Cyril Zakka, MD @cyrilzakka Twitter profile | Pikagi

Pikagi

Cyril Zakka, MD

@cyrilzakka

3,683

Followers

282

Following

418

Media

6,107

Statuses

Medical Doctor | ML Researcher | Postdoctoral Research Fellow @StanfordCTSurg @HiesingerLab

@[email protected]

https://t.co/ykcqoV8bbC

Joined December 2013

Don't wanna be here? Send us removal request.

Pinned Tweet

@cyrilzakka

Cyril Zakka, MD

7 months

There’s been a lot of buzz surrounding multimodal LLMs in medicine. At the @HiesingerLab we’ve been building Almanac Chat for over 6 months alongside our collaborators and we’re excited to showcase the first preview! • Almanac Chat is the first truly multimodal medical LLM

15

55

304

Last Seen Profiles

@MossPresents

@moneymikehr23

@termc1

@little_sub0226

@Slutty_Kashish

@zestylemonvt

@Firdaus43041518

@IneffableChar

@IsaluteTheShits

@WildOnTheMove

@BrazilJonne

@pjmclumsy

@HeyEmilieBoiv

@zachywac

@haqnni

@KimRussell60876

@Esher05

@MalcolmAlick

@rlh0r

@mediablasters1

@BakbakEnt

@zckej

@TessaCopp

@garrascontext

@kokuranggaul

@DiaperedDukeSJ

@neh_pch

@AlbaziAssyrian

@KIMEPerformance

@aerinagauchi

@sanmarinotitans

@PhotoKuro_

@whudjo

@simpBVerfG

@SanchariVijay

@Jas_Bains_12

@cyrilzakka

Cyril Zakka, MD

1 year

I know iOS/macOS ChatGPT apps are all the rage at the moment but looks like at least 50% of them are leaking their private @OpenAI API keys through their property lists/app binaries. (n=10) I've sent emails to the developers, but here's a quick thread: 👇🏻

Tweet media one

Tweet media two

67

311

2K

@cyrilzakka

Cyril Zakka, MD

1 year

I know a lot of folks were quite happy with the iOS 17 NameDrop animation demoed at WWDC so here's a quick pass attempt at replicating it: Code: cc/ @twostraws @b3ll

14

139

2K

@cyrilzakka

Cyril Zakka, MD

1 year

Over winter break I decided to make Notos, an #AI -powered transcription application based on @OpenAI ’s Whisper. Here’s a small demo where it manages to transcribe everything flawlessly! TestFlight:

28

48

469

@cyrilzakka

Cyril Zakka, MD

1 year

In short, never store sensitive data on the client side unless you're comfortable with it being compromised. A sufficiently motivated attacker will be able to reverse-engineer whatever you throw their way.

9

22

373

@cyrilzakka

Cyril Zakka, MD

1 month

I’m sick so I decided it was enough of an excuse to recreate the @newcomputer transitions in Swift. Here’s Iris, my own personal assistant. It has access to all my health/clinical, calendar, location, and weather data. Server runs on my M1 Mac using MLX (cc @awnihannun ) so

13

13

293

@cyrilzakka

Cyril Zakka, MD

2 years

I am pleased to announce that I’ll be joining @StanfordCTSurg at @StanfordMed as a postdoctoral research fellow in the Hiesinger Lab, where I’ll be working at the intersection of robotic surgery, cardiac imaging, and #AI . Looking forward to the next chapter!

Tweet media one

25

9

213

@cyrilzakka

Cyril Zakka, MD

2 years

I’m delighted to announce the launch of @AUBMC_AIM , the program for Artificial Intelligence in Medicine at @AUBMC_Official to advance patient care with AI. A special thanks to @julien_c and our friends over @Huggingface 🤗whose sponsorship made this possible.

9

28

188

@cyrilzakka

Cyril Zakka, MD

1 year

A typical iOS payload has the following structure. We're particularly interested in the 'Info.plist', which is simply a file containing key-values in plain string bundled along with your application. They're typically used to store application metadata/user settings.

Tweet media one

1

3

157

@cyrilzakka

Cyril Zakka, MD

1 year

OTOH storing your🔑 in code might seem like a safer option since human-readable text is turned into a machine-executable binary upon compilation... well not quite. Using a reverse-engineering tool makes your API key plainly visible from inside the compiled executable.

