Vipul Ramtekkar

I like computer vision and robotics. Computer Vision Research Engineer at Honda R&D Japan. Former undergrad @iitbombay.

[Vipul Ramtekkar @VipulRamtekkar]

Had an absolute blast playing with @parimarjan and having a talk with all the talented folks in tech @sh_reya @suzatweet. Thanks a lot @chipro for this!

[ali @waterloo_intern]

it took 1 intern 3 months of continuous work, but eventually, a quantization method that beat every other algo in the market, including @nvidia's official modelopt

to explain why this matters, i ask for exactly 69 seconds of your attention (275 words @ avg reading speed of 238 wpm):

frontier models (like glm52) are huge (~0.8T params). as released, each parameter takes 2 bytes (bf16), so overall size is about 1.6 tb

a b200 has 180gb of memory. a node of 8 gives you 1.44 tb, barely fits weights, much less activations / kv cache

must quantize the model (reduce the size of each individual parameters) to serve. fp8 quantization means each parameter takes 1 byte (fits in 0.8 tb), fp4 takes 1/2 a byte (fits in 0.4 tb)

cutting the model to a quarter its original size is necessary for it to run a) cheap b) fast, and every lab serving models does this.

but, quantization lobotomizes the model if not done correctly (this is why you see people complain about @AnthropicAI nerfing claude or @OpenAI nerfing codex)

there are currently several algorithms (like Nvidia's official model-opt) that attempt to figure how to quantize a model with the least amount of damage.

they find the redundant layers that can be slashed, and sensitive/important layers that need to stay in full-precision.

these algo's have two drawbacks:
1) they take a long time to run
2) they quite often result in a sub-optimal configuration

for the past 3 months, a research (and, as always, waterloo) intern on our model perf team (@the_joshua_hill) came up with a new quant algorithm.

it consistently finds the optimal configuration:
a) in less time than SOTA
b) with more aggressive quant than SOTA
c) scoring higher on benchmarks than SOTA

achieving just one of the above is a feat on its own.

all three...excited for the paper to come out this week

[Yuke Zhu @yukez]

Today, we're introducing SimFoundry, our real2sim2real framework at NVIDIA GEAR that automatically turns real-world scenes into simulation-ready worlds from a single image or video.

Website: https://t.co/JB3kf3GlYm
Paper: https://t.co/pVhE1qWXtU

This work marks a major step for our team toward leveraging simulations and synthetic data for foundation model training and systematic policy evaluation at scale. Code will be open-sourced soon. Stay tuned!

[Matt Krisiloff @mattkrisiloff]

I’m so excited to share this update on @Conception –

We’ve generated the first early human eggs derived from stem cells.

This is a big deal -- the potential to redefine fertility is real.

[PLaMo @PLaMoLLM]

【PLaMo™翻訳】
公式としては、今回のアップデートで「Liteプランのファイル翻訳件数上限引き上げ」を全力で推したいです🔥
無制限のブラウザ拡張に加えてファイル翻訳が最大20件まで利用可能になり、普段遣いに最適です！
月額1000円でコスパ良いので、ぜひお試しください！
https://t.co/HcrP4QBxWt

[Rohan Pandey @khoomeik]

been in ML research for 7 years, wrote a paper on compression & scaling laws, and passed openai's information theory interview

yet the latest 3b1b *still* gave me fresh intuition on entropy

either i'm an impostor or 3b1b is the greatest teacher of all time

[Daisuke Okanohara / 岡野原 大輔 @hillbig]

PLaMo 3.0 Primeの提供を開始しました！
主要ベンチマークで大幅な性能向上を達成し、同価格帯の商用モデルやオープンモデルと比較しても、日本語指示追従、コーディング、ツール利用などで、匹敵、上回る性能を実現しています。さらに256K コンテキスト長への対応してます。
ぜひお試しください！

[Preferred Networks @PreferredNetJP]

【発表】国産フルスクラッチ開発の生成AI基盤モデルPLaMo 3.0 Primeの提供を開始しました。
1⃣API経由またはオンプレで利用可能
2⃣複雑なタスクに対応するReasoningモデル、応答速度の速いNon-reasoningモデルを提供
3⃣高い日本語性能とコストパフォーマンスを両立
4⃣コンテキスト長を64kから256kに拡張
5⃣危険情報に対する安全性向上
https://t.co/r7IeV3OdwC
AIエージェントや企業の実務利用を想定した機能強化がされていますのでおぜひ試しください

[Vipul Ramtekkar @VipulRamtekkar]

@deepigoyal Wait how is it even fair to compare my guess a highly engineered composite metric to raw probably unadjusted heart rate?! A real comparison should pitch raw metrics against each other like raw cerebral blood flow versus raw heart rate to see which signal is actually better.

[Yunsong Zhou @Yunsong_Zhou]

🧐 Simulation has long promised robot pretraining, but breaks at the moment of real-world deployment.

🚀 Today, we introduce SIM1: the first real-to-sim-to-real paradigm where the generative world becomes the same one as reality.

SIM1 produces simulation data whose execution is directly valid in the physical world, enabling policies trained entirely in simulation to transfer zero-shot, at scale.

📈 This unlocks a new scaling law for robotics: we scale intelligence without scaling real-world data.

✨ Few demonstrations in, real-world policies out.
Simulation is no longer a proxy; it is supervision itself.

https://t.co/Kp1YBe5Gmf
https://t.co/GG2SBQfPpG

[vivek @itsreallyvivek]

https://t.co/ZhOgyq7Vgn

[Vipul Ramtekkar @VipulRamtekkar]

@jbhuang0604 Understood! Makes sense, thank you!

[Jia-Bin Huang @jbhuang0604]

Hey! A new vision encoder for robotics is in town 👀🤖

Instead of using models trained on static images (CLIP, SigLIP, DINO), we bake the dynamics-awareness directly into perception.

It transfers well everywhere and boosts real-world OOD success by +22.5%

Check it out👇

[hardmaru @hardmaru]

For over a decade, we’ve accepted that end-to-end backprop is the only way to train deep networks. But holding the entire network in memory all at once is why AI training is hitting a resource wall.

We found a new way to break the network into blocks and train them independently. The trick? Treating the network’s forward pass like a diffusion model denoising a signal.

This reinterpretation slashes the memory needed to train deep models. In our #ICLR2026 paper (https://t.co/PK5h0mqQSo), we matched end-to-end performance across ViTs, DiTs, and LLMs. We did this while training just one isolated block at a time.

[Vipul Ramtekkar @VipulRamtekkar]

This has got to be the funniest demo 😂😂

[Thinking Machines @thinkymachines]

Our model interjects to save Alex's parents from his eccentric ideas.

[fatih kadir akın @fkadev]

My son asking me a lot of questions. It’s a distillation attack obviously.

[Vipul Ramtekkar @VipulRamtekkar]

@pathak2206 @SkildAI Congratulations!!

[Vipul Ramtekkar @VipulRamtekkar]

@krandiash SONIC! And huge congratulations. I would love test it out!

[Vipul Ramtekkar @VipulRamtekkar]

@kamath_sutra @smallest_AI @Sierra_Ventures @3one4Capital @better_capital @UpsparksCapital @SchemaVentures @DeVC_Global @tinyvc @peercheque Wow that's amazing! Remember seeing you in Bakliwal

[Vipul Ramtekkar @VipulRamtekkar]

@simonw That was my first thought! Why would they do it so evidently otherwise.