Deedy Profile picture
Mar 20 8 tweets 2 min read
If you use Purchasing Power Parity to compare income between US and India, $100k/yr = ₹22 lakh.

But using PPP-based comparisons to make real world decisions on where to live is fallacious — it's closer to ₹30-37 lakh, 1.5x that!

🧵 Thread and post!

1/7
Assumptions:
- Expenses for single taxpayers from age 22–30 living in the Bay Area vs Bangalore
- Lifestyle comparisons beyond goods you can buy—such as proximity to family and pollution—also aren't accounted for.
- Savings are treated as absolute, not relative

2/7
Why isn't PPP good enough?

PPP isn't used to compare incomes — its calculated from a weighted basket of goods for GDP comparison.

The weighting and even the goods don't apply to most — iPhones and equivalent cars are still going to be far more expensive in India!

3/7
Many other factors matter for "effective" PPP:
—Stage of life
—Absolute wealth
—Country-specific spending differences
—Similar quality services are priced differently
—Savings should be considered absolute
—Lifestyle differences

Simply dropping a PPP of 22 doesn't work well

4/7
In the US, according to my spending estimates, if you make $100k, you'll save ~$22k a year.

To save an equivalent amount in India, you'd need ₹37 lakh, for an effective PPP of 37.

Post: debarghyadas.com/writes/ppp/

5/7
Why should you equate savings if spending habits are different?

For big-ticket long term spending like ssupporting parents, a house, a US-based college education and international travel, money can be treated as an absolute.

Even if you relax that, I'd say ~₹30 lakh.

6/7
Its worth mentioning that jobs in India also:
- Require longer hours
- Don't increase in salary as much
- Are fewer than the US

This comparison doesn't account for this.

I don't make a recommendation on where you should live—that's more complex and personal.

7/7
Okay I knew there were going to be disagreements with the assumptions: it would help if instead of just pointing out an expense you disagree with, point out an alternate figure.

I cant edit the tweet but I'll update the post with the new assumptions!

• • •

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

Keep Current with Deedy

Deedy 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!

PDF

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 @debarghya_das

Mar 15
What can OpenAI GPT-4 actually do that 3.5 can't?

—Superior reasoning with 50+ page context length
—Next-level coding capabilities
—Vision
—Specific tasks involving long context (taxes)

🧵 Summary of new capabilities the Developer Livestream!
1. Superior reasoning with 50+ page context length

—Summarize articles to sentences with a specific letter
—Find common themes between two blog posts
—Converting blog posts into poems
2. Coding

—Build a Discord GPT bot from scratch
—Correct itself from just error
—Read entire API docs and make changes

e.g. how to use the new API? how to run asyncio in Jupyter?
Read 5 tweets
Mar 15
HUGE NEWS for US immigration today!

Extending H-1B visa grace period from 60 to 180 days was approved at the @WHIAANHPI meeting today and will be announced by USCIS in ~3mos!!

For the 100k+ workers affected by layoffs, this is a huge sigh of relief. Thanks @ajainb!

🧵

1/3
The second recommendation was to grant an EAD and travel documents to all approved I-140 applicants after 5yrs.

If you're confused by this, so was the committee. There was a long discussion and the vote was postponed to the next meeting in June/July. There is hope!

2/3
Recorded livestream of the @WHIAANHPI meeting is here, with important timestamps:

5:57:53—6:06:00: H-1B grace period recommendation
6:06:02—6:28:47: EAD+AP for approved I-140 in 5yrs recommendation
6:59:27—7:14:56: Voting on both policies



3/3
Read 6 tweets
Mar 3
Is OpenAI making money off their ChatGPT API?

YES. They’re making >$2.1 for every $1 in infra cost.

Float16 ops/token (GPT-3) = 2x175B = 350B
A100 FLOPS (F16) (25% util) = 156T/s
Tokens/s = 156T/350B = 446
Tokens/hr = 1.6M
Revenue/hr = $2/Mx1.6M = $3.2/hr
A100 Cost = $1.5/hr
Caveats:
- Assumed no other optimizations
- Inference efficiency may be < 50% peak
- Hardware cost may be further subsidized by Microsoft (I used $12/hr for 8 A100 80GB from LambdaLabs)

Even with no MSFT discount, >11.7% efficiency on public cloud hardware is unit profitable!
Sources:
- FLOPS per token: discuss.huggingface.co/t/understandin…
- A100 spec: nvidia.com/content/dam/en…
- A100 pricing LambdaLabs $12 for 8: lambdalabs.com/service/gpu-cl…
Read 4 tweets
Mar 2
This biggest issue with search with LLMs is that
generative performance != embedding quality

GPT-3 / ChatGPT is impressive for pure generative apps, but OpenAI's embeddings are worse quality + more expensive for use in vector search than

...plain ole info retrieval!

1/8
Traditional search apps use "lexical" methods to retrieve docs — they use keywords and traverse posting lists in an inverted index using simple, basic arithmetic scoring methods.

In literature, BM25 is the most commonly used ranking method.

2/8
In Vector /semantic/neural search, you "encode" documents into long sequences of numbers that represent its meaning, do the same for the query and find the k nearest neighbors.

Modern vector DBs include Pinecone, Weaviate, Redis, Qdrant and Vectara.

3/8
Read 9 tweets
Mar 1
Ever find yourself on a SaaS website with the "Products" and "Solutions" layout.

You scroll through it.

But you still know nothing about the company.

Here's 6 reasons why that happens 🧵

1/7
1. The PR firm making the website knows nothing about the product.

In a game of broken telephone, the Founder may tell marketing to take care of the website, who in turn tells a PR firm who in turn tells their team. The last 2-3 layers may not have even tried the product!

2/7
2. The company is trying to pitch itself as a "category creator" and trying to sell a "vision".

If they simply describe what they do *now*, the total market (TAM) might be un-investable by VC. Therefore, they claim things they don't do while not "really" claiming them.

3/7
Read 7 tweets
Feb 26
This story spans 70 years and is actually ridiculous.

The Salman Khan starring Bollywood movie Hum Dil Chuke Sanam (1999) was based on one of the first documented inter-racial Indian romances between a Romanian and a Bengali in the 1930s Calcutta!

Here's what happened 🧵

1/10
Mircea Eliade was a Romanian author born in 1907 with vivid memories of the WWI as an adolescent. He started writing books in his teens, eventually becoming a professor in Philosophy and Religion at UChicago.

When he was 21, in 1928, he set sail to Calcutta..

2/10
In Calcutta, Mircea studied under Surendranath Dasgupta, a Bengali scholar from Cambridge. In 1930, he fell in love with his then 16yro daughter, Maitreyi Devi—reprehensible at the time—whilst barely speaking the same language.

Maitreyi herself was no ordinary girl...

3/10
Read 11 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

Don't want to be a Premium member but still want to support us?

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

Donate via Paypal

Or Donate anonymously using crypto!

Ethereum

0xfe58350B80634f60Fa6Dc149a72b4DFbc17D341E copy

Bitcoin

3ATGMxNzCUFzxpMCHL5sWSt4DVtS8UqXpi copy

Thank you for your support!

Follow Us on Twitter!

:(