Tweet media one

4

1

154

@cyrilzakka

Cyril Zakka, MD

9 years

Giving away 10 copies of @EquinoxKeyboard . RT and Fav to enter. Giveaway ends Thursday #EquinoxKeyboard

Tweet media one

7

121

135

@cyrilzakka

Cyril Zakka, MD

1 year

I've been working on a small native Mac app ( #Swift / #SwiftUI ) to access #ChatGPT / #GPT natively from anywhere on your Desktop. Simply ⌘+⇧ and spacebar from anywhere on your Mac! If you'd like to help me beta test it, please shoot me a DM!

16

9

153

@cyrilzakka

Cyril Zakka, MD

1 year

Storing your 🔑 within the Info.plist makes it particularly easy to extract since they're made available in plain string without requiring any fancy tooling or much effort.

Tweet media one

Tweet media two

1

0

151

@cyrilzakka

Cyril Zakka, MD

1 month

Super excited to showcase our newest work Almanac Copilot an EHR agent capable of answering questions about your patients and placing orders for you across any modern EHR system. Background: Nearly 75% of clinicians with burnout symptoms pinpoint EHRs as a source due to poor

Tweet media one

6

34

148

@cyrilzakka

Cyril Zakka, MD

6 months

Some more SwiftUI and Metal shader fun to match the San Francisco weather these past few days.

5

5

144

@cyrilzakka

Cyril Zakka, MD

2 years

I know #SwiftUl has been getting a lot of criticism recently (and for good reason) but I really can't get over how much easier it is to build complex Uls compared to UlKit. Here's the new Apple Watch Ultra made entirely in SwiftUl. /cc @stroughtonsmith

4

12

145

@cyrilzakka

Cyril Zakka, MD

6 years

@flyosity I don’t see it moving

5

2

123

@cyrilzakka

Cyril Zakka, MD

4 years

There's been a lot of misinformation going around about #COVID ー19 especially on social media. So I decided to create a WhatsApp chatbot that answers your most pressing questions about the virus. Oh and it's powered by the latest and greatest in #NLP #BERT /cc @huggingface

9

32

134

@cyrilzakka

Cyril Zakka, MD

8 months

To unlock AI capabilities across a range of surgical procedures and medical imaging workflows, I built Avicenna, an augmented reality DICOM viewer with first class support for on-device AI inference, 3D reconstruction, annotations, and more, courtesy of the @HiesingerLab . To

Tweet media one

8

14

113

@cyrilzakka

Cyril Zakka, MD

5 years

Made a simple image viewer in SwiftUI. Good practice for chaining gestures. Features: 1) zoom 2) rotate 3) double-tap to reset the zoom or switch between content modes 4) long-press for action sheet 5) drag to dismiss or pan

1

5

111

@cyrilzakka

Cyril Zakka, MD

5 months

Our research on ‘Almanac: Retrieval Augmented LLMs for Clinical Medicine’ is finally out on @NEJM_AI ! We demonstrate that LLMs augmented with tools (browser, calculators) outperform ChatGPT-4 & others on a novel dataset of 300+ questions assessed by board-certified physicians!

6

25

107

@cyrilzakka

Cyril Zakka, MD

1 year

Thrilled to announce Almanac Chat, a new multi-institutional research collaboration to explore the potential uses and limitations of large multimodal language models for clinical medicine! Find out more:

Tweet media one

4

34

101

@cyrilzakka

Cyril Zakka, MD

5 years

SwiftUI Modal View with fullscreeen and half-modal capabilities. Can be extended to enable/disable DragGesture as needed. Enjoy! 🎉 /cc @J_Carlos_Perez @vatsal_manot @Dimillian

6

21

99

@cyrilzakka

Cyril Zakka, MD

1 year

The newest version of our Almanac preprint is out, and just in time for our demo at the @StanfordAIMI Symposium 2023! Almanac is a retrieval-augmented LLM that provides up-to-date and verifiable answers to medical queries. Link: We benchmark our

6

16

89

@cyrilzakka

Cyril Zakka, MD

1 year

@dabidoYT @OpenAI Definitely a server (although code obfuscation can make reverse engineering harder). If you don’t want to deploy your own I’ve personally used CloudKit database before!

