Microsoft (United States) • Personal Information Management and User Behavior, Information Retrieval and Search Behavior, Mobile Crowdsensing and Crowdsourcing
Ben and his co-authors call these constraints and preferences “standing instructions” (arxiv.org/pdf/2311.09796), enabling AI models to remember details like your preferred cuisine or budget whenever you ask for restaurant recommendations.
Ben and his co-authors call these constraints and preferences “standing instructions” (arxiv.org/pdf/2311.09796), enabling AI models to remember details like your preferred cuisine or budget whenever you ask for restaurant recommendations.
Personalization is another critical area of Ben’s research. By allowing users to explicitly state their general preferences and building models around them, AI can provide personalized responses and recommendations.
Personalization is another critical area of Ben’s research. By allowing users to explicitly state their general preferences and building models around them, AI can provide personalized responses and recommendations.
Ben’s effort to apply logical reasoning to a particularly complex field – identifying loopholes in the US federal tax code – promises to help people navigate and interpret complex documents in a wide range of domains (www.taxnotes.com/featured-ana...).
Ben’s effort to apply logical reasoning to a particularly complex field – identifying loopholes in the US federal tax code – promises to help people navigate and interpret complex documents in a wide range of domains (www.taxnotes.com/featured-ana...).
In the workplace, addressing complex intents and reasoning often requires making sense of vast amounts of documents and business practices grounded in human interactions, individual artifacts, and evolving knowledge that changes over time.
In the workplace, addressing complex intents and reasoning often requires making sense of vast amounts of documents and business practices grounded in human interactions, individual artifacts, and evolving knowledge that changes over time.
Additionally, his research on decomposing complex intents into smaller steps allows AI systems to handle multi-step tasks while keeping humans in the loop for reasonability checks at every decision point (arxiv.org/abs/2305.08677).
Additionally, his research on decomposing complex intents into smaller steps allows AI systems to handle multi-step tasks while keeping humans in the loop for reasonability checks at every decision point (arxiv.org/abs/2305.08677).
One AI challenge is enabling systems to understand and process complex utterances expressed in NL. Ben’s work involves converting NL into machine-readable formats, which is essential for accurate automated interactions with databases, APIs, and chatbots (aclanthology.org/2024.emnlp-m...).
One AI challenge is enabling systems to understand and process complex utterances expressed in NL. Ben’s work involves converting NL into machine-readable formats, which is essential for accurate automated interactions with databases, APIs, and chatbots (aclanthology.org/2024.emnlp-m...).
How can AI move beyond simple responses and reliably address complex tasks? Meet Benjamin Van Durme (www.microsoft.com/en-us/resear...), who is pushing the boundaries of natural language tools so they can handle multi-step processes with contextual understanding.
How can AI move beyond simple responses and reliably address complex tasks? Meet Benjamin Van Durme (www.microsoft.com/en-us/resear...), who is pushing the boundaries of natural language tools so they can handle multi-step processes with contextual understanding.
As we move into 2025, AI is set to transform all aspects of work, but only if we are able to build on what we already know about productivity while being open to learning. This past year taught us that there are new ways AI can drive purpose, sustain persistence, and foster collaboration.
As we move into 2025, AI is set to transform all aspects of work, but only if we are able to build on what we already know about productivity while being open to learning. This past year taught us that there are new ways AI can drive purpose, sustain persistence, and foster collaboration.
By leveraging AI’s strengths in planning and synthesis, this kind of uncoordinated large-scale collaboration will make it possible to harness collective intelligence, enabling groups to solve problems and create knowledge more effectively than ever before (teevan.org/publications...).
By leveraging AI’s strengths in planning and synthesis, this kind of uncoordinated large-scale collaboration will make it possible to harness collective intelligence, enabling groups to solve problems and create knowledge more effectively than ever before (teevan.org/publications...).
But AI’s ability to process input from multiple people at scale will also disrupt collaboration in more profound ways. We already see people using Excel to extract themes from columns of text, where each row represents an idea from a different person.
But AI’s ability to process input from multiple people at scale will also disrupt collaboration in more profound ways. We already see people using Excel to extract themes from columns of text, where each row represents an idea from a different person.
Microsoft’s decades of experience supporting collaboration, managing permissions, and facilitating sharing are proving invaluable as we integrate AI into these spaces.
Microsoft’s decades of experience supporting collaboration, managing permissions, and facilitating sharing are proving invaluable as we integrate AI into these spaces.
Given the importance of conversations, we're embedding AI into collaborative spaces. One of the best Copilot experiences is in Teams meetings, because AI enhances collaboration by summarizing discussion, suggesting action items, even prompting attendees to engage in more meaningful conversations.
