Tejas Kulkarni @tejasdkulkarni Twitter profile

Last Seen Profiles

@AlexanderDBorg

@Crain1Art

@uruuuu__i

@nutridairy

@coc_space

@GroupGraves

@IlesanmiRushdah

@ShakiraKingdom

@_Su_ceren

@kaitlynbristowe

@cornhill5

@llwy416804

@KevinElevinguy

@mon_ti_ra

@Rkkhattak01

@maria_flyn49123

@BsamAlasd

@dorindadeadly

@JeanYip3

@schadenfreude74

@AvengerReturns

@ronniewood

@kiimhynjin

@WeskusTourism

@RIConferences

@EGLLSpotter

@URMC_Neurology

@SPICE41town

@LumidaWealth

@FiorottoBeatriz

@capnaux

@OHGirlWrestling

@TeresaYaw

@AnujSinghBKU

@nemumi_suu

@Garrett_Dorfman

Tejas Kulkarni

@tejasdkulkarni

5 years

depth from unconstrained video (unknown camera parameters). the results look really impressive!

10

212

978

Tejas Kulkarni

@tejasdkulkarni

4 years

Created AGI this weekend

44

4

787

Tejas Kulkarni

@tejasdkulkarni

4 years

#BostonProtests This is the view from our apartment. Never seen anything like this. The crowd size is unbelievable (20 minutes of real time - )

13

226

444

Tejas Kulkarni

@tejasdkulkarni

5 years

Our work shows that adding geometric inductive biases in neural nets enables spatio-temporally consistent (hundreds of steps) object keypoints. This enables agents that play Atari games on a single machine with less than 100k steps + deeply explore hard envs without rewards.

Google DeepMind

@GoogleDeepMind

5 years

Deep RL agents are data hungry and often learn task-specific representations. Our model learns object-centric abstractions from raw videos. This enables highly data-efficient RL and structured exploration.

8

315

1K

4

64

350

Tejas Kulkarni

@tejasdkulkarni

2 years

The rumors are true - AI energy in the valley is on fire. There is a deep understanding and excitement around building products with AI as one of the core moats. Before it always felt like people used AI-first as a cliche to bloat startups but now there is a genuine realization

11

21

327

Tejas Kulkarni

@tejasdkulkarni

2 years

dall-e2 from @OpenAI is not just an interesting AI tool. IMO its the most compelling demonstration of compositionality due to its multimodal nature. While we need a lot more for AGI (videos,3d,abstractions, behaviors etc), this is the first demo that pulls down my timeline of AGI

13

26

322

Tejas Kulkarni

@tejasdkulkarni

3 years

CNNs are all you need -> attention is all you need -> MLPs are all you need -> compute is all you need -> $ is all you need

Neil Houlsby

@neilhoulsby

3 years

New paper from Brain Zurich and Berlin! We try a conv and attention free vision architecture: MLP-Mixer () Simple is good, so we went as minimalist as possible (just MLPs!) to see whether modern training methods & data is sufficient...

17

218

995

14

24

288

Tejas Kulkarni

@tejasdkulkarni

5 years

We have released the code for transporter — a neural network architecture for unsupervised learning of object keypoints (now a NeurIPS paper):

Google DeepMind

@GoogleDeepMind

5 years

Deep RL agents are data hungry and often learn task-specific representations. Our model learns object-centric abstractions from raw videos. This enables highly data-efficient RL and structured exploration.

8

315

1K

0

55

273

Tejas Kulkarni

@tejasdkulkarni

4 years

Just moved to Boston after spending a few exciting and inspiring years at DeepMind. Excited for the next big adventure.

