Aditya Chandupatla
@thechandupatla
Followers
274
Following
931
Media
76
Statuses
1,633
Formerly Engineering @bloomberg @coinbase @tesla @disney | @usc alum | Opinions my own
Mars
Joined September 2013
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
Claudia Sheinbaum
• 567752 Tweets
Fauci
• 397924 Tweets
River
• 130693 Tweets
चुनाव आयोग
• 121802 Tweets
#FixTF2
• 120573 Tweets
Farage
• 117916 Tweets
Venom
• 101847 Tweets
Monica
• 67507 Tweets
Jadue
• 65851 Tweets
Enzo Maresca
• 61434 Tweets
#すとぷり8周年
• 57505 Tweets
First Take
• 53857 Tweets
Bolivar
• 36059 Tweets
Botafogo
• 33783 Tweets
Clacton
• 27629 Tweets
Barriga
• 26138 Tweets
Reform UK
• 25785 Tweets
Başın
• 25218 Tweets
The Strongest
• 23481 Tweets
Colo Colo
• 23118 Tweets
Justin Jefferson
• 22968 Tweets
Stephen A
• 21848 Tweets
ヴェノム
• 21226 Tweets
Talleres
• 18264 Tweets
SHORT N
• 17446 Tweets
Berkshire Hathaway
• 17083 Tweets
산소호흡기
• 16757 Tweets
Djokovic
• 16404 Tweets
Larry Allen
• 14072 Tweets
gracie
• 13240 Tweets
NYSE
• 12852 Tweets
ISABELLE MEDALHA RUY ARAUJO
• 11771 Tweets
RESPIRATOR OUT NOW
• 11198 Tweets
Real ID
• 11185 Tweets
Sri Lanka
• 10979 Tweets
Ducati
• 10654 Tweets
@tunguz
Interview processes are designed for scalability. No other alternative has proven to be as effective as the current version.
8
2
35
@GrantSlatton
Overclocked CPU might not always do productive work. It can have several different outcomes: busy waiting, high context switching, etc.
0
1
28
@arpit_bhayani
Using 1-based indexing is not the problem. Not returning an error when 0 is supplied is the problem. 🤷♂️
1
1
26
@thegarrettscott
It would be interesting to see how much of this is because of raw compute power (more GPU's) vs model/algorithm improvements. My bet is on the former.
5
0
15
@version0chiro
Circumference of a circle is greater than diameter. It's actually more efficient use of space.
Not going to argue on aesthetics.
0
0
11
@Duderichy
You can grep for all the occurrences of pip-install’s from your shells history file sorted by time and pip-freeze them. Recreate the env and pip-install.
0
1
11
@arpit_bhayani
A multi-threaded program is only as good as the number of independent tasks which can be parallelised, i.e., don’t access shared/global data
0
0
9
@arpit_bhayani
Consolidation is an emergent property of an industry maturing. The only differentiating factor, I believe, will be tailor-made databases and frameworks, for domain-specific use cases.
0
0
9
GPT-5 should drop in a week (or in 10 days worst case.) Given GPT-4 completed training in mid 2022, GPT-5 would have completed training in mid-to-late 2023.
0
0
2
0
0
7
@AravSrinivas
If successful, the weights this model will get tuned on, will probably apply universally given India’s traffic
0
0
6
@AravSrinivas
Need one for iPhone’s as well. Augment ‘Look Up’, which currently only offers searching the Web via Safari, with perplexity (or any other third party app)
0
0
6
@soumithchintala
Suffixing the prompt with “shouldn’t result in runtime errors” would perhaps have resulted in less prompts
0
0
6
@Akshat_World
Perhaps we should not associate Rich with Educational, Employment or Healthcare.
0
0
5
@ChrisJBakke
I think the hidden variables here are:
1. The percentage cost for the customer (human consumer vs enterprise), i.e., how much of their earnings is being spent on the product.
2. Corresponding upside to the customer (human consumer vs enterprise) if the value, as advertised by
1
0
4
@karpathy
Soon there’ll be custom assembly primitives/instructions designed to run LLM’s on LLM-native hardware chips 🤖
2
1
5
With Neuralink it will be thought-to-action lmao
2
2
5
@unusual_whales
You’ll see a lot more in 2024 as the “potential” rate-cuts have delayed impact on the financial system
2
0
2
@evelynjunn
Technically, it is better to sprint, complete workout early, and then watch course videos at 2X speed.
1
0
5
@unusual_whales
Feels like there is less-and-less incentive for the working-class to contribute to US economy if majority of taxes are being used to repay back the debt. If one has other alternatives to pursue, they should do so now. Period.
0
0
5
@GergelyOrosz
90+% of startup’s built around LLM’s will get consolidated or run out of VC money.
And so will the demand for NVIDIA chips. Although, big players will still keep buying to train next-gen foundational models.
1
0
4
@arpit_bhayani
Whatever we can do with pass-by-value we can do with pass-by-pointer, the latter is essentially a superset
0
0
3
@EcZachly
Access control and lineage is also important. And I don’t think many companies can afford to hire data analyst, data engineer, and data architect as disparate roles. Everything nowadays is clubbed into “software engineer - data platform”
2
0
4
@chiefaioffice
With better chips, training cost will only reduce.
And once model’s performance converges, the cost of training all prior models will be irrelevant.
If open-source wins, then we will end up with only a handful of models (each optimising for a given metric), or a single ensemble
0
0
4
@arpit_bhayani
If the startup was bootstrapped instead of venture-backed, over-engineering will naturally take a back seat
0
0
4
Current state-of-the-art AI is not widely available only because of lack of safeguards and regulations.
We're sharing our learnings from a small-scale preview of Voice Engine, a model which uses text input and a single 15-second audio sample to generate natural-sounding speech that closely resembles the original speaker.
498
1K
5K
0
0
3
@tobi
Pinning posts allows you to override the recommendation algorithm and boost signal strength.
1
0
3
@arpit_bhayani
Data storage pattern should mirror user requests. It doesn’t matter how optimally you store the data, if the request thread waits for data to be available in the requested format.
Another often overlooked concept is how far your servers are located from users.
0
0
3
Perhaps software engineering is the only role which has evolved from writing code in assembly to build compilers, to writing code in C to build operating systems, to writing code in Python to build distributed applications, to finally writing code in natural language to build AI.
0
1
2
0
0
3
@jerryjliu0
Someone should do this for W-2, 1099, and the likes. Corporations like Turbotax, Sprintax and H&R block have outdated methods.
0
0
3