4

0

78

@cyrilzakka

Cyril Zakka, MD

10 months

Really like this new paper from @giffmana and team. Replacing the softmax-based loss in contrastive learning with a sigmoid-based loss improves scalability and results. Link:

Tweet media one

1

7

75

@cyrilzakka

Cyril Zakka, MD

7 months

Super thrilled to be co-leading the session on Generative AI in Medicine with @RoxanaDaneshjou at the AI+Health conference organized by @StanfordAIMI and @StanfordHAI . Don't forget to register at (CME credits available)

Tweet card media

aimi.stanford.edu

2

10

71

@cyrilzakka

Cyril Zakka, MD

6 months

My biggest flex is DM-ing @julien_c about a ViT-like architecture in January of 2020, 10 months before the release of the actual ViT paper. I couldn't get the training loop to work properly and was neck deep in my second year of medical school.

Tweet media one

2

8

72

@cyrilzakka

Cyril Zakka, MD

3 years

You've probably heard of exclusive #NFTs like the Bored Ape Yacht Club that sell for thousands of dollars... I had a bit of free time this week so I trained a small AI model to generate entirely new ones with a bit of help from @PDillis and @karlhabashy .

6

11

71

@cyrilzakka

Cyril Zakka, MD

1 year

@OfficialLoganK @OpenAI Thanks for reaching out, I just DM’ed you

0

0

64

@cyrilzakka

Cyril Zakka, MD

4 months

Big models are still getting bigger, but on-device domain-specific models are the future, especially for privacy-first applications like medicine. Here's a quick look at Notos to transcribe patient encounters entirely on device with the ability to export to any EHR like Epic.

5

6

68

@cyrilzakka

Cyril Zakka, MD

5 years

Probably the last SwiftUI demo before I release and open source the app I’ve been working on. iOS 13 Music ‘Next Song’ Animation ⬜️⬛️

3

3

50

@cyrilzakka

Cyril Zakka, MD

10 months

Looks like I get to test the @Apple Vision Pro at the end of August! /cc @HiesingerLab

Tweet media one

4

0

46

@cyrilzakka

Cyril Zakka, MD

1 year

Our latest preprint is out! We examine how large language models can improve clinical guideline lookup, tx recommendations, and medical calculations by using external knowledge bases. Our framework Almanac achieves significant gains in factuality! 📄:

Tweet media one

4

11

46

@cyrilzakka

Cyril Zakka, MD

2 years

Was working on something very similar but never got the chance to publish due to finals and graduation. Still a WIP but I've released my code here with some changes inspired by the paper (e.g. sin-cos embedding vs learned):

Tweet card media

GitHub - cyrilzakka/MAE3D: Masked Auto-Encoding for Large Scale Pretraining of Video Data

Masked Auto-Encoding for Large Scale Pretraining of Video Data - cyrilzakka/MAE3D

@_akhaliq

AK

2 years

Masked Autoencoders As Spatiotemporal Learners abs: MAE method can learn strong representations with almost no inductive bias on spacetime (only except for patch and positional embeddings), and spacetime-agnostic random masking performs the best

Tweet media one

1

51

283

1

5

43

@cyrilzakka

Cyril Zakka, MD

1 year

Rode @Cruise for the first time last night and was honestly blown away by the experience. Great work @olivercameron !

@olivercameron

Oliver Cameron

1 year

“Any sufficiently advanced technology is indistinguishable from magic.” DM me if you want an invite to @Cruise .

20

7

147

1

9

42

@cyrilzakka

Cyril Zakka, MD

10 years

Conformity is finally here! Check it out and tell me what you think! RT please, trying to get it developed :) http://t.co/007ri7WKP9

Tweet media one

16

50

40

@cyrilzakka

Cyril Zakka, MD

2 years

Just finished the phenomenal review of #AI in medicine by @EricTopol and @pranavrajpurkar . Here are some key highlights:

Tweet card media

AI in health and medicine

Nature Medicine - AI has the potential to reshape medicine and make healthcare more accurate, efficient and accessible; this Review discusses recent progress, opportunities and challenges toward...