8

2

256

Tejas Kulkarni

@tejasdkulkarni

1 year

@ylecun I see where you are coming from. But technological advancements can have many dimensions - a novel scientific hypothesis and validation of it is one. Before chatgpt, I didn’t see my friends or family mention llm. The technical advance here over other llms was human alignment + UX

5

9

233

Tejas Kulkarni

@tejasdkulkarni

2 years

delusions at an all time high in AI

11

10

203

Tejas Kulkarni

@tejasdkulkarni

10 months

Image to 3D: We are solving this problem at an unprecedented pace. This is just the beginning and we won't stop until we get to human-level performance to disrupt the market. Join us if you want to help create the next breakthrough:…

1

30

199

Tejas Kulkarni

@tejasdkulkarni

3 years

We are looking for a machine learning engineer to train large scale generative models. You will be part of a small but growing team of engineers and scientists building models to capture an unprecedented level of detail and diversity of physical objects.

Common Sense Machines, Inc. is looking for a Machine Learning Engineer in Massachusetts Ave,...

We are looking for a machine learning engineer to train generative models on large scale datasets and compute infrastructure. You will be part of a small but growing team of engineers and scientis...

jobs.smartrecruiters.com

2

35

178

Tejas Kulkarni

@tejasdkulkarni

2 years

diffusion model test on videos cooking over the weekend

16

2

156

Tejas Kulkarni

@tejasdkulkarni

2 years

It is an amazing time to be alive -- humanoids coming; generative AI getting better at producing real/imaginary stuff; LLMs interpolating and (non-humanly) generalizing in the space of human knowledge; AI folding proteins. AGI is already here and it's merely unraveling itself.

9

11

158

Tejas Kulkarni

@tejasdkulkarni

2 years

what causes these spikes while training diffusion (convs+attn) models (sometimes)? the results globally gets worse after the spike but the optimizer quickly bounces back and improves. interestingly the bounce back is very quick so is this some phase transition in the weights?

13

12

150

Tejas Kulkarni

@tejasdkulkarni

6 months

It is interesting to see how the model lifts the 2D sketch into the 3D mesh space, while preserving the same conceptual abstractions. Generated using Cube v1.3:

1

20

139

Tejas Kulkarni

@tejasdkulkarni

6 years

I don't understand many of the complaints about DL and DeepRL. Personally they are super useful and key components for modeling agents. So people working on other aspects of AI should either use them or demonstrate better alternatives. There is no other way to productively engage

7

18

135

Tejas Kulkarni

@tejasdkulkarni

3 years

The most loving AGI I will ever help create

1

0

130

Tejas Kulkarni

@tejasdkulkarni

4 years

Dreamcoder from Kevin Ellis et al. is an exciting+promising direction to efficiently learn new abstraction by solving tasks - . This method seems like a great direction for architecture search, option discovery in RL, programming tasks and more.

3

32

127

Tejas Kulkarni

@tejasdkulkarni

5 months

3

2

125

Tejas Kulkarni

@tejasdkulkarni

6 years

Our NIPS paper on self-supervised intrinsic image (shape,reflectance,lighting) decomposition. This is work from MIT days. Someday this might become useful for robot manipulation or image editing -- .

2

28

121

Tejas Kulkarni

@tejasdkulkarni

3 years

We are hiring a RE to build ultra-realistic simulations of everyday objects using deep learning & computer graphics/3D-CV. You will be part of a growing startup that is positioned to create real-world impact while being on the human-level AI path. @CSM_ai

3

14

120

Tejas Kulkarni

@tejasdkulkarni

7 years

Checkout our JMLR paper on a new and simple inference algorithm called Variational Particle Approximations --

0

34

109

Tejas Kulkarni

@tejasdkulkarni

5 years

two cool papers on learning discrete objects. if more people start taking this problem seriously we can solve it within the next few years at scale.

