Sotwe logo Sotwe logo
praneethsamine's profile banner image
praneethsamine's profile image

Praneeth Samineni

Verified account

@praneethsamine

cs & math @georgiatech ‘30 365 days of robot learning 23/365
Atlanta, GA
Joined January 2022
503 Following
215 Followers
192 Posts
praneethsamine's profile image
Praneeth Samineni Verified account @praneethsamine
@kaycidot I've been pretty loyal to claude code remote control but might need to check this out
0
0
0
0
55
praneethsamine's profile image
Praneeth Samineni Verified account @praneethsamine
@AadhavThaniga bro's grinding out animations every day, u using manim?
0
0
0
0
8
praneethsamine's profile image
Praneeth Samineni Verified account @praneethsamine
Link to paper: https://t.co/1m5Jis6Vzt
0
0
0
0
68
praneethsamine's profile image
Praneeth Samineni Verified account @praneethsamine
Day 22/365 of robot learning. Today I finally decided to start reading LoRA, a paper I've been putting off for a while but think is pretty foundational, especially since I want to contribute to pretraining as robotics foundation models continue to scale. LoRA was introduced in 2021 as LLMs began to scale as well, with models like GPT-3 reaching 175B parameters, making full fine-tuning computationally expensive. It builds on the broader idea of parameter-efficient fine-tuning (PEFT), where we adapt models without updating all their pretrained weights. In their GPT-3 experiments, LoRA reduced trainable parameters by up to 10,000x and GPU memory requirements by 3x compared to full fine-tuning! The main idea I grasped so far is to freeze the pretrained weights Φ₀ and learn a much smaller set of parameters Θ to represent the updates: Φ = Φ₀ + ΔΦ(Θ), where |Θ| ≪ |Φ₀|. More specifically, LoRA represents the weight update through two smaller matrices: W = W₀ + BA Where B ∈ ℝ^(d×r) and A ∈ ℝ^(r×k), with r ≪ min(d,k). Instead of optimizing d×k parameters, we're only optimizing r(d+k), based on the paper's hypothesis that weight updates during adaptation have a low intrinsic rank, meaning they can be represented using much smaller matrices. I've been slumped with projects and hw, so I plan to take this weekend to work through this paper in the background, but the robot learning grind doesn't stop!
praneethsamine's tweet photo. Day 22/365 of robot learning. Today I finally decided to start reading LoRA, a paper I've been putting off for a while but think is pretty foundational, especially since I want to contribute to pretraining as robotics foundation models continue to scale. LoRA was introduced in 2021 as LLMs began to scale as well, with models like GPT-3 reaching 175B parameters, making full fine-tuning computationally expensive. It builds on the broader idea of parameter-efficient fine-tuning (PEFT), where we adapt models without updating all their pretrained weights. In their GPT-3 experiments, LoRA reduced trainable parameters by up to 10,000x and GPU memory requirements by 3x compared to full fine-tuning! The main idea I grasped so far is to freeze the pretrained weights Φ₀ and learn a much smaller set of parameters Θ to represent the updates: Φ = Φ₀ + ΔΦ(Θ), where |Θ| ≪ |Φ₀|. More specifically, LoRA represents the weight update through two smaller matrices: W = W₀ + BA Where B ∈ ℝ^(d×r) and A ∈ ℝ^(r×k), with r ≪ min(d,k). Instead of optimizing d×k parameters, we're only optimizing r(d+k), based on the paper's hypothesis that weight updates during adaptation have a low intrinsic rank, meaning they can be represented using much smaller matrices. I've been slumped with projects and hw, so I plan to take this weekend to work through this paper in the background, but the robot learning grind doesn't stop!
