Eric Horvitz
            
            @erichorvitz
Followers
                29K
              Following
                3K
              Media
                249
              Statuses
                3K
              Chief Scientific Officer, Microsoft
              
              Seattle WA USA
            
            
              
              Joined March 2011
            
            
           Promise & risk of AI in biology: We share in @ScienceMagazine our efforts to identify & mitigate DNA screening vulnerabilities. Grateful to a fabulous team of collaborators, across organizations. 
          
            
            lnkd.in
              This link will take you to a page that’s not on LinkedIn
            
                
                2
              
              
                
                15
              
              
                
                59
              
             Trusting AI in medical applications: Nat'l Academy of Medicine panel on "Building Trust: AI Governance & Ethics in Medicine and Health"  https://t.co/HMEr6AsV0r 
            @theNAMedicine @johnhalamka @DrCaliff_FDA @VarditRavitsky @kbjohnsonmd
          
          
                
                0
              
              
                
                0
              
              
                
                2
              
             Highly recommended presentations & discussions on frontier research projects. @StanfordHAI @Stanford @landay
          
           🎥 Missed this year's Hoffman-Yee Symposium? You can now watch all sessions of the research presentations on our YouTube channel. Dive back into the rich and thought-provoking discussions here:  https://t.co/jIhVtz6PJC 
            
            
                
                0
              
              
                
                1
              
              
                
                5
              
             Can’t beat them? Connect your broker to trade like them. 
          
                
                0
              
              
                
                34
              
              
                
                366
              
             Artificial intelligence is spreading faster than any technology in history. Yet today’s map shows a clear divide. @Microsoft's AI Economy Institute’s new AI Diffusion Report assesses where the world stands today, while laying out the steps to expand access, build skills, and make 
          
                
                37
              
              
                
                91
              
              
                
                420
              
             Agents and economics: Platform for studying worlds that we can expect. Focus: Exploration of two-sided agentic marketplace. 
           🌻 Announcing New Agents + Economics Research from Microsoft! AI agents are starting to shop and buy for us. At the same time, agents are representing and providing customer support on behalf of businesses. We believe that these two sides will soon collide, and... 1/n 
            
                
                1
              
              
                
                0
              
              
                
                4
              
             Introducing the Synthetic Biology Risk Consortium (SBRC), building on collaborations sparked during Paraphrase effort. Work includes def'n of "sequences of concern" to address biosecurity risks of synthetic nucleic acids created via in silico design.  https://t.co/NoSTzpKPaJ 
          
          
                
                0
              
              
                
                1
              
              
                
                6
              
             Promising direction, led by intern @martinagvilas, on the path to learning how to better guide test-time reasoning. Enjoyed working with Martina and a great team of co-authors @MSFTResearch
          
           But not all reasoning traces are equal ⚖️ → some contain productive steps that lead to correct solutions ✅, while others deviate into overthinking, fail to converge, or exhibit inconsistent reasoning patterns ❌ 
          
                
                0
              
              
                
                0
              
              
                
                6
              
             Tandem Training for Language Models:  https://t.co/fiLq7VvN5v  “With an eye on long-term futures, we pursue methods that encourage models to produce solutions that remain intelligible to weaker collaborators.” Intelligibility as robust handoff. @cervisiarius @ashton1anderson
          
           🚨New paper alert! 🚨 Tandem Training for Language Models  https://t.co/Emzcgf1KHx  Actions & thoughts of AI w/ superhuman skills will be hard for humans to follow, undermining human oversight of AI. We propose a new way to make AI produce human-understandable solutions. How?👉🧵 
            
                
                0
              
              
                
                0
              
              
                
                7
              
             Upcoming panel at the @GalienFdn Patient Summit: "From Promise to Practice: Navigating AI's Potential in Enhancing Patient-Centered Care" More here:  https://t.co/uZnDoYJ3U3 
          
          
                
                0
              
              
                
                1
              
              
                