Google DeepMind

@GoogleDeepMind

5 years

Interested in unsupervised object decomposition & representation learning? We're excited to share two new approaches: MONet, which uses sequential decomposition & more recently IODINE, which uses iterative refinement MONet: IODINE:

10

291

898

1

15

108

Tejas Kulkarni

@tejasdkulkarni

2 years

Good inductive biases are not useless after all. This says more about the task that NERF primarily aims to solve — *dense* view based reconstruction is not as hard as it seems visually

AK

@_akhaliq

2 years

Plenoxels: Radiance Fields without Neural Networks abs: project page: propose a view-dependent sparse voxel model, Plenoxel, that can optimize to the same fidelity as NeRFs without any neural networks

11

166

902

4

5

102

Tejas Kulkarni

@tejasdkulkarni

4 years

Let’s hope that in 2020 more people/voices in AI admit to not knowing. This should open up the space for new ideas and expressions

1

6

103

Tejas Kulkarni

@tejasdkulkarni

1 year

GPT-4 is extremely powerful for coding -- its next level. I suspect coding workflows will now start escaping the trudges of IDEs -- programmers will spend more time on browsers. Can't wait to start using APIs. It just created me a #Unity game with prey/predator moving in…

0

17

89

Tejas Kulkarni

@tejasdkulkarni

2 years

Definitely some interesting points but also not true in real world ML. It's surprisingly quite the opposite in our experience: the more we replaced symbolic modules with DL components the better it got on real world data. DL is definitely not hitting any wall by a long shot.

Gary Marcus

@GaryMarcus

2 years

Deep Learning Is Hitting a Wall. What would it take for artificial intelligence to make real progress? #longread in ⁦ @NautilusMag ⁩ on one of the key technical questions in AI.

62

248

843

3

4

90

Tejas Kulkarni

@tejasdkulkarni

2 years

Exciting that basic Nerf as a piece of technology is now becoming a commodity. Understanding videos/sensors with rich appearance, dynamics,motion,topologies,multi modality and pre-trained prior is the real 3D problem now — needs cascaded self supervised learning

NVIDIA AI Developer

@NVIDIAAIDev

2 years

Trained in minutes, rendered in seconds. ⏱️ Our Instant NeRF from NVIDIA Research turns 2D photos into 3D scenes in the blink of an #AI . #nerfies #neuralnetworks #SX70 #GTC22

18

280

1K

3

7

89

Tejas Kulkarni

@tejasdkulkarni

3 years

❄️😍

3

2

91

Tejas Kulkarni

@tejasdkulkarni

6 years

Will neural nets escape the curse of end to end learning? Interesting paper in that direction ...

Nicolas Pinto

@npinto

6 years

"Shallow Learning For Deep Networks" -- an amazing paper submitted to ICLR 2019! Wow.

2

50

208

3

14

89

Tejas Kulkarni

@tejasdkulkarni

6 months

The real promise of AI agents won’t really show up with RAG or computer-LLM abstractions. Need intrinsic motivation at scale over multimodal domains and action spaces. Don’t hold your breadth or get worried/disheartened about AGI being built and controlled by large players - no…

6

14

88

Tejas Kulkarni

@tejasdkulkarni

1 year

Things just got more interesting in code generation land.

GitHub - salesforce/CodeGen: CodeGen is a family of open-source model for program synthesis....

CodeGen is a family of open-source model for program synthesis. Trained on TPU-v4. Competitive with OpenAI Codex. - salesforce/CodeGen

github.com

2

18

87

Tejas Kulkarni

@tejasdkulkarni

29 days

I took an image from the original Pokémon artists to see how far can go:

73

10

87

Tejas Kulkarni

@tejasdkulkarni

2 years

Didn't think markov chains would be cool again in the image domain. Diffusion models feels like a full circle back to the beginning of grad school 😂

3

1

79

Tejas Kulkarni

@tejasdkulkarni

2 months

Single Image → 3D → AI Re-texturing. Made using

3

17

79

Tejas Kulkarni

@tejasdkulkarni

6 years

If you are curious about how DeepRL, GANs and program synthesis fit together in the long term, come to our ICML oral at 5 pm in A7 (deep learning adversarial section)

Tejas Kulkarni

@tejasdkulkarni

6 years

Work with @yaroslav_ganin , Igor, @arkitus and @OriolVinyalsML . DeepRL + GANs + Program Synthesis + Robots

