trevor (taylor’s version)
@tmychow
Followers
2,250
Following
603
Media
125
Statuses
1,666
building @moonglow_ai ; prev @hazyresearch
um-helat
Joined July 2017
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
Sinwar
• 657539 Tweets
Rosé
• 618202 Tweets
García Luna
• 446749 Tweets
Bluesky
• 238783 Tweets
#يحيي_السنوار
• 218017 Tweets
Calderón
• 211183 Tweets
スーパームーン
• 189865 Tweets
ブルースカイ
• 183737 Tweets
Tyler
• 96792 Tweets
Marca
• 91882 Tweets
ALRIGHT TPOP STAGE
• 74395 Tweets
غير مدبر
• 55807 Tweets
#OurPreciousFOURTHTurns20
• 53076 Tweets
HAPPY FOURTH DAY
• 52213 Tweets
Monsalve
• 50336 Tweets
Rafah
• 45147 Tweets
#Hamas
• 35185 Tweets
#TUDUMNaLata
• 31487 Tweets
ابو ابراهيم
• 28382 Tweets
対象作品
• 23334 Tweets
Obolo
• 21286 Tweets
冬月くん
• 20057 Tweets
Gisela
• 14977 Tweets
#لن_تبقي_لكم_دبابات
• 14639 Tweets
Macaya
• 14527 Tweets
Karen Nyamu
• 14076 Tweets
ふまふさん
• 13860 Tweets
الصهاينه العرب
• 10281 Tweets
Pinned Tweet
2/3 Start your experiments in a Jupyter notebook on your own computer. To use more compute, click a button to switch from your local machine to your preferred cloud provider.
We handle the DevOps under the hood so you can focus on what matters: research e.g. a CIFAR speedrun!
1
1
13
Never knew I needed this in my life till now, but behold the three most based and most recent Fed chairs:
234
2K
9K
On the history of macroeconomic thought:
(Link to the actual post in the next tweet)
4
106
402
Want to be Medlock-pilled? Don't know where to start? Presenting "Mr Medlock and the Classics": .
Or: "Trevor has nothing better to do except collate 6 months of
@jdcmedlock
tweets +
@LittleKeegs0
/
@DavidSligar
/
@MattBruenig
content into a socdem summary"
15
42
271
case in point: i transferred from cambridge to stanford 2 years ago, and lo and behold, what am i up to now?
yc s24!
Stanford and Berkeley students start companies at 5x the rate of Oxford and Cambridge.
Why the smartest, most technical people in the UK should take more risk 🚀
96
196
1K
18
3
207
the dwarkesh podcast taking over the lightcone is strikingly plausible
it doesn't require believing in sci-fi; it just requires believing in straight lines on a graph
you can be part of making it happen!
I've only gotten 15 applications so far.
If you're really good, please apply!
I'm looking for a partner/COO.
Every week, we will publish worldview-shaking insights of (formerly) obscure intellectuals, reaching an audience that often touches into the millions and which grows
40
53
484
4
8
192
3
0
103
what happened to political philosophy? why aren't there any compelling visions of the future any more?
If there were, then people might not have got so caught up in crypto nonsense and maga land.
35
10
90
1
6
96
1/3 Today, I'm excited to show you what
@leilavclark
and I have been up to. We've been working on
@moonglow_ai
: Colab on your cloud!
One problem I faced when doing ML research was that moving my experiments from local notebooks to cloud GPUs.
Moonglow fixes this.
8
13
83
@lefttailguy
@ipsumkyle
@rexjetter
it can go lower (see landec as the classic example), but the bigger thing i was gesturing at is that taking your a-levels comes after you do interviews and get your offer, not before
8
0
80
@nominalthoughts
interesting - my experience at both Cambridge and Stanford is that everyone is super low key about this!
2
0
72
💥“what causes polysemanticity?" was accepted to re-align/bgpt @ iclr!💥
tl;dr: with more features than neurons in a model, each neuron has to represent many features -> interpreting that is hard
but even with enough neurons, you may still get one neuron for many features!
4
8
62
@JosephPolitano
@balajis
@jdcmedlock
it's not just bad in expectation, it's that there's a pure (excluding counterparty risks) arb on the other side
3
0
60
Updated
@open_phil
plans for supporting
@GiveWell
's recs this year:
Going to $350M, up $50M from last yr, but less than we'd planned, because GW found better opportunities than we expected but our assets are down ~40%. GW recs now ~20% more cost-effective.
1
14
95
1
3
58
i think an aspect of boarding school i really underrated while i was there (but have noticed since going to uni) as to why i enjoyed it is that knowing you'll be living with the same 14 people for 5 years means there is common knowledge that everyone is playing a repeated game
3
1
59
this!
every time we press pause, we burn through the lead that major labs have, which makes the race after the pause more competitive
thus there are many moments (including right now) where a pause may be actively counterproductive (even ignoring credibility re crying wolf)!
