Going to take a shot at building a world class AI team out of India ( Preferably Hyderabad ). I'll build publically this time, Follow on to see how it unfolds!
But what happens when you pair the power of claude with the builders of an app that understand the domain nuances... Isn't that the best of both worlds π₯
Claude can clone the interface in 45 minutes. It cannot clone the year of real user behavior baked into your data model, the edge cases your team has already solved, the trust your customers have built with you, or the fine tuned AI layer trained on your specific domain.
A prototype and a production system are not the same thing. Anyone who has shipped both knows this.
The moat was never the code. It was always the insight behind it, the data feeding it and the relationships keeping it alive.
@hiarun02@mrgirish said this at a talk -- " 20 years ago I used to listen to the song Roja on my casette. Today I listen to it on my phone. The need to listen to Roja has never gone away" . I think pinpointed focus on the right problems is an endless pursuit of challenges and opportunity
- Claude co-work observations -
Seeing this more and more now with peers, saying that it doesn't run reliably. Skills just being a routing logic to prompt sheets doesn't seem to be helpful. πΆβπ«οΈ
I'm betting that verticalized versions of claude code with harnesses to ensure determinism and transparency will be a kick-ass opportunity. We have been building Asthra AI with this in mind for US FDA submissions. π¨βπ¬
@elonmusk It needs better tooling for multimodal experiences. I asked it to do research on a website and it did. But when I then asked it to generate an image based on this context it crashed
6. Data is still a moat, and custom datasets built creatively on public submissions and regulations continues to be a compounding asset for us. ( this might need a blog post itself ) β¨
No biology degree. Worked with a university. Reduced his dog's tumor by 75%.
This is what AI-enabled medical Innovation looks like.
https://t.co/rp0Iq72ljU is building to make US FDA approvals faster for the next wave of innovators. 15 submissions in, here's what we've learned
5. When all competitors are saying the same thing it becomes difficult to cut through the noise. The only way ahead is to keep the bar for innovation high and continue to adapt to the improved closed and open source models π
4. It doesn't look like writers themselves will get replaced. in fact, expertise and nuance will continue to be celebrated in this category. we designed asthra to be an execution arm for writers who can bring their directives to our AI and draft instantly π¨βπ¬
3. privacy is a theme even for small biotechs. to build trust with them we have been able to show them how over 90% of data processing happens exclusively within our cloud with SLMs that are domain and tasked tuned for regulatory submissions in life sciences π
2. CMC submissions ( how a medicine is made ) are super closed and really complex at the same time. we have capabilities to do it, but it seems like it's best to do for large organisations, so that we can justify private deployments and ensure their data is safe π
1. The fastest way to build trust has been by doing historical submissions that our clients have already done, so that they can compare it to manual submissions. Got some positive response for this and we are already helping biotechs with active US FDA submissions they have π