2

11

41

@cyrilzakka

Cyril Zakka, MD

2 years

In SF until April 30 for my medical electives. I visited @kevin_zakka at the @berkeley_ai lab today 🧠🦾

Tweet media one

Tweet media two

Tweet media three

1

1

39

@cyrilzakka

Cyril Zakka, MD

1 year

Putting the ChatGPT API to work on a third party iOS keyboard makes for quick content generation in virtually any app! Sorry for the spam @rohanshad

3

7

38

@cyrilzakka

Cyril Zakka, MD

4 years

Preprint Alert! "MammoGANesis: Controlled Generation of High-Resolution Mammograms for Radiology Education" is now on arXiv! 📄 Paper: ✍🏼 Blog: And just in time for #BreastCancerAwareness month 🎀 Read on to learn more: 1/7

1

9

36

@cyrilzakka

Cyril Zakka, MD

3 years

Created a small iOS app to read and interpret 12-lead ECGs. The underlying Al model was trained on close to 15k ECGs to detect atrial fibrillation, myocardial infarctions, atrioventricular blocks, etc.

4

6

35

@cyrilzakka

Cyril Zakka, MD

5 years

iOS 13 Music Action Sheet recreated in SwiftUI. Couldn’t recreate the stacked card effect when expanded though..

2

1

34

@cyrilzakka

Cyril Zakka, MD

1 year

Excited to announce that I’ll be presenting at #AATS2023 in Los Angeles on Augmented Reality and Telesurgery in CTS. Looking forward to meeting everyone! /cc @HiesingerLab

Tweet media one

3

0

34

@cyrilzakka

Cyril Zakka, MD

4 months

The current crop of generative AI applications in healthcare are only scratching the surface of what's possible. Documentation and question-answering are just the lowest hanging fruit.

2

3

33

@cyrilzakka

Cyril Zakka, MD

5 years

@sarahfrier @business If you switch it to creative, you have all the benefits of a business account without making your contact info public

0

7

30

@cyrilzakka

Cyril Zakka, MD

11 months

The @HiesingerLab just got a nice little delivery this morning. Can you guess what's coming next?🤖🫀 Thanks @IntuitiveSurg for the donation!

Tweet media one

Tweet media two

0

4

33

@cyrilzakka

Cyril Zakka, MD

2 months

I have a **124.5M parameter** LLM that obtains a 32% on MedQA. We need to move away from these benchmarks to really push the field forward. There’s no reasoning involved, and none that are actually relevant to the way these models will be deployed in clinic.

@maximegmd

Maxime G, M.D

2 months

Internistai 7b: Medical Language Model Today we release the best 7b medical model, outperforming GPT-3.5 and achieving the first pass score on the USMLE! Our approach allows the model to retain the same general capabilities as the base model.

15

65

303

6

3

32

@cyrilzakka

Cyril Zakka, MD

7 months

Dropping something super exciting in the next two weeks… almost 6 months of work!

2

0

32

@cyrilzakka

Cyril Zakka, MD

2 months

Looking forward to meeting everyone there #wwdc24

Tweet media one

1

0

31

@cyrilzakka

Cyril Zakka, MD

6 months

We finally have all the components to continue our research into autonomous robotic surgery using the DaVinci Research Kit!

Tweet media one

Tweet media two

1

0

31

@cyrilzakka

Cyril Zakka, MD

1 year

Especially thrilled to announce that the @HiesingerLab has been awarded a @StanfordHAI seed grant to pursue the development of SuTr, a foundational #AI vision model for #surgicalAI . If you’d like to help out by contributing surgical videos please get in touch!

@HiesingerLab

Hiesinger Lab

1 year

We are excited to share that the Hiesinger Lab has been selected for a @StanfordHAI seed grant to support the creation of SuTr, a foundational vision model for #surgicalAI . Get in touch if you’d like to help by contributing surgical videos! Learn more:

0

1

6

1

3

30

@cyrilzakka

Cyril Zakka, MD

3 years

Prompted an AI model to generate pictures of #Beirut in the style of Van Gogh. I'm honestly impressed by the results 🖼 You can see a recognizable mix of the architectural styles seen all over the city - and even some Arabic calligraphy! Made me a little bit nostalgic...

3

4

30

@cyrilzakka

Cyril Zakka, MD

1 year

I made a native macOS application to experiment with the Llama/Alpaca family of models. You can find it here: . Pull-requests are welcome! Demo running in real-time on an M1 Pro:

1

0

30

@cyrilzakka

Cyril Zakka, MD

10 years

Mercury, a fluid quick compose solution for iOS 7! If you want to help me develop it, hit me up with a DM :) http://t.co/9RiGrZcJnt

1

17

26

@cyrilzakka

Cyril Zakka, MD

2 years

A few pictures from around London today.