0

19

46

0

22

78

Tejas Kulkarni

@tejasdkulkarni

7 years

biggest challenge for Deep RL is sample complexity and reusable abstractions. the end-to-end agent mentality is to blame

6

9

75

Tejas Kulkarni

@tejasdkulkarni

7 years

Deep learning,probabilistic programming, RL, explicit low-mid-high level perception, all need to come together if we want common sense in AI

3

17

74

Tejas Kulkarni

@tejasdkulkarni

4 years

Computer Graphics has come a long way in the last 30 years. In the next 30 years it will solve Computer Vision.

Pixar Pioneers Win $1 Million Turing Award (Published 2020)

Ed Catmull and Pat Hanrahan created computer techniques that remade animation, special effects, virtual reality and artificial intelligence.

www.nytimes.com

2

9

73

Tejas Kulkarni

@tejasdkulkarni

2 years

CommonSim-1 @CSM_ai supports adding physics to NeRF rendered assets in blender. Would love to add more plugins into @nvidiaomniverse @UnrealEngine @unity depending on bandwidth and interest. Read more here -

1

11

74

Tejas Kulkarni

@tejasdkulkarni

1 year

It zero-shot discovers all entities in Montezuma's revenge. #AGI ;D

Tejas Kulkarni

@tejasdkulkarni

1 year

Significant progress towards resolving the age old gestalt problem in perception. From @MetaAI

0

30

3

14

73

Tejas Kulkarni

@tejasdkulkarni

3 years

@PyTorchLightnin reorganizes @PyTorch code significantly better. IMO its one of the best designed DL frameworks out there. I also predict that with integrations to MLFlow/Ray etc and multiple hardware/accelerator support, it will eat up a lot of MLOps tools/startups.

Lightning AI ⚡️

@LightningAI

3 years

Excited to announce our partnership with @facebookai to standardize research and production deep learning code at @Facebook with @PyTorchLightnin ! If you're a company still on the fence, Lightning is great for research & production!

2

47

272

0

10

72

Tejas Kulkarni

@tejasdkulkarni

1 year

AGI driven capabilities that will be shipped in 2023 (ones that will cross uncanny valley on several dimensions): 1. Multimodal 2D/3D generation with precise control. 2. Complex code generation. 3. Chatbots for UIs. 4. Generative datasets for training. 5. LLM enhanced search. 🚀

2

5

69

Tejas Kulkarni

@tejasdkulkarni

7 years

How visual entities arise and compose to create perception is the most fundamental question in AI. Capsules strike -

1

25

72

Tejas Kulkarni

@tejasdkulkarni

1 year

Solve hard but valuable problems. Otherwise AI is after you.

5

6

68

Tejas Kulkarni

@tejasdkulkarni

7 years

Some foundational conceptual frameworks for AI are (1) Horde by Sutton et al, (2) algorithmic complexity (3) core knowledge by Spelke et al

2

11

70

Tejas Kulkarni

@tejasdkulkarni

2 years

Amongst other things, we are building a foundation 3D world simulator @CSM_ai . It is early days but we are going to see extremely rapid progress. We are a small startup but moving very fast - and also deeply believe in open source. Reach out: jobs @csm .ai or DM

1

8

69

Tejas Kulkarni

@tejasdkulkarni

4 years

Differentiable volume rendering has come a long way! Impressive results from Mildenhall, Srinivasan, Tancik et al. -

2

9

68

Tejas Kulkarni

@tejasdkulkarni

5 years

Come see our talk at the Deep RL workshop (11.30am, 220E)! Object oriented (deep) RL/planning based approaches are powerful but require painful hand-crafting. We present a fully unsupervised approach towards this with Catalin Ionescu, @avdnoord , Andriy Mnih and @VladMnih

0

13

67

Tejas Kulkarni

@tejasdkulkarni

1 year

@gdb Ability to fine-tune gpt-3.5-turbo would be a game changer

3

65

Tejas Kulkarni

@tejasdkulkarni

11 months

Super proud of this product release from @CSM_ai Now we are one step closer to translating the web to realistic or stylistic 3D worlds -- with infinite variations. Outputs are compatible to your favorite game engine or 3D editor. Try it out! It is improving quite rapidly.