Given the importance of conversations, we're embedding AI into collaborative spaces. One of the best Copilot experiences is in Teams meetings, because AI enhances collaboration by summarizing discussion, suggesting action items, even prompting attendees to engage in more meaningful conversations.
But the nature of knowledge artifacts is evolving. Much of the knowledge people return to is now generated through conversation (w/ people or AI) and reused not only by people but also by AI for grounding. Microsoft was a document company, but now must become a dialog company (vimeo.com/927715632).
But the nature of knowledge artifacts is evolving. Much of the knowledge people return to is now generated through conversation (w/ people or AI) and reused not only by people but also by AI for grounding. Microsoft was a document company, but now must become a dialog company (vimeo.com/927715632).
Currently, most conversations with AI are ephemeral. People must reestablish context every time they return to an AI system. This is at odds with the fact that work takes place over time. Fifty years ago, Microsoft began its journey by helping people create documents to digitally persist knowledge.
Currently, most conversations with AI are ephemeral. People must reestablish context every time they return to an AI system. This is at odds with the fact that work takes place over time. Fifty years ago, Microsoft began its journey by helping people create documents to digitally persist knowledge.
But we are learning that AI can go even further by prompting people to fully articulate their goals and explore new directions. Right now, AI simply obeys commands. Looking forward, it will also challenge people to think differently (cacm.acm.org/opinion/ai-s...).
But we are learning that AI can go even further by prompting people to fully articulate their goals and explore new directions. Right now, AI simply obeys commands. Looking forward, it will also challenge people to think differently (cacm.acm.org/opinion/ai-s...).
People can now directly express their purpose, rather than having to translate it into computer-understandable actions. But expressing purpose is still hard to do. Decades of research show that recognition is easier than recall, and prompt support strategies are proving to be powerful tools here.
People can now directly express their purpose, rather than having to translate it into computer-understandable actions. But expressing purpose is still hard to do. Decades of research show that recognition is easier than recall, and prompt support strategies are proving to be powerful tools here.
Work has always been 1) purposeful, 2) persistent, and 3) collaborative, but AI is fundamentally redefining how computing supports these essential aspects. Let’s look at how.
Work has always been 1) purposeful, 2) persistent, and 3) collaborative, but AI is fundamentally redefining how computing supports these essential aspects. Let’s look at how.
But the early glimpses we got this year of how people are starting to use Copilot suggests AI will teach us new ways to think about productivity as well. If 2024 was the year of taking what we know and applying it to how we use AI at work, 2025 can be the year of taking a fresh look at work given AI
But the early glimpses we got this year of how people are starting to use Copilot suggests AI will teach us new ways to think about productivity as well. If 2024 was the year of taking what we know and applying it to how we use AI at work, 2025 can be the year of taking a fresh look at work given AI
In 2025, Microsoft will turn 50, and the wealth of knowledge the company has built over half a century of pioneering productivity solutions will supercharge how AI helps people summarize email and write documents.
In 2025, Microsoft will turn 50, and the wealth of knowledge the company has built over half a century of pioneering productivity solutions will supercharge how AI helps people summarize email and write documents.
The fact that we’re already observing significant impact is interesting because, honestly, 2024 was also the year of using AI in the most boring ways possible. Right now, people are pretty much just using AI to summarize email threads and write documents for them, if they're using it at all.
The fact that we’re already observing significant impact is interesting because, honestly, 2024 was also the year of using AI in the most boring ways possible. Right now, people are pretty much just using AI to summarize email threads and write documents for them, if they're using it at all.
I know there's a lot there to digest, but don't worry. In the new year I'll continue my annual tradition of walking through the slides. Consider this a sneak peak to get ahead of the game.
I know there's a lot there to digest, but don't worry. In the new year I'll continue my annual tradition of walking through the slides. Consider this a sneak peak to get ahead of the game.
I am at #NeurIPS this week! We are hiring full-time applied scientists in @microsoft.com Office of Applied Research, particularly those pushing the boundaries of model finetuning. Catch me for a chat or drop by Microsoft booth!
I am at #NeurIPS this week! We are hiring full-time applied scientists in @microsoft.com Office of Applied Research, particularly those pushing the boundaries of model finetuning. Catch me for a chat or drop by Microsoft booth!
Hello Bluesky! We're hiring amazing scientists in the productivity division at Microsoft. Come work with me, @teevan.bsky.social, and a whole legion of awesome researchers.
Hello Bluesky! We're hiring amazing scientists in the productivity division at Microsoft. Come work with me, @teevan.bsky.social, and a whole legion of awesome researchers.