Tweet media one

Tweet media two

Tweet media three

1

0

29

@cyrilzakka

Cyril Zakka, MD

4 years

Needed a dynamic card animation for an app I’m working on. Shadows and card adjust according to the finger’s position. #SwiftUI

3

0

29

@cyrilzakka

Cyril Zakka, MD

10 months

Had the pleasure of testing out the Apple Vision Pro today. Overall a great experience with many exciting opportunities!

Tweet media one

Tweet media two

Tweet media three

0

1

27

@cyrilzakka

Cyril Zakka, MD

11 months

Been playing around with making my own DICOM viewer/parser from scratch in pure Swift and after fully optimizing it, I'm finally happy with the results. After benchmarking, the Metal impl. was the clear winner with 8x speedup over Accelerate and 18x over a naive impl!

Tweet media one

Tweet media two

4

0

28

@cyrilzakka

Cyril Zakka, MD

4 months

Unsure what the long term vision of these medical scribing companies is, but transcription has become so commoditized you can recreate their core offerings privately and on-device entirely from scratch in about a day. Here’s a demo from almost a year ago.

@cyrilzakka

Cyril Zakka, MD

1 year

Over winter break I decided to make Notos, an #AI -powered transcription application based on @OpenAI ’s Whisper. Here’s a small demo where it manages to transcribe everything flawlessly! TestFlight:

28

48

469

5

3

28

@cyrilzakka

Cyril Zakka, MD

6 months

Excited to be presenting alongside @rohanshad on the state of AI in the operating room at the @StanfordCTSurg Transplantation Summit in Vail, CO! The @HiesingerLab will finally be whole again! Register here:

Tweet card media

Stanford Cardiothoracic Transplantation Summit

The Stanford Cardiothoracic Transplantation Summit three-day symposium will bring together the entire spectrum of the transplant team, including cardiothoracic surgeons, pulmonologists, cardiologis...

events.stanford.edu

0

4

25

@cyrilzakka

Cyril Zakka, MD

9 years

Presenting #EquinoxKeyboard , coming soon to an Appstore near you! http://t.co/gY6byTt3SN

Tweet media one

9

38

23

@cyrilzakka

Cyril Zakka, MD

5 years

Starting to get the hang of SwiftUI I think 😅

2

1

26

@cyrilzakka

Cyril Zakka, MD

2 years

I've been working a lot with with cardiac echos recently and thought I might tweet this out to help out some #AI #ML folks working in #Radiology and #Cardiology . You can get a pixel-perfect mask of the scanning view in as little as 4 lines of code:

Tweet media one

Tweet media two

1

3

26

@cyrilzakka

Cyril Zakka, MD

3 months

I heard it might be dropping next week?

@DahlanAdil

Adil Dahlan

5 months

Another great work by @cyrilzakka 👏 ! Retrieval model > generative model! Cannot wait to see Almanac v2 😎!

0

0

6

1

1

23

@cyrilzakka

Cyril Zakka, MD

29 days

Exciting week! We're also releasing an early preview of MediSyn, a family of text-diffusion models exploring the potential for generative models to synthesize medical data across many modalities (2D and 3D) and specialties (dermatology, radiology, pathology, GI, ophthalmology,

@joseph_cho1

Joseph Cho

29 days

Introducing MediSyn, a pair of text-guided diffusion models for generating high-fidelity and diverse medical 2D and 3D images across medical specialties and imaging modalities. 1/n

Tweet media one

Tweet media two

3

6

24

2

5

25

@cyrilzakka

Cyril Zakka, MD

5 years

Introducing Scribe: Medical Image Annotation Made Easy. 📝 Code available Sep 17, 2019.

Tweet media one

3

6

24

@cyrilzakka

Cyril Zakka, MD

9 months

Medical AGI has been achieved internally with @pranavrajpurkar and @Michael_D_Moor

Tweet media one

2

0

24

@cyrilzakka

Cyril Zakka, MD

2 years

Took the last 2 weeks of 2021 as time off from everything medicine, AI, and research-related to visit my family in London. It’s honestly been a nice break but it’s time to get back to work. Exciting things coming soon! 🤗

1

0

24

@cyrilzakka

Cyril Zakka, MD

2 years

Recently found out @kevin_zakka has a txt file where he keeps a bunch of his favorite quotes. So for Christmas I made Quōtl - a simple app for him to store quotes.

