Chaitanya Asawa

Early Impact Results from GPT-5.5 We’ve been evaluating @OpenAI’s newest model across Abridge workflows like note generation and clinical decision support. In healthcare AI, it’s not one model, it’s orchestration. Every model is rigorously tested before use. The results include a relative 25% lift in clinical quality coverage and 30% more concise responses.

@HarryStebbings Besides the value of oiling the machine points, important to hire people who "want to play ball", and raise your ambition!

I have never met a top-performing CEO who likes the role of HR. They are here to slow us down and instill meaningless process.

It’s a privilege to help build toward a future where clinical evidence is timely, relevant, and shaped by the conversation at the point of care. Abridge is bringing NEJM and JAMA evidence directly into clinical workflows, right where decisions happen. This adds a new layer of rigor to care delivery, making evidence more accessible, actionable, and aligned with how clinicians actually practice.

Slop = AI without context

Me when all my lab mates are congratulating me on Anthropic deploying our work but actually they didn't cite us...

people's motivations are the fun part of Abridge :)

We conflate mission-driven with "company works in [X] sector" when it's actually more a comment of the people in the company & highest order bit of why they are there

The most valuable companies in the world may soon be enterprise AI firms – not consumer apps. This wasn't obvious a few years ago. For decades, consumer tech always won but AI is quietly rewriting that order.

@kahnfessions do it!! we'll get a real nice whiteboard

Working with this build team has been a career highlight. I have considered moving back to SF just to whiteboard with them.

@IUcordes @CpnpDan @HealthcareAIGuy It's live today if your organization approves it & has UTD licenses!

@CpnpDan @HealthcareAIGuy Usually a 12 month runway from Abridge marketing to go-live 🫠

NEW: Abridge just launched clinical decision support powered by UpToDate. It connects real-time clinical conversations to trusted evidence -- bringing patient-specific context into CDS.

𝗘𝘃𝗶𝗱𝗲𝗻𝗰𝗲, 𝘀𝗵𝗮𝗽𝗲𝗱 𝗯𝘆 𝘆𝗼𝘂𝗿 𝗰𝗼𝗻𝘃𝗲𝗿𝘀𝗮𝘁𝗶𝗼𝗻. Knowledge alone isn’t enough, it needs context. Abridge connects patient dialogue, history, and the clinical moment to surface what matters, right when it’s needed.

@tanayj Yeah here's a couple of past blog posts abridge.com/ai/science-ai-…, e.g. comparing our own voice stack to leading ASR models at the time, and abridge.com/ai/science-con…! We'll mention more publicly in the coming future!

@c_asawa yes apologies! have you released a post with approach/evals publicly by chance?

AI App Layer Companies that have built their own models: • Cursor: Composer 2 • Cognition: SWE 1.6 • Fin: Apex 1.0 • Hippocratic AI: Polaris 3 • Decagon: collection of own models that run 80% of model requests • Sierra: constellation of models some of which are proprietary

@nikillinit Or bundling CDS into a product that provides a wide variety of capabilities, as we're doing at Abridge – not only about not paying extra, but building a more seamless in-workflow product with the context you care about!

are free products that advertised to doctors good or bad? One dimension to think about this is cost vs. utility UpToDate costs $550+/year per physician. OpenEvidence being free means any physician can access AI-assisted evidence lookup regardless of where they practice or how much their employer spends on IT. I think it’s bad to keep stacking fees on top of doctors/providers - you can think of this as a way to shift those expenses to pharma instead. My belief is that getting AI tools in the hands of doctors as a clinical decision assist is important and a net benefit for society. So we should try to reduce the barriers to doing that. If using an advertising based model speeds that up, then that’s probably net good. This also puts competitive pressure on the legacy providers who've been charging $550+/seat for decades. UpToDate now has to justify its price against a free alternative that physicians are voluntarily choosing. That kind of market pressure is healthy and means UpToDate has to provide $550 of value that feels worthwhile.

@emilyzsh lol no wisdom just a random reflection based on a few recent conversations haha

@c_asawa preach chai is spreading wise knowledge for the youth

Early career people often don't know what they want – because they've never worked – so they end up optimizing for a legible thing, most often comp. Over time factors such as mission, culture, company size may increase in value w experience, such that they regret earlier moves.

@rahulgs Propietary data wagmi – models can only go so far without the right data, whether to connect the dots in reasoning re context graphs, or best performance if you train your own

seems obvious but: things that are changing rapidly: 1. context windows 2. intelligence / ability to reason within context 3. performance on any given benchmark 4. cost per token things that are not changing much: 1. humans 2. human behavior, preferences, affinities 3. tools, integrations, infrastructure 4. single core cpu performance therefore, ngmi: 1. "i found this method to cut 15% context" 2. "our method improves retrieval performance 10% by using hybrid search" 3. "our finetuned model is cheaper than opus at this benchmark" 4. "our harness does this better because we invented this multi agent system" 5. "we're building a memory system" 6. "context graphs" 7. "we trained an in house specialized rl model to improve task performance in X benchmark at Y% cost reduction" wagmi: 1. product/ui 3. customer acquisition 4. integrations 5. fast linting, ci, skills, feedback for agents 6. background agent infra to parallelize more work 7. speed up your agent verification loops 8. training your users, connecting to their systems and working with their data, meeting them where they are

@pgasawa Yeah that's fair! There's also an observability analogy. And then I wonder if inference providers actually provide that all to commoditize the complement (these are the new clouds)

@c_asawa Evals can be part of a loop that actually improves the product! It feels more like a monitoring problem to get feedback from the product and turn that into an eval->improvement loop, which has been successfully monetized (albeit maybe trickier).

The test for agent framework & eval companies: python and unit testing, the spiritual precursors, have not been monetized before. Does that change in the AI era?

New waves create new terms and then people make businesses around them... but if you map to the historical analogies, you have to wonder what's different