Profile picture
François Chollet
@fchollet
, 3 tweets, 1 min read Read on Twitter
In math, the more experienced you are, the more comfortable you get with using shortcut notations and omitting obvious details.

Students like to use verbose, explicit notation, and feel lost when they encounter shortcuts -- thinking they're missing important details.
This reminds me of a pattern I see a lot in grad students in ML -- those least comfortable with abstracting away boilerplate are usually those least familiar with what the boilerplate represents.
I mean, can I really trust a library to compute my gradients? I don't know what's going on behind the scenes... 🤔
Missing some Tweet in this thread?
You can try to force a refresh.

Like this thread? Get email updates or save it to PDF!

Subscribe to François Chollet
Profile picture

Get real-time email alerts when new unrolls are available from this author!

This content may be removed anytime!

Twitter may remove this content at anytime, convert it as a PDF, save and print for later use!

Try unrolling a thread yourself!

how to unroll video

1) Follow Thread Reader App on Twitter so you can easily mention us!

2) Go to a Twitter thread (series of Tweets by the same owner) and mention us with a keyword "unroll" @threadreaderapp unroll

You can practice here first or read more on our help page!

More from @fchollet see all

Profile picture
François Chollet
@fchollet
Machine learning has the potential to make a big difference in solving some of humanity's biggest problems -- making renewables more efficient, optimizing our transportation networks, recycling our trash, making medical care more broadly accessible, accelerating science.
Most of it will take the form of 10-50% improvements to existing processes. It won't always be very sexy, but in aggregate, it will be transformational.

It seems to me we aren't investing nearly enough researcher brain power on this type of impactful, "boring" applied problems.
Meanwhile, what is our community focusing on? What are the media & public focusing on?

Misplaced hype. Fantasies about lofty AGI research that is sure to generate unlimited profits. Fantasies of AI apocalypse peddled by those who need you to be afraid.

It's a bit disheartening.
Read 6 tweets
Profile picture
François Chollet
@fchollet
All content from the first edition of the book is still relevant -- all code examples will run -- except for the appendix about installing software.

I'm working on a 2nd edition that will cover the latest techniques and APIs -- possible release at the end of 2019 or early 2020.
In general, the new APIs *extend* prior Keras workflows rather than upending them -- in particular they make Keras completely flexible and extensible for the needs of deep learning researchers. Prior high-level workflows are still relevant.
As a new user, you won't see a huge difference in the Keras experience in TensorFlow 2.0. But as a researcher, you'll get access to entirely new workflows and features to work more productively.
Read 3 tweets
Profile picture
François Chollet
@fchollet
"Ever bigger steam engines burning ever more coal is the superior approach to build faster trains", they said.

More data + scaling up simple algos will get you better models, obviously. But there are diminishing returns -- meanwhile, new approaches can be transformational
I guarantee you that the dominant AI algorithm of the 2050s will not be a stack of differentiable layers trained with SGD.
At this point we are pretty close to training our language models on the entirety of text corpora that humanity has produced so far. Their language understanding capabilities are still close to zero.
Read 3 tweets

Related threads

Profile picture
K/DA Jim
@nerdsherpa
I saw an article about "Walmart Insulin", which you can get for cheap and how some diabetics without insurance are trying to use that instead of their usual insulin. So I thought I might do a thread about what "Walmart Insulin" is, since I've been diabetic for almost 30 years.
In the olden times, before I got diabetes, they used pig or cow insulin. I know nothing about this dark age. By the time I arrived on the scene, they were making insulin with SCIENCE. They jammed human DNA into E.coli and stuck it in a bioreactor and blammo, humulin.
Humulin is what they're talking about when they say "Walmart Insulin". A type of insulin made from human DNA and bacteria. There are two basic types. R (short acting) and N (long acting). The problem is, Humulin is slow and kind of unpredictable.
Read 19 tweets
Profile picture
John Schindler
@20committee
Co-signed

Not to mention that if this is true (it's almost certainly not, but let's just play along with Cheeto Jesus and his Orange Fantasy World for a moment, shall we?), what Obama told him about DPRK was very, very far from UNCLASSIFIED.

#details
POTUS can declassify and release whatever he wants, but only the mentally unwell would consider telling the world this (if it's true: if) to be wise and helpful to world peace and American national security.
Read 3 tweets
Profile picture
Vitalik Non-giver of Ether
@VitalikButerin
More people should look at using smart contracts for recurring payments on a *per-second* basis for subscription services, ongoing donations, dividends, etc. Can set and change a *payment rate* with a single transaction.
Also employee salaries. If you think about it in a modern cybernetic society where the technological tools to manage flows as first-class objects exist there's no reason for the concept of biweekly or monthly "paydays" to exist.
I recommended @GetGitcoin switch to per-second flows for its grants and CLR matching; would love to see them and others experimenting with things like this.
Read 3 tweets
Profile picture
Steven Dennis
@StevenTDennis
Fun fact: the deficit would be shockingly close to zero today with Clinton-era tax levels.
Back when we had a surplus in 2000, federal revenues were ~20% of GDP.
In 2018, feds spent ~20.3% of GDP
Deficit would have been ~0.3% GDP w/Clinton-2000-era tax $
=*<$100 billion deficit.*
After a succession of tax cuts, revenue has dropped to 16.5% GDP.
Deficit? $779 billion.
2000 arguably an anomaly - top of dotcom boom, etc
But tax cuts as scored by CBO are big part of deficit: Fiscal Cliff deal which extended most of the Bush tax cuts scored as ~$400B deficit for fiscal 2018 alone.
Trump tax law scored as $135B cut in '18; $280B in '19
Read 6 tweets
Profile picture
foone
@Foone
More films should start like Under Siege 2: Dark Territory.
SHOW THE NAMES WHILE THE FUCKING SHUTTLE LAUNCHES!

SURE IT'S FREE NASA FOOTAGE BUT FUCK IT, WHO CARES, IT LOOKS AWESOME!
Then as the intro credits end we get to see a satellite slowly unfolding, in bad 1995 CGI
the most important part of the movie: BEHOLD MY UNLABELED CD-R!
Read 30 tweets
Profile picture
Dani Bostick
@danibostick
Yesterday Sarah Huckabee Sanders called Trump's accusers 'liars.' Here is why that is so dangerous. (Thread)

vanityfair.com/news/2017/10/w…
A very small percentage of sexual violence victims come forward. There are many reasons why: fear of retaliation, character assassination, being portrayed as a liar. 2/
The Trump White House has fortified those fears in victims everywhere by engaging in the very behavior victims fear. 3/
Read 12 tweets

Trending hashtags

#smarticites #Dishonored2 #BeglariadRead #SNS #Hitman #bigdaga #Glacier #bigdata #taekook #ygk #StorageGateway #Dishonored #PolicingInCollapse #opendara #opemdata #dysfunctional #details #worldbuilding #SQS #KinesisFirehose #AWSConfig #IanLivetweetsHisResearch #freedom #YoonminWeek18_D3 #Healing
Follow Us on Twitter!

Did Thread Reader help you today?

Support us! We are indie developers!

This site is made by just three indie developers on a laptop doing marketing, support and development! Read more about the story.

Become a Premium Member ($3.00/month or $30.00/year) and get exclusive features!

Become Premium

Too expensive? Make a small donation by buying us coffee ($5) or help with server cost ($10)

Donate via Paypal Become our Patreon

Thank you for your support!