Fwiw, I think this is a bad idea.
Models aren’t actually dangerous yet. Risks “crying wolf.” Keep the powder dry for if/when we face real xrisk.
32
15
272
9
4
52
what should training compute be spent on?
4 years ago: mostly parameters
2 years ago: an even split of parameters and data
now: mostly on data
here's a brief history of how pretraining recipes have changed, and why!
(link in qt)
1
9
51
@PradyuPrasad
suppose each morning, there's new work
at every level of the hierarchy, someone needs to make a decision
thus by the time it arrives at an analyst's desk, it is quite late in the day
since there is no marginal cost to your boss from you getting it done overnight, you stay late
1
0
41
@DavidDuvenaud
at least for me:
@shakoistsLog
I tend to interpret ranges of probabilities as bid-ask spreads on a binary event contract, and it’s what I mean when I give them
a bit awkward here because of issues re the marginal utility of money in the case where it pays out, but normally I think this is a useful construct
0
0
30
2
0
39
r/be
Tell me a decision that changed the trajectory of your life but you CAN'T say choosing your faith, getting married, or having a kid.
3K
565
6K
1
0
37
when your co-founder is so horrified by your git hygiene she makes a whole blog post about it:
Why shouldn't you commit large files to Git?
When working on
@moonglow_ai
, I learned that many people use Git as a one-off way to transfer files remotely.
But this is bad practice! To explain why, I wrote a short blog post that explains how Git works. Link below!
2
1
12
2
1
39
people underrate how nice the sun is!
could not agree more
stanford has created a culture of intense — almost delusional — optimism on campus, where students’ default mindset is that they can go achieve whatever they want and live their dreams... then they go actually do it
plus it's fun and sunny
9
98
2K
0
0
33
Delighted to announce that I'm part of the 13th cohort of Emergent Ventures grant winners! Many thanks go to
@TejasReal
and
@leopoldasch
for advice on EV and
@tylercowen
for picking me! And do check out the other incredibly talented people in this cohort:
7
2
32
2/🧵 in our new post,
@basilhalperin
,
@zmazlish
and i argue that one way of forecasting ai timelines is by looking at asset prices
7
5
30
does anyone actually have empirical data on "a non-trivial percentage of ppl who have SWE jobs cannot solve fizzbuzz"?
4
1
31
But in all seriousness, I remember not being entirely clear on what some of the socdem ideas were and why they were being advocated for when I first joined twitter, so hopefully this annotated compilation of links is a helpful starting point for anyone who is interested!
2
1
27
oh ffs - this was for the scholarship exam, and it is an entirely reasonable thing to ask when evaluating if prospective students can write well and pass ideological turing tests
35% of UK prime ministers have been educated at one school: Eton.
This was an entrance exam question set by the school in 2011.
744
7K
38K
4
2
31
Leopold Aschenbrenner’s SITUATIONAL AWARENESS predicts we are on course for Artificial General Intelligence (AGI) by 2027, followed by superintelligence shortly thereafter, posing transformative opportunities and risks.
This is an excellent and important read :
0
415
3K
0
1
31
@shakoistsLog
I tend to interpret ranges of probabilities as bid-ask spreads on a binary event contract, and it’s what I mean when I give them
a bit awkward here because of issues re the marginal utility of money in the case where it pays out, but normally I think this is a useful construct
0
0
30
1/🧵 the arrival of transformative artificial intelligence could be one of the most important things to happen, ever.
that's why some of the smartest people i know are working on making sure this goes well!
but when is this going to happen and how can we estimate this timeline?
1
3
29
@besttrousers
probably not tbh - basically all publicity is bad, and this is just as likely to reflect badly as "all EAs are just doing it as a bit"
1
0
28
Yeah, it's crazy long at 14000 words: .
My contention: "we have advanced beyond Wicksell in two respects only: first, we have a more secure grasp on the quantitative magnitudes involved; second, we have gone one layer of microfoundations beyond Wicksell".
1
2
26
NKs: “there’s no financial constraint on government spending. It’s real resources that will prove binding.”
*inflation occurs amid crunch on real resources, as opposed to financing*
Critics: “Aha. NK fails its first big test. MMT succeeds.”
????
MMTers: “there’s no financial constraint on government spending. It’s real resources that will prove binding.”
*inflation occurs amid crunch on real resources, as opposed to financing*
Critics: “Aha. MMT fails its first big test.”
????
37
159
1K
1
4
27
fwiw this is one of my gripes with how social groups are stereotyped at school/uni wrt "jocks vs nerds"
not saying there's no relationship there, but basically all the top tier athletes i know are also very smart!
