David Ma Profile picture
20 Jul, 6 tweets, 1 min read
This synthesizes what I am looking for in an ML eng hire.

Even with normal models, the way you define the ML problem is 80% of the insights needed to solve a product problem.

What is “defining the ML problem”?

Famous examples:
Product problem: Airbnb ranking listings.

Naive problem definition: trying to predict whether a listing was booked based on a search session.

Woke problem definition: pair a booked listing with a non-booked listing. Predict which of the two was booked
Problem: unsupervised language modeling

Naive: given a prompt, predict the next word

Woke (Bert): mask word in a passage, predict missing word; given two sentences, predict if they followed each other in original context;
Problem: train an AI to play Go

Naive (alphago): value net predicts the winner based on board position

Woke (katago): value net predicts the owner of a specific point on the board by the end of the game based on current position. Result: able to play for points like human.
By the way, none of the “naive” methods mentioned above were “bad”. They were picked to illustrate how reframing the problem can yield better results.
In fact, the naive formulations were genius the first time somebody came up with it.

• • •

Missing some Tweet in this thread? You can try to force a refresh

Keep Current with David Ma

David Ma Profile picture

Stay in touch and get notified when new unrolls are available from this author!

Read all threads

This Thread may be Removed Anytime!


Twitter may remove this content at anytime! Save it as PDF for later use!

Try unrolling a thread yourself!

how to unroll video
  1. Follow @ThreadReaderApp to mention us!

  2. From a Twitter thread mention us with a keyword "unroll"
@threadreaderapp unroll

Practice here first or read more on our help page!

More from @madavidj

16 Oct
Alright my thoughts on AMMs vs Orderbooks since ya'll still debating...
They are different *market microstructure* (mm).

Microstructure is the set of rules that govern how participants interact with each other.

Different mm favor different participants types.

Orderbooks, dark pools, otc, auctions, are all valid tradfi mm with their use cases.
The best market microstructures *must* be well-balanced or else one party is gonna realize and stop trading there.

The chosen rules make liquidity and trade volume possible in the first place.
Read 10 tweets
27 Sep
I get cold emails every couple weeks asking how I got into quant trading and if I have any advice for aspiring quants.

Here are a few thoughts in that direction:
If you’re interested in $, that’s cool. Achieving a safety net for my family was also my first priority after college.

But, the market has since adjusted and simply getting into QF isn’t gonna make u rich. Your expected TC might be a bit higher than big tech.
I think being employed as an alpha researcher in a big HF is good for creds, but bad for learning.

Few teach for free. Most teach the minimum you need to know.

Work is silo’d. Most of what I learned came from personal thinking and reading.

Here’s my recc for learning:
Read 6 tweets
12 Sep
Smart contracts mainly create shared value for the world.

That value created cannot be hoarded by the contract authors. (see Thiel on value capture)

A thread... 1/8
2. The reason is simple: smart contracts are open source.

See: @SushiSwap, @SwerveFinance, @CreamdotFinance, and many others are building on top of the OGs like @UniswapProtocol @compoundfinance @CurveFinance

Whatever value was being captured by contracts can be forked, but...
3. People still figuring out what the moats can be. Team? community? user mindshare? memes?

Even liquidity might not be a moat. As seen with Sushi, "liquidity has no loyalty" @AndreCronjeTech

But even Sushi wont capture rent if it doesn't innovate:
Read 8 tweets
19 Aug
“Ideas are cheap, execution is everything”

My interpretation of this:
A successful startup starts with one high level idea, but requires many follow-up Execution Ideas
Our main idea was a bot for rentals leasing.

But follow-up Execution Ideas that made us who we are include: how we chose the ML tech, how we designed the human-bot UI to work seamlessly together, how we hired our operators, how we “pivoted” on our target market segment, etc
Read 7 tweets
14 Aug
As far as I can tell in 5 mins, this team is trying to leverage GANs to compress video chats so hard that 1) latency can be on the order of frames, and 2) anybody with a internet connection can video chat.

How I think one could achieve this:
App takes enough photos of you to construct a profile.

You exchange profiles with the other person. This may or may not be large. But it’s a one time exchange.

When you video chat, your phone synthesize your expressions into the most compact set of bits to send over.
How many bits do you really need? Perhaps not that many. Your range of facial expressions isn’t that large.

This representation then gets re-hydrated by your pre-exchanged profile to recreate a convincing video chat.
Read 4 tweets
30 Jul
As a recovering quant, I share this cautious sentiment about data driven decisions.

Ways data driven decision can be worse than careful thinking:
Precision is not accuracy

A repeatable and precise measurement of a system may not be the measurement that matters most.
Bias towards the measurable

Tendency to measure what is easy to measure. And ignore what is hard to measure.
Read 8 tweets

Did Thread Reader help you today?

Support us! We are indie developers!

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

Become a Premium Member ($3/month or $30/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!

Follow Us on Twitter!