Common Sense Machines

@CSM_ai

11 months

We're thrilled to announce a breakthrough in 3D world generation. Now, transform ANY image - AI-generated, concept art or real world shots - into high-resolution game-engine ready 3D asset. Check it out: 🎈 Public Showcase on Discord: 🤖 Generate your own…

8

80

342

2

6

65

Tejas Kulkarni

@tejasdkulkarni

6 years

This talk from Malik made an impression on me during grad school — — I don’t see how model free, model based or unsupervised learning methods scale meaningfully for sensory motor domains without learning such reusable abstractions.

1

7

64

Tejas Kulkarni

@tejasdkulkarni

3 years

The joy of exploration

1

0

64

Tejas Kulkarni

@tejasdkulkarni

4 years

This paper raises interesting questions regarding data-driven/implicit and hand-crafted/explicit inductive biases

Aran Komatsuzaki

@arankomatsuzaki

4 years

An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale When pre-trained and transferred to CV tasks, Vision Transformer attains excellent results compared to SOTA CNNs while requiring much fewer computational resources to train.

5

208

940

1

5

63

Tejas Kulkarni

@tejasdkulkarni

3 years

reinforcement learning

6

0

63

Tejas Kulkarni

@tejasdkulkarni

2 months

If I can create stuff like this in 3D then anyone can. I am excited about the future where a 5 person team can build a AAA level world. Made using the latest in . Original image/session from (Nalya Bermeo):

57

6

60

Tejas Kulkarni

@tejasdkulkarni

4 years

Proud day for USA. Glad to see my 3mo old daughter be born in a world where gender barriers and perceived glass ceilings are breaking.

Women for Election

@women4election

4 years

How it went: How it’s going:

248

20K

92K

1

0

60

Tejas Kulkarni

@tejasdkulkarni

2 years

When I was at DeepMind I had a feeling that there would be a network effect of startups due to the sheer diversity and concentration of AI talent. @DeepMind should have just started an incubator -- had even suggested this a few times ;) --but tricky to organize and align for sure

Insider Tech Business

@insidertechbiz

2 years

Meet the DeepMind mafia: These 17 alumni from Google's AI research lab are raising millions for their own startups, from climate to crypto

0

8

32

2

3

61

Tejas Kulkarni

@tejasdkulkarni

6 years

DeepRL=DL+Search+update rules. What matters now is to improve the search by sample efficiently learning models. It’s very hard from pixels by sticking to 1 technique. But perhaps the combination of DL, ‘classical’ perception, planners and program induction can take us closer.

1

3

58

Tejas Kulkarni

@tejasdkulkarni

2 years

To do this, we are training a large action temporal 3d diffusion model. Preparing as fast as we can to run this on ~256 A100 80G GPU cluster now. Reach out if you want to work on this - it’s just the beginning and we are only scratching the surface. #CommonSim -1 @CSM_ai

Tejas Kulkarni

@tejasdkulkarni

2 years

Amongst other things, we are building a foundation 3D world simulator @CSM_ai . It is early days but we are going to see extremely rapid progress. We are a small startup but moving very fast - and also deeply believe in open source. Reach out: jobs @csm .ai or DM

1

8

69

2

7

59

Tejas Kulkarni

@tejasdkulkarni

8 years

thesis defense on June 13. Sad to leave MIT but also quite excited to join @DeepMindAI

1

57

Tejas Kulkarni

@tejasdkulkarni

6 years

a reasonable bet is that this natural curriculum will converge with the path for building AGI

