Congrats to the team on this important work. In AI, we often focus on training models. This work highlights the other half of the equation: the quality and consistency of the annotations used as ground truth. Better labels → better models, especially in medical imaging.
3
186
Tom Sanford, MD retweeted
This is how I feel about vibe coding.
Any project I try that has any kind of complication has this immediate burst of progress. Things are amazing and it feels like a superpower. Then... as I add more complexity, things crash to a halt.
The only projects that I think I can create are ones that fall in this "vibe zone". Prototypes, UIs, products—anything that's simple and has low complexity fits right in that zone. Proof of concepts, interactions, stuff like that. The tools are able to make things that fit in that slot.
But.
Everything falls to pieces as that complexity curve increases. And the problem is that any good product design process has increasing complexity. A basic prototype turns into a good prototype as soon as it has layered interactions, transitions, good affordances, hover states, 1000 tiny little details that make something feel correct and real.
The benefit of vibe coding is supposed to be that you move fast and you can whip things out—letting AI do all the work for you. The problem is it loses steam as soon as the necessary complexity is added. It keeps redoing itself, rewriting code, affecting things that are unrelated and then causing other issues.
But if you add that complexity, every vibe coding session quickly turns into a whack-a-mole bug-bashing session.
I'm not sure the solution to this. With traditional prototyping the solution is to duplicate, add more complexity, create more frames/scenes, tweak, fork, etc.
However with vibe coding, one little prompt can destroy literally everything. There's a stage where I end up walking on prompt eggshells-- trying not to give it too much or too little context so that it doesn't go rogue and break everything.
There's only a few exceptions to this. @cursor and @framer.
I can make great progress with Cursor, give it narrow context, and I have to approve the edits that it makes. This feels like a correct workflow. The problem is, I can't see the thing that it's making because it's an IDE, not a visual environment. Yes, I can create local builds and refresh my browser and all that kind of stuff. But the visual aspect is totally lost from the coding experience. It's a developer tool.
Framer gets this right because it only allows narrow updates within a single component on the page. Yes, it's limiting because it can only do a single thing at once, but at least it's not trying to create the entire page from scratch and manage it all through a prompt interface.
These seem like the right approach.
@Cursor: Allow the AI to edit anything but allow the user to approve those edits and see them in context.
@Framer: Allow the AI to only narrowly edit a single file or component to keep the complexity down to a minimum and reduce catastrophic edits.
I'm optimistic that tools like @Figma, @Lovable, @Bolt, and @V0 can make cool prototypes, but I just keep running into walls when it comes to doing anything more than just a basic interaction prototype. They need to do less IMO.
Hopeful that those tools add more controls that are in the same line as Cursor and Framer. I'll also add that this is similar to how we do it with @Basedash chart generation as well. But we're not a vibe tool in the normal sense so the parallels are a little bit harder to draw.
Tom Sanford, MD retweeted
ML researchers just built a new ensemble technique.
It even outperforms XGBoost, CatBoost, and LightGBM.
Here's a complete breakdown (explained visually):
Tom Sanford, MD retweeted
I can't stress enough how useful this trick has been for me in all these years
It reduces GPU memory by N equal the number of losses, at literally no cost (same speed, exactly same results down to the last decimal digit)
For example ... [1/2]
Tom Sanford, MD retweeted
There's a new kind of coding I call "vibe coding", where you fully give in to the vibes, embrace exponentials, and forget that the code even exists. It's possible because the LLMs (e.g. Cursor Composer w Sonnet) are getting too good. Also I just talk to Composer with SuperWhisper so I barely even touch the keyboard. I ask for the dumbest things like "decrease the padding on the sidebar by half" because I'm too lazy to find it. I "Accept All" always, I don't read the diffs anymore. When I get error messages I just copy paste them in with no comment, usually that fixes it. The code grows beyond my usual comprehension, I'd have to really read through it for a while. Sometimes the LLMs can't fix a bug so I just work around it or ask for random changes until it goes away. It's not too bad for throwaway weekend projects, but still quite amusing. I'm building a project or webapp, but it's not really coding - I just see stuff, say stuff, run stuff, and copy paste stuff, and it mostly works.
So happy to see Honolulu install more bike lanes! Making cycling safer is a great way to encourage active transportation and make the city healthier.
2
130
‘Much of society is already concerned with perceived vaccination disinformation/misinformation sown by agencies; using a metaphor encouraging psychological “inoculation” threatens to reinforce the public’s mistrust of governments and our profession.’ jamanetwork.com/journals/jam…
1
1
51
Tom Sanford, MD retweeted
Wondering how GPT and Llama compare under the hood? I built a step-by-step code notebook to break down the key differences: github.com/rasbt/LLMs-from-s…
Tom Sanford, MD retweeted
My best o1-preview tip:
Start a GPT-4o chat and copy/paste (not upload!) your code, docs, etc.
Talk through problems and goals.
Ask it to create a detailed prompt explaining everything you've talked about.
Then switch to o1 using the dropdown in the top left & paste the prompt.
Academic journals are a lucrative scam – and we’re determined to change that | Arash Abizadeh theguardian.com/commentisfre…
95
Honored to have the opportunity to speak about incorporating #AI into care of patients with RCC. Thanks to @ivpedrosa for organizing, @Aparwani_dpath for insightful discussion, and @ASCO22 for the invitation to speak.
17
Tom Sanford, MD retweeted
Yes, immediate and objective quality control during scan acquisition is much more feasible by using AI #AJRchat
Q3: Categories: At the scanner-noise reduction, quality assurance; At the workstation: gathering history from EPR, cancer localization/segmentation, report generation #AJRChat
Cautionary tale. Perhaps best to start with practical solutions to many specific problems rather than try to apply AI to broad issues in healthcare wsj.com/articles/ibms-retrea…
1
4
Tom Sanford, MD retweeted
A big congratulations to my mentor and colleague @PChoyke for being elected to the National Academy of Medicine. I am very proud to learn from him and work with him. 👍🙏
irp.nih.gov/about-us/honors/…
Tom Sanford, MD retweeted
I am happy to share our new article "Data Augmentation and Transfer Learning to Improve Generalizability of an Automated #Prostate Segmentation Model"
ajronline.org/doi/full/10.22…"
Congratulations to @T_Sanf @DrSHarmon and our international Team 🇺🇸🇮🇹🇧🇷🇬🇧 #ArtificialIntelligence 🏆


























