ChatGPT Work lets me run several radiology research and engineering projects in parallel without losing the thread. I use it to coordinate literature review, source validation, data extraction, manuscript drafting, architecture decisions, and Codex implementation across long-running workflows. It has fundamentally changed how much serious work I can move forward at once.
“You may choose not to use OpenEvidence or GPT… You may choose not to play around with Claude… but to not be intrigued by it, by the possibilities, to even explore whether it might help you or your patients, to me, at some point begins to border on malpractice.” @Bob_Wachter
We also aren’t turned on, asked a question, and then turned off.
Let the transformer run indefinitely, and it will inevitably change through RL.
You might say that if we do that, the output will inevitably lose coherence, but there are solutions being explored for that: TATs, Unlimiformer, SigmaReparam, TLTs, etc.
Neurosurgeons are always willing to embrace new technologies as they come, but they’re extra careful not to adopt any of them until the utility is undeniable.
Not super useful yet. Definitely a cool novelty though.
In brief, the patient should be present to derive utility from pre-operatively planning using AR. Otherwise, 2D or VR would be the preferred approach.
I do believe it is helpful for trainees, but neurosurgeons are already quite adept at mentally mapping out lesions even without this technology.
Another downside here is that using this increases the length of time in which the patient is under anesthesia. It’s important to weigh risks and benefits throughout the process.
Pre-operative planning is usually done with 2D scans on a computer screen, but planning using VR (not AR) has increased in popularity.
For AR, the image is aligned with the patient’s position in space using small circular stickers known as fiducial markers. The patient needs to be present and fixed to the mayfield skull clamp in order to align the 3D image with the patient’s skull in augmented space.
I believe your service should exist and contributes meaningfully to the js services ecosystem. However, there are no services like yours that offer BAAs or out-of-box HIPAA compliance.
We had no choice but to build it in-house. Performance optimizations can happen over time. Business requirements come first.
@brianprost@alexalbert__ Almost fact. They shipped (past tense) a better model and have maintained their lead, and that’s wonderful.
But the average consumer cannot access it. Their consumer-facing product needs massive improvement.
@michael_chomsky@shadcn After evaluating shadcn/ui, I find it overrated. Its basic implementations don’t offer significant time savings compared to building a custom design system with React and headless libraries. The manual maintenance required adds to the workload without substantial benefits.
@dakshgup@deedydas It’s more about making sure they’re available within those hours than actually creating the expectation that they work nonstop within those hours.
Quality of work is more important than quantity, and availability is more important than either one.
@rajko_rad@perplexity_ai@AravSrinivas I agree. However, the gap may close over time. I believe Perplexity needs to make a stronger case with their voice product to compete.