Max Kleiman-Weiner

@maxhkw

6 years

Cognitive development milestones over the first 18 months of human life. A Gantt chart for building human-like AI from the DARPA Machine Common Sense challenge to build core cognition of objects, agents, and places into machines:

2

62

209

3

7

54

Tejas Kulkarni

@tejasdkulkarni

1 year

@OfficialLoganK @OpenAI pretty cool! already shipped it with Coder ->

Common Sense Machines

@CSM_ai

1 year

CSM-Coder () and ChatGPT-API is now live! From early feedback Coder is getting better (human-feedback) at complex blender/python/roblox coding but lacks explanations like ChatGPT. Now you get best of all worlds! Access for free:

7

30

225

1

6

54

Tejas Kulkarni

@tejasdkulkarni

4 years

I will be at the @NeurIPSConf to present this work tomorrow and more (Wednesday, 10:45 AM -- 12:45 PM @ East Exhibition Hall B + C. Poster #124 ). Come say hello if you are around!

Google DeepMind

@GoogleDeepMind

5 years

Deep RL agents are data hungry and often learn task-specific representations. Our model learns object-centric abstractions from raw videos. This enables highly data-efficient RL and structured exploration.

8

315

1K

0

6

55

Tejas Kulkarni

@tejasdkulkarni

2 years

You can get #dalle to compose logic if you let it 'talk' to another pre-trained network. i manually interfaced them as there are no python APIs yet (detic<>dalle). goal: "blue block on top of yellow block which is next to three green blocks, 3d rendering". first prediction:

2

3

55

Tejas Kulkarni

@tejasdkulkarni

20 days

Startup life is incredibly intense but only stressful if you emotionally/physically latch onto projected expectations. Once you realize this deeply, then you become unstoppable and ironically also an optimal reward maximizer.

6

1

54

Tejas Kulkarni

@tejasdkulkarni

1 month

The biggest realization at GDC: Deep Learning / Scaling Laws is about to hit gaming and most people are unaware of what this will mean. Feels like CVPR/NeuRIPS/EMNLP in the times of bayesian non-parametrics. Deep RL was the last time this happened (AlphaGo / Atari etc) but…

5

6

54

Tejas Kulkarni

@tejasdkulkarni

3 years

After having a child I am deeply skeptical of generalization via sparse-data extrapolation as the bootstrapping mechanism. In the first 9mos, there is 100s of hours of experience of a small # toys. So large-scale data-driven learning -> generalization seems more plausible.

2

0

52

Tejas Kulkarni

@tejasdkulkarni

2 months

Graph of the century

1

5

50

Tejas Kulkarni

@tejasdkulkarni

4 years

inverse graphics in the brain:

0

9

49

Tejas Kulkarni

@tejasdkulkarni

5 years

Here’s a person blind from birth abstracting their proprioceptive modality into objects and perspective geometry. Both of these abstractions are beyond data, especially perspective geometry. [Ref: …]

3

11

47

Tejas Kulkarni

@tejasdkulkarni

2 years

CommonSim-1 @CSM_ai supporting photorealistic rendering of Neural Radiance Fields with traditional textured assets (blender). Read more here -

0

13

48

Tejas Kulkarni

@tejasdkulkarni

7 years

We are organizing a NIPS workshop on learning disentangled representations. Submissions due Nov 5!

Learning Disentangled Representations

An important facet of human experience is our ability to break down what we observe and interact with, along characteristic lines. Visual scenes consist of separate objects, which may have different...

sites.google.com

0

12

46

Tejas Kulkarni

@tejasdkulkarni

6 years

Why are polls being used here to make the final decision? Isn't it common sense to change the name to create a more inclusive vibe in the community?

2

45

Tejas Kulkarni

@tejasdkulkarni

6 years

Work with @yaroslav_ganin , Igor, @arkitus and @OriolVinyalsML . DeepRL + GANs + Program Synthesis + Robots

Oriol Vinyals

@OriolVinyalsML

6 years

RL + GANs: Program synthesis with an agent that uses a paint program to fool a discriminator. Paper+Blog:

4

207

563

0

19

46

Tejas Kulkarni

@tejasdkulkarni

3 years

"30 trillion parameters on 512 NVIDIA V100 Tensor Core GPUs". That's ~1/4 the total number of synapses in the human brain.

