

5

jason

@agikoala

11 months

note to self: anyone who volunteers to participate in a race is probably pretty fast

0

5

jason

@agikoala

10 months

lost to elementary schoolers today in pickup soccer. not sure how that is possible as I watched seven virgil van dijk highlight reels in the past week

0

5

jason

@agikoala

1 year

2. When optimizing for number papers, it's easy to optimize for getting a paper into conferences and not any better than that.

1

0

5

jason

@agikoala

1 year

@colinraffel i really enjoyed your performance in "Her"

2

0

4

jason

@agikoala

1 year

Thought about making an anonymous alt account but decided not to. My relevation was that the intersection of sentences that move society forward and sentences that should be said anonymously is basically 0.

0

1

4

jason

@agikoala

1 year

@agihippo lol

0

4

jason

@agikoala

1 year

the best dating app is twitter

1

0

3

jason

@agikoala

1 year

@agihippo Course not

0

4

jason

@agikoala

10 months

. @agihippo

0

1

4

jason

@agikoala

1 year

@agihippo @sur4js im invited to your hot tub next time im in singapore right?

1

0

4

jason

@agikoala

1 year

@agihippo Rematch next time i see you. I will study up

1

0

3

jason

@agikoala

1 year

really liked a quote from wojciech on lex fridman's podcast: we live in a simulation running in our brain. psychedelics change some hyper parameters of the simulation, allowing us to look back on experience and see that things we took for granted are changeable

1

0

3

jason

@agikoala

6 months

@hwchung27 Apparently the test failed, so thank god for tests lol

0

3