Param Popat

ML @apple SPG (3D Deep Learning, Reinforcement Learning) | @ColumbiaCompSci @Columbia

[Param Popat @parampopat]

Excited to see Apple Intelligence bring powerful AI capabilities into everyday experiences — including Describe a Shortcut, announced today.
I had the opportunity to work on this feature, which makes Shortcuts more approachable by letting users simply describe what they want to automate, and then helping assemble the steps on their behalf. Even better, users can describe tweaks or additions, and Shortcuts can adjust from there.
I’ve always loved Shortcuts because it gives people a way to make their devices feel truly personal. Bringing natural language, tool use, and automation together in a privacy-first Apple experience has been an especially meaningful problem to work on.
Proud of the teams behind this launch, and excited for users to try it. https://t.co/vFh7ENHoBs

[Param Popat @parampopat]

Oof, what a gold mine! Absolutely loved this!

[Jeff Dean @JeffDean]

Performance Hints

Over the years, my colleague Sanjay Ghemawat and I have done a fair bit of diving into performance tuning of various pieces of code. We wrote an internal Performance Hints document a couple of years ago as a way of identifying some general principles and we've recently published a version of it externally.

We'd love any feedback you might have!

Read the full doc at: https://t.co/jej95g236P

[Tim Cook @tim_cook]

Battersea Power Station is merry and bright with Christmas trees designed on iPad this holiday season! Thanks to Munya Chawawa and our collaborators for sharing their creativity—every submission brought joy to all of us at Apple!

[Param Popat @parampopat]

@JeffDean But can it make a usd splat out of an image?

[Tim Cook @tim_cook]

Life is about the memories we make.

[Aleksei Petrenko @petrenko_ai]

Excited to share our new pre-print https://t.co/FtT3hkHKeG

We train a digital agent that solves diverse day-to-day tasks from the AppWorld benchmark by interacting with its stateful environment using API calls. AppWorld is hard! The previous best open-weight agent (Llama 3 70B) reached only a 7% success rate on the hardest test split. Our RL algorithm, LOOP - a PPO variant with Monte Carlo baselines - achieves a 45.7% success rate, 24% over the base Qwen 2.5 32B, and 9% higher than a much larger OpenAI o1.

[Param Popat @parampopat]

@svlevine Happy Pi Day!

[Param Popat @parampopat]

@PurvaTendulkar @cvondrick Congrats, Purva and Carl!

[Awni Hannun @awnihannun]

All of the LLM examples in MLX-examples now support quantized models out of the box:

- Mistral, Mixtral, Llama, Phi2, Qwen
+ Code, Chat, and Instruct variants

https://t.co/OEjf28RKKr

[Param Popat @parampopat]

Will Buxton! @wbuxtonofficial

[non aesthetic things @PicturesFoIder]

who would you choose?

[Param Popat @parampopat]

@YashviDhruvv Why does my phone autocorrect it to “Wash”

[Param Popat @parampopat]

@tejasvisoi And that you can listen to what your phone's microphone is taking in through Live Listen.

[Param Popat @parampopat]

@RushiBhimani95 Why would you use a toughened glass on top? It's such a beautiful screen with rich colors

[Param Popat @parampopat]

@ClementDelangue @humairawins There are so many podcasts out there, I guess one opportunity would be to create a short series with 10ish episodes of 1 hour each going through each founder's story. Sort of like a Masterclass on learning from their lessons.

[Param Popat @parampopat]

@goodfellow_ian Glad that everything’s in under control now!!

[Zhou Yu @Zhou_Yu_AI]

My lab is looking for a postdoc to work on some exciting LLM and Multimodal work. Please fill this simple form if you are interested: https://t.co/dG0YM6OF8A Colleagues and friends, please retweet to spread the words.

[Param Popat @parampopat]

@Coffee_and_NLP Calculator App!

[Param Popat @parampopat]

@BhosalePratim Have you tried asking ChatGPT/Bing/Bard (choose your poison) for ideas?

[Param Popat @parampopat]

@stale2000 I think it's still just "sf". Hoping it'll change as the date comes closer

[David Berthelot @D_Berthelot_ML]

New paper TRACT - Faster diffusion model sampling
- Single-step diffusion SotA for CIFAR10 and ImageNet64 with L2 loss without architecture changes
- Up to 2.4x FID improvement
https://t.co/UOMlLb0plG