praneethsamine's profile image
Praneeth Samineni Verified account @praneethsamine
Day 21/365 of robot learning. I finally finished reading OpenVLA today! Unlike ACT or Diffusion Policy, the paper focused less on developing new algorithms and more on scaling existing architectures, improving data mixtures, and optimizing training, so it was pretty easy to get through. It almost felt like a better application of the Bitter Lesson to robotics, borrowing a lot of the approaches that made LLM scaling successful. I also couldn't help reading bits of π0.7 after watching Sergey Levine's recent conversation with Aditya Agarwal about zero-shot generalization and adapting to new tasks through simple language instructions rather than task specific fine-tuning, which OpenVLA relied on for new setups. Here are a few key takeaways and thoughts on OpenVLA: 1. Scaling robot learning OpenVLA was pretrained on 970K robot demonstrations, carefully filtering and reweighting datasets based on task diversity and performance. Surprisingly, they trained for 27 epochs, unlike the 1-2 epochs typically used for LLM/VLM pretraining. It feels like scaling laws are starting to kick in, but we're still bottlenecked by how much human demonstration data we can collect. This made me wonder whether we even have enough global GDP to collect the data necessary to reach LLM-level generalization in robotics. 2. Why only 224×224 resolution? They found that increasing image resolution to 384×384 didn't improve performance in their experiments and also took 3x longer to train, so they just stuck with 224×224. I guess this worked for the robot manipulation tasks within their experiments, but I couldn't help but wonder what happens when a task depends on small/distant visual details? For example, what if a robot were instructed to shoot a Nerf gun at a faraway target? dw I'm not trying to build a terminator. 3. LoRA and generalization OpenVLA was the only approach to achieve at least 50% success across all tested downstream tasks in their fine-tuning experiments, making it a strong general-purpose option. However, Diffusion Policy still produced smoother and more precise trajectories for narrower but highly dexterous tasks. They also found that LoRA nearly matched full fine-tuning while updating only 1.4% of parameters, taking just 10-15 hours on a single A100, an 8x reduction in compute. I realized I should probably read up on LoRA soon after seeing this. 4. A surprising quantization result Their inference results initially confused me: bfloat16: 71.3% success, 16.8GB int8: 58.1% success, 10.2GB int4: 71.9% success, 7.0GB Why was int4 matching bf16 while outperforming int8? It turns out int8 introduced inference overhead that slowed the control loop to 1.2Hz, while int4 reached around 3Hz. The performance difference was largely due to control latency rather than quantization accuracy itself. This is definitely something I'll experiment with when quantizing models for edge inference. Overall, reading OpenVLA kinda confirmed how important scaling is in robotics, and made me more optimistic about how combining generalist robot policies with higher-level agentic systems might actually work out. From reading this paper, I also realized I want to revisit LoRA, FlashAttention, and speculative decoding to better understand large-scale GPU training and inference optimization. I went through some of these ideas while reading through the DeepMind scaling book this past summer, but there's still a lot I want to understand more deeply.
praneethsamine's tweet photo. Day 21/365 of robot learning. I finally finished reading OpenVLA today! Unlike ACT or Diffusion Policy, the paper focused less on developing new algorithms and more on scaling existing architectures, improving data mixtures, and optimizing training, so it was pretty easy to get through. It almost felt like a better application of the Bitter Lesson to robotics, borrowing a lot of the approaches that made LLM scaling successful. I also couldn't help reading bits of π0.7 after watching Sergey Levine's recent conversation with Aditya Agarwal about zero-shot generalization and adapting to new tasks through simple language instructions rather than task specific fine-tuning, which OpenVLA relied on for new setups. Here are a few key takeaways and thoughts on OpenVLA: 1. Scaling robot learning OpenVLA was pretrained on 970K robot demonstrations, carefully filtering and reweighting datasets based on task diversity and performance. Surprisingly, they trained for 27 epochs, unlike the 1-2 epochs typically used for LLM/VLM pretraining. It feels like scaling laws are starting to kick in, but we're still bottlenecked by how much human demonstration data we can collect. This made me wonder whether we even have enough global GDP to collect the data necessary to reach LLM-level generalization in robotics. 2. Why only 224×224 resolution? They found that increasing image resolution to 384×384 didn't improve performance in their experiments and also took 3x longer to train, so they just stuck with 224×224. I guess this worked for the robot manipulation tasks within their experiments, but I couldn't help but wonder what happens when a task depends on small/distant visual details? For example, what if a robot were instructed to shoot a Nerf gun at a faraway target? dw I'm not trying to build a terminator. 3. LoRA and generalization OpenVLA was the only approach to achieve at least 50% success across all tested downstream tasks in their fine-tuning experiments, making it a strong general-purpose option. However, Diffusion Policy still produced smoother and more precise trajectories for narrower but highly dexterous tasks. They also found that LoRA nearly matched full fine-tuning while updating only 1.4% of parameters, taking just 10-15 hours on a single A100, an 8x reduction in compute. I realized I should probably read up on LoRA soon after seeing this. 4. A surprising quantization result Their inference results initially confused me: bfloat16: 71.3% success, 16.8GB int8: 58.1% success, 10.2GB int4: 71.9% success, 7.0GB Why was int4 matching bf16 while outperforming int8? It turns out int8 introduced inference overhead that slowed the control loop to 1.2Hz, while int4 reached around 3Hz. The performance difference was largely due to control latency rather than quantization accuracy itself. This is definitely something I'll experiment with when quantizing models for edge inference. Overall, reading OpenVLA kinda confirmed how important scaling is in robotics, and made me more optimistic about how combining generalist robot policies with higher-level agentic systems might actually work out. From reading this paper, I also realized I want to revisit LoRA, FlashAttention, and speculative decoding to better understand large-scale GPU training and inference optimization. I went through some of these ideas while reading through the DeepMind scaling book this past summer, but there's still a lot I want to understand more deeply.
