Santiago Romero-Brufau

Asst. Prof., machine learning applied to health informatics. Mayo Clinic, Harvard School of Public Health. Views my own. Retweet not endorsement.

[Santiago Romero-Brufau @SantiagoRB]

@IsaacKing314 The version I had in mind was geared to thinking about ML prediction models, it was a version of this graph:

Where the user can move a few things:
- model accuracy (separates the distributions)
- prevalence
- classification threshold

And on the side it‘d show ROC and PR curves

[Santiago Romero-Brufau @SantiagoRB]

@Jayson_Marwaha @surginformatics @npjDigitalMed @HarvardDBMI @BIDMCSurgery @bratogram @GUH_WHCSurgRes @landmaad @willjg @DBatesSafety @jkvedar We will be doing a course this summer.

[Harvard Chan Executive and Continuing Education @HarvardChanECPE]

Revamp healthcare with our Implementing Health Care AI into Clinical Practice program !
Gain skills to integrate AI seamlessly and enhance patient care. Perfect for clinicians and leaders ready to innovate.
Learn more: https://t.co/3bufTM3IyL
#HealthcareInnovation #AI

[Santiago Romero-Brufau @SantiagoRB]

Cone join our course this August to learn how to implement AI solutions into clinical practice, besides building a cool machine-learning model.

[Harvard Chan Executive and Continuing Education @HarvardChanECPE]

Revamp healthcare with our Implementing Health Care AI into Clinical Practice program !
Gain skills to integrate AI seamlessly and enhance patient care. Perfect for clinicians and leaders ready to innovate.
Learn more: https://t.co/3bufTM3IyL
#HealthcareInnovation #AI

[Santiago Romero-Brufau @SantiagoRB]

Agree with this. LLMs+prompting are generally so much better than what most users can train from scratch. For most of our projects at @MayoClinicENT that require NLP, we’re mostly planning to use LLM APIs, or fine-tune open-source LLMs.

[Andrej Karpathy @karpathy]

I think this is mostly right.
- LLMs created a whole new layer of abstraction and profession.
- I've so far called this role "Prompt Engineer" but agree it is misleading. It's not just prompting alone, there's a lot of glue code/infra around it. Maybe "AI Engineer" is ~usable, though it takes something a bit too specific and makes it a bit too broad.
- ML people train algorithms/networks, usually from scratch, usually at lower capability.
- LLM training is becoming sufficently different from ML because of its systems-heavy workloads, and is also splitting off into a new kind of role, focused on very large scale training of transformers on supercomputers.
- In numbers, there's probably going to be significantly more AI Engineers than there are ML engineers / LLM engineers.
- One can be quite successful in this role without ever training anything.
- I don't fully follow the Software 1.0/2.0 framing. Software 3.0 (imo ~prompting LLMs) is amusing because prompts are human-designed "code", but in English, and interpreted by an LLM (itself now a Software 2.0 artifact). AI Engineers simultaneously program in all 3 paradigms. It's a bit 😵‍💫

[Santiago Romero-Brufau @SantiagoRB]

Yoshua Bengio (Turing awardee with LeCun and Hinton) speaks out in favor of regulation of AI agents.

[w̸͕͂͂a̷͔̗͐t̴̙͗e̵̬̔̕r̴̰̓̊m̵͙͖̓̽a̵̢̗̓͒r̸̲̽ķ̷͔́͝ @anthrupad]

summary of Yoshua Bengio article on AI safety:

He thinks instrumental convergence can potentially result in AIs exhibiting dangerous behaviors and that we can avoid this if we ban agents and only make question-answering 'STEM' AI's.

article:
https://t.co/xOq5ra7ZoD

[Santiago Romero-Brufau @SantiagoRB]

@sullyj3 @robertwiblin To be fair to my past self, I hadn’t given it much thought because most doom scenarios I had ran into were of the type “what if they become sentient and start wanting X?” Eliezer’s post made me realize that you don’t need that. It can all be a natural consequence of optimization.

[Santiago Romero-Brufau @SantiagoRB]

@sullyj3 @robertwiblin Very clearly. It was this:
https://t.co/GoSnNTbEbt

I hadn’t considered orthogonality, instrumental convergence, self-improvement as an instrumental goal, and how many ways there are to get power+control with just “intelligence”.

[Santiago Romero-Brufau @SantiagoRB]

@kareem_carr I completely agree. I think there’s a big difference between teaching practitioners and teaching the future generations of methodologists/statisticians, and those are often confused.

[Santiago Romero-Brufau @SantiagoRB]

@kareem_carr Alternative interpretation: performance is great for most datasets “in a lab setting”, but it’s easier to keep applying the lab approach to weirder “problems” than to deal with the problems of models for the real world.

[Santiago Romero-Brufau @SantiagoRB]

@JorgeGalindo Excelente tabla resumen. La tenéis en inglés?

[Santiago Romero-Brufau @SantiagoRB]

@diegoagusan @Egocrata sí pero a través de que la cultura es distinta. Este tipo, con el mismo machete, presentando el mismo riesgo, en EEUU la poli le hubiera disparado “por si acaso”. Y nadie habría dicho que era uso excesivo de fuerza.

[Santiago Romero-Brufau @SantiagoRB]

@Egocrata pues yo me alegro de que una situación así termine sin un tiro

[Santiago Romero-Brufau @SantiagoRB]

@jfalbertos y sin embargo, ninguna de esas coaliciones alcanza el 50%

[Santiago Romero-Brufau @SantiagoRB]

@jfalbertos Basado en tu propio artículo yo diría que no le va a salir muy bien: dedicas tres frases a “la agenda social”, y la mayoría del artículo se refiere a la cuestión nacional.

[Santiago Romero-Brufau @SantiagoRB]

@Egocrata no podían proponer cosas útiles, como una reforma laboral?

[Santiago Romero-Brufau @SantiagoRB]

@Mercedesvsm @Genko Los mossos detienen a la gente y no le han contestado ni se le han mandado el mail a los que están haciendo la solicitud.

[Santiago Romero-Brufau @SantiagoRB]

@DannyBate4 It’s reviews all the way down.

[Santiago Romero-Brufau @SantiagoRB]

@jordievole Dirán lo que quieran, pero el conseguir una entrevista así es importante, y muy interesante verla. Enhorabuena.