@Lan_Dao_
Sweden keeps IQ records on all their professions.
Not only are soccer players really smart, it turns out that the better the player, the more likely he is to be smart
16
9
187
2
0
27
i love when people make
@neelnanda5
-style walkthroughs + there's a loose end from our paper i've been meaning to deal with: i decided to record a short one of me extending one of our plots!
(+ it gave me a chance to stress test some new
@moonglow_ai
features we've added)
💥“what causes polysemanticity?" was accepted to re-align/bgpt @ iclr!💥
tl;dr: with more features than neurons in a model, each neuron has to represent many features -> interpreting that is hard
but even with enough neurons, you may still get one neuron for many features!
4
8
62
0
2
28
0
2
28
if you “leave [a] top tier university”, I highly doubt that doing a startup is anywhere close to the risk-reward Pareto frontier
The biggest problem of venture capital right now is we have made entrepreneurship too safe a career path.
Starting a company should be for the crazy, for the obsessed.
Instead, it is for the rational.
Leave top tier university, raise VC. Own your own time, work on something
345
182
2K
2
0
27
for stuff that's going to be big, you're almost always early even if you think you're late (because it's going to keep growing really fast)
this is mostly inspired by leopold's comment on NVDA, but also because i thought i was late on taylor in 2011 and sabrina in 2015
1
0
24
incredibly cool that there's a JEP article on how Open Phil does global health and wellbeing cause prioritisation!
0
2
24
this post mostly makes me want every high schooler to read cal newport's books/blog posts on college admissions, because it is really unfortunate if people continue to see grinding extra-curriculars as the only path into top colleges
i finally have a post about something topical!
math team, and other horrible things you do to get into stanford
85
140
2K
3
0
24
it is hard to quantify the value of marketmakers, since historical comparisons are marred by all sorts of secular trends
but on may 3rd 2023, there was a partial outage on a European exchange which meant MMs couldn’t quote on certain stocks, providing a nice natural experiment
2
1
24
Suppose a small chance time travel is possible. If we can alter the past, we can make sure all mistakes are null and void. This seems really good. And if it got into the wrong hands, this seems really (see: existentially) bad. Ergo, a new cause area!
/s
8
0
22
often ppl say we should swap [insert math class] with stats; my guess is that if we did it, we would acc get more confused ppl in matt's replies, who cargo-culted thru school by repeating "outlier" and "wrong average", forgetting whether it is the mean/median that is misleading
"63% of Americans do not have $500 in the bank to pay for an emergency healthcare bill."
This is inaccurate.
Median household net worth is $192k, including $8k in checking accounts.
321
270
3K
3
0
23
@joodaloop
eh in part, but at least in the us, children of immigrants don't beat out children of natives in earnings if they live in the same area i.e. the dominant force for why children of immigrants do better is because immigrants move to higher opportunity areas than natives
3
0
20
A concept I really like comes from this post: for every field, there are dull intro classes and niche specialised classes. But between those are "essential classes [which] nail the 80/20 of their subject", and they give you the mindset of the field!
1
2
20
i love waymo!
in spite of taking longer and being more expensive than an uber, the experience is simply too magical and seamless such that it's worth every penny
1
1
22
some people think LLMs are stupid because they feel they reason differently to LLMs. idk if this means i'm uniquely introspective or uniquely stupid, but my thoughts often feel very LLM-esque e.g. lots of stochastic parroting, big gains from 0-shot to ICL, finding CoT useful etc.
0
1
22
my experience is that my density/pace of learning is highest in internships rather than uni - why don't more places do the uwaterloo co-op model?
4
0
20
1/🧵
Index funds! Love 'em or hate 'em, you just can't ignore 'em. Today, more assets are held by passive funds than by active funds in the US!
Some have called them "socialist". If so, they're socialist both in the best and worst sense of the word!
1
2
17
@AgustinLebron3
i've often wondered if an intro stats course taught based on simulation and bootstraps would do significantly better than current ones, which often feel a bit like memorising a litany of tests
1
0
17
to be clear, garden leaves are exempt, so this isn't nearly as disruptive to things like quant finance as one might expect at first glance
“New noncompetes banned for all workers as of effective date.”
The FTC just issued a complete ban on all employee non-competes. No carveouts, no limitations, no exceptions.
Wow!