8
51
1
30
3K
1
11
0
12
646
praneethsamine's profile image
Praneeth Samineni Verified account @praneethsamine
@aliteracy It made my roommate start doing nuerotech research
0
0
0
0
21
praneethsamine's profile image
Praneeth Samineni Verified account @praneethsamine
@aarav_kothari_x I didn’t know snake robotics was a thing
0
0
0
0
15
praneethsamine's profile image
Praneeth Samineni Verified account @praneethsamine
@Cortixofficial You’re right, but that is still bottlenecked by how many people we can record, so it’ll be hard to scale it as much as llms which aren’t even that robust in the first place. I guess potential solutions include rl since that’s bottlenecked by compute
0
1
0
0
21
praneethsamine's profile image
Praneeth Samineni Verified account @praneethsamine
Day 21/365 of robot learning. I finally finished reading OpenVLA today! Unlike ACT or Diffusion Policy, the paper focused less on developing new algorithms and more on scaling existing architectures, improving data mixtures, and optimizing training, so it was pretty easy to get through. It almost felt like a better application of the Bitter Lesson to robotics, borrowing a lot of the approaches that made LLM scaling successful. I also couldn't help reading bits of π0.7 after watching Sergey Levine's recent conversation with Aditya Agarwal about zero-shot generalization and adapting to new tasks through simple language instructions rather than task specific fine-tuning, which OpenVLA relied on for new setups. Here are a few key takeaways and thoughts on OpenVLA: 1. Scaling robot learning OpenVLA was pretrained on 970K robot demonstrations, carefully filtering and reweighting datasets based on task diversity and performance. Surprisingly, they trained for 27 epochs, unlike the 1-2 epochs typically used for LLM/VLM pretraining. It feels like scaling laws are starting to kick in, but we're still bottlenecked by how much human demonstration data we can collect. This made me wonder whether we even have enough global GDP to collect the data necessary to reach LLM-level generalization in robotics. 2. Why only 224×224 resolution? They found that increasing image resolution to 384×384 didn't improve performance in their experiments and also took 3x longer to train, so they just stuck with 224×224. I guess this worked for the robot manipulation tasks within their experiments, but I couldn't help but wonder what happens when a task depends on small/distant visual details? For example, what if a robot were instructed to shoot a Nerf gun at a faraway target? dw I'm not trying to build a terminator. 3. LoRA and generalization OpenVLA was the only approach to achieve at least 50% success across all tested downstream tasks in their fine-tuning experiments, making it a strong general-purpose option. However, Diffusion Policy still produced smoother and more precise trajectories for narrower but highly dexterous tasks. They also found that LoRA nearly matched full fine-tuning while updating only 1.4% of parameters, taking just 10-15 hours on a single A100, an 8x reduction in compute. I realized I should probably read up on LoRA soon after seeing this. 4. A surprising quantization result Their inference results initially confused me: bfloat16: 71.3% success, 16.8GB int8: 58.1% success, 10.2GB int4: 71.9% success, 7.0GB Why was int4 matching bf16 while outperforming int8? It turns out int8 introduced inference overhead that slowed the control loop to 1.2Hz, while int4 reached around 3Hz. The performance difference was largely due to control latency rather than quantization accuracy itself. This is definitely something I'll experiment with when quantizing models for edge inference. Overall, reading OpenVLA kinda confirmed how important scaling is in robotics, and made me more optimistic about how combining generalist robot policies with higher-level agentic systems might actually work out. From reading this paper, I also realized I want to revisit LoRA, FlashAttention, and speculative decoding to better understand large-scale GPU training and inference optimization. I went through some of these ideas while reading through the DeepMind scaling book this past summer, but there's still a lot I want to understand more deeply.
