Eric Horvitz

Here's a demo of Lumiere, the research prototype that inspired Clippy (the Office Assistant): t.co/tHWwZcBhOB As background: The shipped version operated quite differently. The Office team decided to execute the Bayesian inference component only when there was a user-initiated search. So there was no autonomous initiative for the Bayesian component. So where did the infamous Clippy behaviors come from? The activations that folks noted and joked about were based in a separate, legacy system that was hooked up to Clippy--relatively simplistic rule-based tips that were a holdover from the "Tip Wizard" system that had been used in earlier versions of Office, where the rules gently lit up a lightbulb icon on a toolbar. So, there were actually two minds behind Clippy--one Bayesian and only executing when a question was asked, and the other a legacy rule-based system designed originally for a quieter bell icon.

Here's the demo of Lumiere, on the core technology that led the Office Assistant (aka "Clippy"): t.co/tHWwZcBhOB As background: The Microsoft Research prototype inspired the product, but the shipping product was quite different because of challenges with scaling the service. A lightweight variant of the Bayesian inference was used to predict the best assistance to provide based on streams of events and words. And a design choice was made by the Office team: the Bayesian inference component would only come to life when there was a user-initiated search. So there was no autonomous initiative for the Bayesian component. The product team didn't compute the probability that a user needed help, the way we did in the prototype. So where did the infamous Clippy behaviors come from? All of the activation that folks noted and joked about was based in rules firing in a separate, legacy system: the relatively simplistic tips that were a holdover from the "Tip Wizard" system that had been used in earlier versions of Office--where the rules gently lit up a lightbulb icon on a toolbar. A decision was made when building the shipping version of Clippy to connect those legacy rules to Clippy (and friends). So, there were actually two minds behind Clippy--one Bayesian and only executing when a question was asked, and the other a legacy rule-powered system designed for a quieter bell icon.

Nature Digital podcast on clinician-AI collaboration. open.spotify.com/episode/3mta95…

That graphic makes me feel like a member of the Mod Squad. Looking forward to sharing my conversation with @AnnieDuke on the Decision Education podcast. Skills for decision making under uncertainty will only grow with the increasing influence of AI in our lives. I'm deeply supportive of the goals & efforts of the Alliance for Decision Education. @AllDecisionEd

RadFusion: Confidence-based control of clinical report generation t.co/gOEClCQFn6 AI-generated radiology report generation has focused on creating a single, best report. But clinical settings have different costs for misses vs. false alarms. Thus, we may want different “operating points,” that is, different thresholds for when findings should be reported.

RadFusion makes that tradeoff explicit & controllable. It combines the strengths of classification + generation: calibrated classifier scores provide threshold control, a radiology report generator provides clinical detail, and an LLM reconciles the two. Change the thresholds for reporting findings → change the sensitivity/specificity tradeoff → generate a report consistent with the selected operating points. Details: t.co/gOEClCQFn6 Collaboration with Ying Jin, @NoelCodella, John Corring, Mu Wei, Dinei Florencio @Microsoft

Just recorded an exciting session with host @AnnieDuke. Looking forward to sharing the conversation.

Wow. x.com/zachary_horvit…

:-) x.com/Rohit_Writes/s…

High-stakes decisions with language models We introduce 'decision-analytic steering' in this study: t.co/sRpbqujSIX In real-world decisions, different types of errors rarely have equal consequences. So, a system making recommendations under uncertainty must do more than predict what is likely to be true. It must also reflect the relative costs and benefits of acting or failing to act. As an example, in emergency medical triage, missing a patient who needs immediate care is often far more costly than sending someone to the emergency department who ultimately did not need to go. It's critically important to consider that asymmetry. We set out to investigate opportunities for decision-analytic steering by generating ROC curves for families of model. These sample curves showing the true positive and false positive rate for two models on a set of emergency medicine triage cases.

A recent Nature Medicine study provided a striking example: ChatGPT Health was found (Feb, 2026, @NatureMedicine) to have under-triaged more than half of the cases that physicians judged to require immediate emergency care. Was the underlying model simply unable to recognize the emergencies? Our analyses suggest a more nuanced and actionable answer.

Exciting results for AI in mathematics and theoretical computer science @OpenAI openai.com/index/ten-adva…

Nice summary. msn.com/en-us/technolo…

Full editorial: science.org/doi/10.1126/sc…

Congratulations, Wajeeha. Enjoyed our collaborations. Onward! @wajeeha__ahmad x.com/wajeeha__ahmad…

New study on AI and biosecurity: We tested whether updated screening tools can detect short fragments of AI-engineered sequences. The results are encouraging, but also underscore the need for sustained red-teaming, research, and safeguards as AI-biology capabilities advance. Article: t.co/pak563SIkE @IBBIS_bio @Microsoft

Podcast on clinician-AI collaboration study: cassyni.com/events/Cfw9Yuc…

A few years back. @MSFTResearch x.com/hbou/status/20…

@NoamShazeer Noam central!

Test of time awards are the most impressive honors for publications. Longuet-Higgins is the 10yr impact award for CVPR papers. Congrats to the ResNet team. Seems like yesterday. @MSFTResearch x.com/msftresearch/s…

Important finding. x.com/polynoamial/st…

In diagnostic settings, it's important to understand how AI models weigh uncertainties and tradeoffs--for example, the cost of false alarms vs. cost of missed diagnoses. @KhurramYam describes methods for identifying implicit utilities, and for testing whether models can follow x.com/KhurramYam/sta…

Editorial: “A narrowing window to understand AI,” published in this week’s issue of Science science.org/doi/10.1126/sc… @cervisiarius @EPFL @MSFTResearch

Editorial: “A narrowing window to understand AI,”published in this week’s issue of Science science.org/doi/10.1126/sc… @cervisiarius @EPFL @MSFTResearch

Strengthening biosecurity in the era of AI blogs.microsoft.com/on-the-issues/… @MSFTIssues

Patrick Winston’s last lecture on “How to Speak.” In 2019, as I was getting mic’d to give a lecture at MIT, Patrick slowly came down the auditorium steps with a cane. He handed me a copy of his new book, and we caught up briefly. My host later told me he was very ill and had

At Build, we announced the general availability of Microsoft Discovery, a platform for accelerating scientific and engineering R&D. Discovery orchestrates leading-edge scientific models through cycles at the heart of science: evidence → hypotheses → analysis & experimentation

AI and the firm: Empowering organizations to climb the gradient to their own AI frontiers. @satyanadella x.com/satyanadella/s…

On related work & directions with agents: opportunities and prototypes described in this video are still as fresh as they were decades ago. Enjoy! microsoft.com/en-us/research… x.com/erichorvitz/st…