38
419
3K
2
1
18
this is absolutely obscene and also deeply unnecessary - just get a college student to read your essays and work through some past SAT exams!
total cost: $100
I don’t shock easily but these stories are way beyond what I had ever heard. Why shouldn't colleges ask applicants to report on all coaching expenditures above $5,000 and then revoke admission for anyone caught lying?
via
@wealth
123
161
717
3
0
17
upside of YC: we got $200k AWS credits for free to run ML experiments with
downside of YC: i have to figure out how to get AWS set up 😭
no more!
you can now run your Jupyter notebooks using
@awscloud
compute without leaving your IDE via
@moonglow_ai
- hmu to get set up!
been working a lot on making
@moonglow_ai
better recently...
most excitingly, we now support
@awscloud
! if you've got AWS credits to spend and want to easily run your notebooks on some of their GPUs straight from VSCode, let me know - we'd be excited to get you set up!
2
3
13
0
0
16
IMO for some model-dependent ε:
ρ > 0 implies neartermist EA,
ρ ≈ 0 & EV < ε implies progress studies,
ρ = 0 & EV > ε implies longtermist EA.
ρ is the discount rate and EV = U(S)*[P(S|try) - P(S|¬try)] is the utility of survival x the change in chance of survival if we try.
Some thoughts on Effective Altruism (EA) and Progress Studies (PS) as both are influential and generally related to Silicon Valley's increased influence on the discourse 1/n
3
19
219
2
1
16
spending two years away has made me forget how much I love London as a city, but an hour back and I’m reminded of why it feels so much like home!
1
0
16
1/🧵 Today is the 15 year anniversary from when Bear Stearns bailed out two of its hedge funds, marking one of the first publicly visible signs of the GFC!
In the midst of crypto's 2008, what exactly is going on? And how does it compare to the GFC?
1
2
16
yup!
but i’m not going into quant trading for the money
i’m doing it to improve market liquidity and narrow bid-ask spreads!
0
0
17
as Stigler used to say: “if you never miss a plane, you’re spending too much time at the airport”
good to know that I don’t spend too much time at the airport!
1
0
17
I might write sth longer about transferring at some point, but as I’m finishing my first year at stanford, I can pretty confidently say that my experience of changing unis very much fits this i.e. my median day here is at least as good as my +1 stdev day when I was at cambridge
4
0
16
@EAheadlines
the last 18 months (this included) have really convinced me that more people should take this sort of reasoning seriously:
2
1
17
fun fact!
stanford requires all ugrads to do 18 gen-ed classes, broken down into:
- 4 STEM
- 14 non-STEM (7 humanities/social sciences + 2 writing + 2 civic education + 3 foreign language)
tl;dr: this sucks (e.g. ppl aren't engaged in mandatory ones, grad ones don't count etc.)
3
0
14
Supposedly there is an NFT market for this tweet: , so go and make your offers and I'll sell it at some point!
0
0
15
vote
@besttrousers
!
(2) Apricitas founder and Economics Data Guy
@JosephPolitano
vs
(1) guy with the coolest pants
@besttrousers
23
2
58
3
1
15
🎉 we're rolling out native support for A40s and A100s on
@moonglow_ai
today 🎉
to start with, we're giving out ~5 hours of A40 or ~1 hour of A100 compute for free!
you can now try us without bringing your cloud, and with cold start times of ~5 seconds, you can get going ASAP
1
2
16
last final over; officially done with my first year at stanford!
2
0
16
@JosephPolitano
@balajis
@jdcmedlock
yeah that too - but from the pov of someone who wants to take the other side of the bet, i’d just think it’s insane that this is a risk free return of 1900% in 90 days
0
0
15
and he isn’t subtle about it!
like the first line of the second paragraph is literally: “This, however, is emphatically not the economic problem which society faces.”
Hayek: All the information you'd need to start planning an economy cannot be gathered into one place to even begin to be processed.
Like half the people who comment on Hayek, for some reason: What if we had a really big CPU?
37
93
662
1
1
16
man, that's disappointing
i would have thought stanford students are familiar enough with MOSS not to pull this sort of stuff
Llama-3-V Github and HF both down after evidence of them stealing
@OpenBMB
's model is out.
Sorry bros I don't think we should let it be a closed chapter in your lives, to venture forth to new 10x engineering grifts.
@mustafaaljadery
,
@AkshGarg03
,
@siddrrsh
– speak up please.
18
68
688
0
0
15
an econ instinct which i think is probably correct and underrated in how much it shapes people's thinking is: the relative scarcity of corner solutions i.e. by default, we assume that we get an interior solution and that the optimal amount of x is rarely 0
3
0
15
self-recommending!
Virtually nobody is pricing in what's coming in AI.
I wrote an essay series on the AGI strategic picture: from the trendlines in deep learning and counting the OOMs, to the international situation and The Project.
SITUATIONAL AWARENESS: The Decade Ahead
278
918
4K
3
0
15
i was lucky enough to intern with the cause prio team last summer, and it was fab!
everyone at OP is incredibly kind and smart (shoutout esp to my manager
@chris_topian
), and the work you'll do is intellectually interesting and can have a positive impact on global wellbeing!
1
0
14