praneethsamine's tweet photo. Day 21/365 of robot learning. I finally finished reading OpenVLA today! Unlike ACT or Diffusion Policy, the paper focused less on developing new algorithms and more on scaling existing architectures, improving data mixtures, and optimizing training, so it was pretty easy to get through. It almost felt like a better application of the Bitter Lesson to robotics, borrowing a lot of the approaches that made LLM scaling successful. I also couldn't help reading bits of π0.7 after watching Sergey Levine's recent conversation with Aditya Agarwal about zero-shot generalization and adapting to new tasks through simple language instructions rather than task specific fine-tuning, which OpenVLA relied on for new setups. Here are a few key takeaways and thoughts on OpenVLA: 1. Scaling robot learning OpenVLA was pretrained on 970K robot demonstrations, carefully filtering and reweighting datasets based on task diversity and performance. Surprisingly, they trained for 27 epochs, unlike the 1-2 epochs typically used for LLM/VLM pretraining. It feels like scaling laws are starting to kick in, but we're still bottlenecked by how much human demonstration data we can collect. This made me wonder whether we even have enough global GDP to collect the data necessary to reach LLM-level generalization in robotics. 2. Why only 224×224 resolution? They found that increasing image resolution to 384×384 didn't improve performance in their experiments and also took 3x longer to train, so they just stuck with 224×224. I guess this worked for the robot manipulation tasks within their experiments, but I couldn't help but wonder what happens when a task depends on small/distant visual details? For example, what if a robot were instructed to shoot a Nerf gun at a faraway target? dw I'm not trying to build a terminator. 3. LoRA and generalization OpenVLA was the only approach to achieve at least 50% success across all tested downstream tasks in their fine-tuning experiments, making it a strong general-purpose option. However, Diffusion Policy still produced smoother and more precise trajectories for narrower but highly dexterous tasks. They also found that LoRA nearly matched full fine-tuning while updating only 1.4% of parameters, taking just 10-15 hours on a single A100, an 8x reduction in compute. I realized I should probably read up on LoRA soon after seeing this. 4. A surprising quantization result Their inference results initially confused me: bfloat16: 71.3% success, 16.8GB int8: 58.1% success, 10.2GB int4: 71.9% success, 7.0GB Why was int4 matching bf16 while outperforming int8? It turns out int8 introduced inference overhead that slowed the control loop to 1.2Hz, while int4 reached around 3Hz. The performance difference was largely due to control latency rather than quantization accuracy itself. This is definitely something I'll experiment with when quantizing models for edge inference. Overall, reading OpenVLA kinda confirmed how important scaling is in robotics, and made me more optimistic about how combining generalist robot policies with higher-level agentic systems might actually work out. From reading this paper, I also realized I want to revisit LoRA, FlashAttention, and speculative decoding to better understand large-scale GPU training and inference optimization. I went through some of these ideas while reading through the DeepMind scaling book this past summer, but there's still a lot I want to understand more deeply.
praneethsamine's profile image
Praneeth Samineni Verified account @praneethsamine
Day 20/365 of robot learning. I spent today experimenting with @dreamlabs_ai Dream-1 model on my so101 arm, trying to get it to complete a simple pick-and-place task without any fine tuning. Since I'm running a single arm on my macbook instead of the typical two-arm setup the model was designed around, I had to configure everything manually. After 20+ runs, a few changes somewhat improved its performance: -Experimented with mapping my so101 as the model's left arm instead of the right, which lead to fewer hallucinations. -Checked the gripper's tcp with dreamcli tcp; it needs to sit right at the fingertips (moving it inward into the gripper made things worse). -Repositioned my top and wrist cameras to better match the YAM setup the model was trained on. -Added table collision constraints and joint limits to prevent the arm from pressing into the table. In my best run so far, the arm reached my deodorant, but the gripper closed at the tip of it. This was the same issue I encountered in sim funnily enough. It worked there, but not in real life. It still hasn't completed the full pick-and-place task, but it's definitely making some progress! Also, huge thanks to @NamraByTheSea from Dream Labs for reaching out and helping me debug the setup! @dreamlabs hoping I can get some extra credits since the API wasn't working with my single arm setup :) Going to get back to reading papers soon and try out @nvidia 's newly open-sourced ASPIRE framework to experiment with agentic robot control!