ZeRO-Infinity and DeepSpeed: Unlocking unprecedented model scale for deep learning training -...

Since the DeepSpeed optimization library was introduced last year, it has rolled out numerous novel optimizations for training large AI models—improving scale, speed, cost, and usability. As large...

www.microsoft.com

2

7

44

Tejas Kulkarni

@tejasdkulkarni

3 years

This critique implies that vision is not solved and language models are not grounded. Grounded unsupervised deep learning is still the holy grail. Just a few more layers on AlexNet was supposed to solve it, then Deep RL, then VAE/GAN, then Nerf ... needs a new DL breakthrough

Serge Belongie

@SergeBelongie

3 years

“These models are really castles in the air. They have no foundation whatsoever.” -J. Malik

13

39

276

0

1

44

Tejas Kulkarni

@tejasdkulkarni

2 years

a new recipe by mixing rasmalai and donuts. #dalle

5

2

42

Tejas Kulkarni

@tejasdkulkarni

1 year

ChatGPT-API was out a few hours ago and now its integrated and live on our website.

Common Sense Machines

@CSM_ai

1 year

CSM-Coder () and ChatGPT-API is now live! From early feedback Coder is getting better (human-feedback) at complex blender/python/roblox coding but lacks explanations like ChatGPT. Now you get best of all worlds! Access for free:

7

30

225

2

1

41

Tejas Kulkarni

@tejasdkulkarni

7 years

nice paper making connections between policy grads and Q values from @OpenAI (Schulmann et al) --

0

8

41

Tejas Kulkarni

@tejasdkulkarni

3 months

Humans are coming out pretty good now. [Cube generation with high resolution & lowest pixel alignment]

7

6

40

Tejas Kulkarni

@tejasdkulkarni

2 years

Variations of the environments that my daughter saw when she was 2mo old (head cam). Most of what she learned was from passive observations--one reason why I am excited about generative models. This type of AGI seems to learn from a lot of data but from a relatively small env set

2

0

40

Tejas Kulkarni

@tejasdkulkarni

9 months

Everything here except the grassy terrain is AI generated using CSM-Cube x Mixamo x Unreal Subscribe now:

3

4

39

Tejas Kulkarni

@tejasdkulkarni

3 years

You will always be remembered as one @pmddomingos

Pedro Domingos

@pmddomingos

3 years

The words “racist” and “sexist” are so destructive and unhelpful that in the future their use will be frowned on.

22

4

101

2

39

Tejas Kulkarni

@tejasdkulkarni

6 months

Unfortunately AGI has become more of a meme to capture attention, recruit the smartest folks, do a regulatory capture, deemphasize creativity and momentum amongst AI builders, gain more social media attention by fear mongering, raise a lot of capital due to fear of winner take…

Louis Anslow

@LouisAnslow

6 months

Here is Nick Bostrom admitting that AI panic feels out of control right now, "like a wrecking ball" that could "destroy the future"

33

70

371

3

10

38

Tejas Kulkarni

@tejasdkulkarni

4 years

GPT-3 is re-emphasizing in a strong way that neural neural networks are not just pattern recognizers. There is a weird and interesting geometry between weights and discrete symbols that large-scale computation traverses (sub-symbolic is not quite the right word)

2

0

39

Tejas Kulkarni

@tejasdkulkarni

2 years

Use CommonSim-1's @CSM_ai mobile/web app/APIs to create objects models and synthetic training data to generate 6DOF mesh/NeRF trackers. Now your vision systems and robots can start seeing the real world in radiance fields! Read more -

1

3

39

Tejas Kulkarni

@tejasdkulkarni

5 years

Meditation has been the biggest personal breakthroughs and I could not have imagined how deeply you can explore within yourself.