                2
              
             Paraphrasing a protein is like paraphrasing a sentence—the words change, but the meaning holds. Microsoft scientists applied this concept to toxic proteins, revealing a blind spot in DNA screening and leading a cross-sector team to develop a fix.  https://t.co/kKAZwNOsN4 
          
          
                
                1
              
              
                
                10
              
              
                
                30
              
             A pleasure to collaborate with Adam Clore at @idtdna on the Paraphrase project. @MSFTResearch
          
           Recently, IDT Principal Scientist Adam Clore helped test detect AI-designed protein sequences that could be biosecurity threats. The results of the study, led by @Microsoft CSO, @erichorvitz, are now published in @ScienceMagazine. đź’»Read more in MIT @techreview here: 
            
                
                0
              
              
                
                1
              
              
                
                5
              
             Laude Institute call for Moonshots. First phase of moon: Abstracts due October 25.  https://t.co/wu2fKgxgft 
          
          
            
            moonshots.laude.org
              Seed funding and multi-year labs for proven academic researchers using AI to tackle species-level challenges. Apply for Moonshot funding in civic discourse, scientific acceleration, frontline...
            
                
                0
              
              
                
                0
              
              
                
                4
              
             How do we decide what not to share? I explore “information hazards” in this blog--on when openness in science risks misuse. The Paraphrase Project led to a first-of-its-kind publishing model, accepted as a precedent by Science: "When AI Meets Biology" 
          
            
            microsoft.com
              Microsoft researchers reveal a confidential research effort that explored how open-source AI tools could be used to bypass biosecurity checks—and helped create fixes now influencing global standards.
            
                
                0
              
              
                
                5
              
              
                
                25
              
             Some background on the Paraphrase Project, published in Science last week, including some of the behind-the-scenes story, my blog on the disclosure dilemma, and a conversation with Bruce Wittmann, Tessa Alexanian, and James Diggans. @msftresearch
          
           Microsoft researchers reveal a confidential research effort that explored how open-source AI tools could be used to bypass biosecurity checks—and helped create fixes now influencing global standards.  https://t.co/QV56g8vCb7 
            
            
                
                6
              
              
                
                14
              
              
                
                86
              
             Full paper available at  https://t.co/AG2ULosTYK. 
          
          
            
            erichorvitz.com
              Strengthening nucleic acid biosecurity screening against generative protein design tools.
             Promise & risk of AI in biology: We share in @ScienceMagazine our efforts to identify & mitigate DNA screening vulnerabilities. Grateful to a fabulous team of collaborators, across organizations. 
          
                
                0
              
              
                
                0
              
              
                
                2
              
             The work continues on with advancing AI + biology responsibly, Much more to do. We need to stay on it. @MSFTResearch @TwistBioscience @idtdna @SecureDNA @Battelle @LLNL_SciTech @RTX_News @aclid @NIST @NTI_WMD @StanfordHAI
          
          
                
                0
              
              
                
                0
              
              
                
                1
              
             Our solution: a novel perpetual access framework—a non-profit org (@IBBIS_bio) serves as gatekeeper, with data store & process supported by an endowment, ensuring responsible access for scientists in perpetuity. Grateful this approach was accepted by @ScienceMagazine editors. 
          
                
                1
              
              
                
                0
              
              
                
                3
              
             In a critical phase of the project, we weighed the “information hazard” of sharing technical details. We faced the tension of supporting scientific progress in AI & biosecurity while guarding against misuse. We drew input from experts in gov’t, academia, industry, nonprofits. 
          
                
                1
              
              
                
                0
              
              
                
                1
              
             Guided by failures to detect, brought to light by intensive AI red-teaming, we developed “patches.” The vulnerabilities and mitigations were shared confidentially with DNA synthesis companies worldwide via the Int’l Gene Synthesis Consortium (IGSC). 
          
                
                1
              
              
                
                0
              
              
                
                0
              
             We explored how AI protein-design tools can “paraphrase” toxic & viral proteins—reformulating amino acid sequences while aiming to preserve structure and potential function. We found that paraphrased proteins could fly through DNA synthesis screening tools. 
          
                
                1
              
              
                
                0
              
              
                
                0