10
72
4
27
7K
8
51
1
30
3K
praneethsamine's profile image
Praneeth Samineni Verified account @praneethsamine
@TSC @dreamlabs_ai Thanks, I’ll definitely check that out!
0
1
0
0
21
praneethsamine's profile image
Praneeth Samineni Verified account @praneethsamine
Day 20/365 of robot learning. I spent today experimenting with @dreamlabs_ai Dream-1 model on my so101 arm, trying to get it to complete a simple pick-and-place task without any fine tuning. Since I'm running a single arm on my macbook instead of the typical two-arm setup the model was designed around, I had to configure everything manually. After 20+ runs, a few changes somewhat improved its performance: -Experimented with mapping my so101 as the model's left arm instead of the right, which lead to fewer hallucinations. -Checked the gripper's tcp with dreamcli tcp; it needs to sit right at the fingertips (moving it inward into the gripper made things worse). -Repositioned my top and wrist cameras to better match the YAM setup the model was trained on. -Added table collision constraints and joint limits to prevent the arm from pressing into the table. In my best run so far, the arm reached my deodorant, but the gripper closed at the tip of it. This was the same issue I encountered in sim funnily enough. It worked there, but not in real life. It still hasn't completed the full pick-and-place task, but it's definitely making some progress! Also, huge thanks to @NamraByTheSea from Dream Labs for reaching out and helping me debug the setup! @dreamlabs hoping I can get some extra credits since the API wasn't working with my single arm setup :) Going to get back to reading papers soon and try out @nvidia 's newly open-sourced ASPIRE framework to experiment with agentic robot control!
praneethsamine's profile image
Praneeth Samineni Verified account @praneethsamine
Day 19/365 of robot learning. Messing around with @dreamlabs_ai api today. I'm going to try getting it running on my so-101 robot arms soon
2
17
1
1
5K
10
72
4
27
7K
praneethsamine's profile image
Praneeth Samineni Verified account @praneethsamine
@karthikvetrive3 @dreamlabs_ai I followed the regular setup guide to run it in mujoco. But you can also check out the sandbox, they already have everything configured for you
0
0
0
0
12
praneethsamine's profile image
Praneeth Samineni Verified account @praneethsamine
@IsaacSin12 @Ryan_Resolution @makermodsai @dreamlabs_ai Yoo this is nice
1
2
0
0
90
praneethsamine's profile image
Praneeth Samineni Verified account @praneethsamine
Wasn't industry supposed to scale the main line of theory while academia explored the branches? But robotics is taking way too many branches even in industry 😭 Is this what happens when academia leads industry or just a sign of rapid innovation? ig the architecture wars aren't ending anytime soon
0
8
0
1
816
praneethsamine's profile image
Praneeth Samineni Verified account @praneethsamine
paper links: openvla: https://t.co/05vUn0NlA1 pi0.7: https://t.co/Axytbn8nlL
0
0
0
0
234
praneethsamine  retweeted
makermodsai's profile image
MakerMods Robotics @makermodsai
Dream-1 with the Maker and Metal Arm This is the best combo to try out zero shot tasks!
1
15
2
0
819
praneethsamine's profile image
Praneeth Samineni Verified account @praneethsamine
@adityaag Agreed, this is helpful to hear. I've been debating whether to pursue a phd after undergrad or go all in on applying and scaling current robot learning approaches through industry research
0
1
0
0
600
praneethsamine  retweeted
AneeshDan's profile image
Aneesh Dantuluri @AneeshDan
We won 2nd place @ycombinator's Make Something Agents Want Hackathon in the @insforge track! We built Cortex: Cursor for training robot arms autonomously. After facing the tedious process of training a policy from scratch at @HackMIT last month, we instantly got to work on finding a faster method. Cortex lets agents control the whole pipeline, from data collection to deployment. Big thanks to my teamates @praneethsamine and @suhasdev23 for building with me. Also, thank you to @nemild, @hanghuang_, and @tonychang430 for organizing the hackathon and track.