Bill Gates

@BillGates

5 years

When it comes to meditation, @andypuddicombe turned me from a skeptic to a believer.

314

828

5K

0

39

Tejas Kulkarni

@tejasdkulkarni

2 months

Inverse Pixar has been a long standing dream. I am incredibly excited about this frontier, especially after Sora's release. It presents a radically new way to animate game-engine worlds - with multi-object interactions. It is the beginnings of a true 3D world simulator with…

Common Sense Machines

@CSM_ai

2 months

We are excited to preview a breakthrough in 3D generative AI. Video models are incredible but do not directly allow for controllability and simulation. We are closing this gap by predicting 4D meshes from video inputs. Join the waitlist at to receive…

9

41

247

0

4

38

Tejas Kulkarni

@tejasdkulkarni

2 years

an army of monkeys wearing VR goggles on a busy street on mars, with a desert storm approaching in the back. digital art in the style of salvador dali #dalle

2

1

38

Tejas Kulkarni

@tejasdkulkarni

5 months

@ylecun Agree with your sentiment but it also felt too polarizing as a grad student. Non symbolic theories were not taken as intellectually seriously as other ones. I remember suffering through silly/unnecessary social wrath when I got overjoyed after seeing the results of AlexNet in my…

1

36

Tejas Kulkarni

@tejasdkulkarni

6 years

While tracing the history of where RL algorithms originated, I came across this seminal paper again from 1960 -- . Footnote 18 starts to vaguely introduce the idea of temporal difference learning.

2

4

38

Tejas Kulkarni

@tejasdkulkarni

10 months

Image to 3D (Unity) worlds is going to rapidly accelerate in the coming weeks. More broadly this is going to revive Deep RL research. Generating realistic+infinite 3D environments was the key bottleneck in its way.

nakashun / Witchpot

@nakashun_t

10 months

画像→3Dモデル生成AIのCSM( @CSM_ai )で作ったオブジェクトでゲームステージ作り Midjourney→CSM→Unity(stable diffusion for unity) のワークフローかなり良さそう 3D生成の待ち時間、生成クオリティなど粗は探せばいくらでもありますが、どんどん進化していく前提で考えるとめっちゃ未来が楽しみ

1

70

265

2

4

38

Tejas Kulkarni

@tejasdkulkarni

1 year

After this, if anyone still takes @ESYudkowsky seriously, then you should deeply question why. This guy is using an ultra dangerous narrative and has some deep psychological issues to work through.

Joshua Achiam ⚗️

@jachiam0

1 year

@RichardMCNgo I don't think this is a misunderstanding. If you say, "The people who are building AI are going to kill my children and your children, violence is acceptable and even necessary to shut this down," that is not. Goddamn. Okay.

9

6

111

7

2

36

Tejas Kulkarni

@tejasdkulkarni

2 years

when AGI reaches enlightenment. #dalle (different prompt with lots of variations)

2

0

36

Tejas Kulkarni

@tejasdkulkarni

2 years

tiger wearing a saree in a lab mixing chemicals, portrait in the style of Raja Varma #dalle

1

2

35

Tejas Kulkarni

@tejasdkulkarni

1 year

TEDx Boston | 3D World Simulation: The last domino before we get AGI | Tejas Kulkarni ... via @YouTube

3D World Simulation: The last domino before we get AGI | Tejas...

As AI learns to master building simulators and game worlds, it will result in a Cambrian explosion of new capabilities for both physical and digital tasks, p...

www.youtube.com

2

7

36

Tejas Kulkarni

@tejasdkulkarni

4 months

Nice book from @maxsbennett tracing how intelligence may have emerged from DNA -> ribosomes -> proteins -> … evolution of neurons -> TD / back prop … -> simulation engines -> language/theory of mind etc — all from a reductionist point of view. Definitely worth a read. Thanks…

2

8

36