Tweet media one

Tweet media two

@kevin_zakka

Kevin Zakka

3 years

Slightly tangent, but reminds me of this beautiful quote that I motivate myself with every day:

Tweet media one

1

6

55

1

1

23

@cyrilzakka

Cyril Zakka, MD

1 year

@_akhaliq I might be wrong but hasn’t this already been demonstrated with GANs and diffusion models before? Here’s an example in a recent publication by @StanfordAIMI

Tweet media one

3

0

22

@cyrilzakka

Cyril Zakka, MD

4 years

This was terrifying to experience. Hospitals are being overrun with injured people. We need blood donations and disaster relief. Help however you can.

@FadyRoumieh

Fady Roumieh

4 years

Explosion in #Beirut #Lebanon - view from Annahar building. Praying for everyone’s safety 🙏🏼😓

2K

48K

105K

4

5

23

@cyrilzakka

Cyril Zakka, MD

3 years

@Dimillian Bit of a controversial opinion but I hate any and all code examples with ‘foo’ and ‘bar’. It’s always made things harder to understand instead of using more concrete examples

1

1

23

@cyrilzakka

Cyril Zakka, MD

2 years

📝: How Do Vision Transformers Work? () Interesting paper by by @xxxnell that tries to explain the successes and pitfalls behind ViTs. A small thread:

Tweet card media

How Do Vision Transformers Work?

The success of multi-head self-attentions (MSAs) for computer vision is now indisputable. However, little is known about how MSAs work. We present fundamental explanations to help better...

2

4

21

@cyrilzakka

Cyril Zakka, MD

1 year

Transcription happens entirely on-device so it’s perfect for certain sensitive contexts like medical dictation. Next on my todo list: - Optional on-device fine-tuning based on transcription edits. - Ability to add custom fine-tuned models from the @huggingface hub.

1

0

21

@cyrilzakka

Cyril Zakka, MD

1 month

@RoxanaDaneshjou I’m also increasingly worried about non-medical or non-technical people reading too much into ‘medical-adjacent’ preprints that look impressive but have no real use cases and tout it as the next big thing.

3

1

22

@cyrilzakka

Cyril Zakka, MD

3 years

Day made 😍

Tweet media one

1

0

21

@cyrilzakka

Cyril Zakka, MD

2 years

And w/ support from @Gradio , we can now deploy our demos faster than ever. Enjoy generating an unlimited number of mammograms or OCTs, with more models coming soon! 🤱🏻: 👁: Kudos to @abidlabs , @ak92501 for making this possible!

1

6

21

@cyrilzakka

Cyril Zakka, MD

8 months

If you're going to work at the intersection of AI and Medicine, the least you can do is your due diligence and consult someone in the medical field. Here's an autonomous cath lab simulator I came across recently.

2

3

20

@cyrilzakka

Cyril Zakka, MD

9 years

Introducing the Equinox Keyboard Get downloading :D

9

21

17

@cyrilzakka

Cyril Zakka, MD

3 years

@GiorgioPatrini @Miles_Brundage I might be wrong but it seems to me that attention and MLP are just general versions of CNNs and other more task-specific architectures. If you throw enough data at them, they learn the intrinsic biases of the architectures they replace.

2

2

21

@cyrilzakka

Cyril Zakka, MD

4 years

@hardmaru @arxiv_org Love how the arXiv ID is the first few digits of pi

0

0

21

@cyrilzakka

Cyril Zakka, MD

5 years

I’ll release the code soon. I just need to figure out some animation curves and how to the image back to sourceRect after dismissal. Also shoutout to @mecid and @Dimillian - learning a lot from their respective blog posts 😊

@cyrilzakka

Cyril Zakka, MD

5 years

Made a simple image viewer in SwiftUI. Good practice for chaining gestures. Features: 1) zoom 2) rotate 3) double-tap to reset the zoom or switch between content modes 4) long-press for action sheet 5) drag to dismiss or pan

1

5

111

2

1

19

@cyrilzakka

Cyril Zakka, MD

1 year

@jeremyphoward In a way, I welcome this initiative for medical generative AI. The amount of poorly tested models for clinical AI I’ve seen out there is going to cause a lot of harm.

4

1

20

@cyrilzakka

Cyril Zakka, MD

5 days

Apple’s on-device neural network is a base model that uses adaptors (LoRA) that are swapped based on the task being performed. Inference uses grouped query attention, speculative decoding.