AneeshDan's tweet photo. We won 2nd place @ycombinator's Make Something Agents Want Hackathon in the @insforge track! We built Cortex: Cursor for training robot arms autonomously. After facing the tedious process of training a policy from scratch at @HackMIT last month, we instantly got to work on finding a faster method. Cortex lets agents control the whole pipeline, from data collection to deployment. Big thanks to my teamates @praneethsamine and @suhasdev23 for building with me. Also, thank you to @nemild, @hanghuang_, and @tonychang430 for organizing the hackathon and track.
2
9
1
1
196
praneethsamine's profile image
Praneeth Samineni Verified account @praneethsamine
@AneeshDan @ycombinator @insforge @HackMIT Let’s gooo
0
0
0
0
19
praneethsamine's profile image
Praneeth Samineni Verified account @praneethsamine
@SaidevDhal @dreamlabs_ai Yup, will check out!
0
1
0
0
24
praneethsamine's profile image
Praneeth Samineni Verified account @praneethsamine
@DsouzaJovian @dreamlabs_ai good luck!
0
1
0
0
29
praneethsamine's profile image
Praneeth Samineni Verified account @praneethsamine
@DanielLenton1 @dreamlabs_ai Thanks!
0
0
0
0
35

Last Seen Users on Sotwe

FiKradangn97201's profile image
กระดังงาลนไฟ(ช่องหลักนะคะ)
Seen from Thailand
eaftey01's profile image
Eaftey01
Seen from Turkey
fenixwryght's profile image
Serpifeu
Seen from Spain
Fatima0abdulla's profile image
Fatima abdulla
Wwnn_w's profile image
Wwna_na_ja_ 🦋
Seen from Thailand
DiscplsofDesire's profile image
Disciples of Desire
Seen from United States
1percentaeon's profile image
1percentaeon 💹🧲 Verified account
Seen from United States
girlfucki's profile image
Girl fucking for thy money Verified account
Seen from India
utara_anak99158's profile image
couple ds
Seen from Malaysia
piercednipples0's profile image
Beautiful Piercings
Seen from Turkey

Trends for you

1
#WorldSeries
Under 10K tweets
2
Ethan
Under 10K tweets
3
Haiti
Under 10K tweets
4
Hila
Under 10K tweets
5
Luka
Under 10K tweets
6
#CulturalCreditMatters
Under 10K tweets
7
Amen Thompson
Under 10K tweets
8
Bioware
Under 10K tweets
9
LeBron
Under 10K tweets
10
#TheLastDriveIn
Under 10K tweets

Most Popular Users

elonmusk's profile image
1
Elon Musk Verified account
@elonmusk
241.5M followers
barackobama's profile image
2
Barack Obama Verified account
@barackobama
118.8M followers
cristiano's profile image
3
Cristiano Ronaldo Verified account
@cristiano
114.5M followers
realdonaldtrump's profile image
4
Donald J. Trump Verified account
@realdonaldtrump
111.8M followers
narendramodi's profile image
5
Narendra Modi Verified account
@narendramodi
107.1M followers
rihanna's profile image
6
Rihanna Verified account
@rihanna
98.7M followers
nasa's profile image
7
NASA Verified account
@nasa
92.3M followers
justinbieber's profile image
8
Justin Bieber Verified account
@justinbieber
91.8M followers
katyperry's profile image
9
KATY PERRY Verified account
@katyperry
90M followers
taylorswift13's profile image
10
Taylor Swift Verified account
@taylorswift13
84M followers
ladygaga's profile image
11
Lady Gaga Verified account
@ladygaga
75.5M followers
imvkohli's profile image
12
Virat Kohli Verified account
@imvkohli
73.5M followers
kimkardashian's profile image
13
Kim Kardashian Verified account
@kimkardashian
70.9M followers
youtube's profile image
14
YouTube Verified account
@youtube
68.8M followers
neymarjr's profile image
15
Neymar Jr Verified account
@neymarjr
66.5M followers
billgates's profile image
16
Bill Gates Verified account
@billgates
65.3M followers
selenagomez's profile image
17
Selena Gomez Verified account
@selenagomez
63.1M followers
theellenshow's profile image
18
The Ellen Show
@theellenshow
62.2M followers
cnn's profile image
19
CNN Verified account
@cnn
61.7M followers
x's profile image
20
X Verified account
@x
60.7M followers