2

3

33

@cyrilzakka

Cyril Zakka, MD

2 months

New side quest unlocked: Just soldered my first components to a PCB and have to admit it’s not that much different from wound cauterization

0

0

19

@cyrilzakka

Cyril Zakka, MD

1 month

I've been fascinated by the possibilities opened up by devices like the @Humane pin, but creating a new ecosystem from the ground up doesn't feel like the right move. Instead on app launch, various app services are fetched for display and added to a global context used by the

3

2

18

@cyrilzakka

Cyril Zakka, MD

6 years

@jeremyphoward Just out of curiosity, instead of stochastic gradient descent with restarts, have you tried having an array of learning rates and occasionally picking a larger one with ‘e’ probability? A kind of exploration problem a la RL if you will.

4

1

18

@cyrilzakka

Cyril Zakka, MD

7 months

This work wouldn’t be possible without the constant feedback and guidance of some great folks such as @pranavrajpurkar @Dr_ASChaudhari @RoxanaDaneshjou @rohanshad @jure @Michael_D_Moor and our friends at @huggingface @julien_c @katieelink @wightmanr . You can find out more about

0

1

18

@cyrilzakka

Cyril Zakka, MD

4 years

Definitely going to improve on it in my free time and I’ll release a small blog post/notebook if there’s enough interest. 🤗

4

0

18

@cyrilzakka

Cyril Zakka, MD

9 years

I'm just going to leave this here.. Slowly inching towards release day! /cc @CortexDevTeam http://t.co/F7gsITmyJz

Tweet media one

4

11

17

@cyrilzakka

Cyril Zakka, MD

6 months

Did someone say foundation model for cardiac MRI with zero/few-shot disease detection? Weights out soon! Great work from @HiesingerLab alum @rohanshad

@rohanshad

Rohan Shad, MD

6 months

I’m excited to finally showcase our work on developing a generalizable deep learning system for cardiac MRI. This is the culmination of over 3 years of research spanning my time as a postdoc in @HiesingerLab and now as a cardiac surgery resident at @pennsurgery , and I couldn’t be

Tweet media one

Tweet media two

Tweet media three

Tweet media four

4

18

100

0

3

17

@cyrilzakka

Cyril Zakka, MD

1 year

Seriously considering a CS PhD before residency but every candidate I’ve met is more despondent than your typical medical student 😅

0

0

16

@cyrilzakka

Cyril Zakka, MD

4 months

. @RoxanaDaneshjou and I have been arguing this for a while now. If you have to prompt your model a certain way to get good results on a benchmark, you’re just fitting to the benchmark, making it effectively useless.

@MimansaJ

Mimansa Jaiswal

4 months

@OfirPress I really like this part. One of the hardest parts about releasing benchmarks for me is, someone always responds with "you could have prompted it XYZ way", which feels like some form of fitting on benchmark for me.

Tweet media one

0

0

13

2

2

16

@cyrilzakka

Cyril Zakka, MD

4 months

I can only imagine what will happen with LLMs in healthcare without careful deployment and testing.

@curtlanglotz

Curt Langlotz

4 months

Companies are being held liable for the behavior of their chatbots: "A passenger has won his fight after the airline refused him a retroactive discount, claiming it wasn’t responsible because the promise was made in error by the airline’s online chatbot."

1

2

11

3

0

16

@cyrilzakka

Cyril Zakka, MD

5 years

Preview of the next #SwiftUI Modal update: Add unlimited modals 🔥 Just need to clean up the code and add interactive animations.

2

1

16

@cyrilzakka

Cyril Zakka, MD

3 years

I don’t usually post personal pictures on here but I finally got vaccinated against Covid! #GetVaccinated #Pfizer

Tweet media one

0

0

16

@cyrilzakka

Cyril Zakka, MD

1 year

As part of the @HiesingerLab 's commitment to open-source, along with the democratization of medical machine learning, we are releasing a set of tools to facilitate DICOM anonymization:

Tweet card media

GitHub - cyrilzakka/Phantom: Effective DICOM anonymization in Python

Effective DICOM anonymization in Python. Contribute to cyrilzakka/Phantom development by creating an account on GitHub.

1

6

15

@cyrilzakka

Cyril Zakka, MD

3 years

Watching @DrAlbert using the DaVinci surgical system is humbling.

2

0

16