Nasrin Mostafazadeh

I love Europe! Not only because the entire European economy is driven by consumer retail brands, but also because they operate under quite different set of constraints than US. All of this makes their approach towards enterprise AI adoption much more nuanced than just hyped-up “tokenmaxxing”! I’m thrilled to join the stage of @HumanXCo in Amsterdam at the end of this summer to talk about what it takes to 𝙩𝙪𝙧𝙣 𝘼𝙪𝙜𝙢𝙚𝙣𝙩𝙚𝙙 𝙄𝙣𝙩𝙚𝙡𝙡𝙞𝙜𝙚𝙣𝙘𝙚 (𝘼𝙄) 𝙞𝙣𝙩𝙤 𝙊𝙧𝙜𝙖𝙣𝙞𝙯𝙖𝙩𝙞𝙤𝙣𝙖𝙡 𝙄𝙣𝙩𝙚𝙡𝙡𝙞𝙜𝙚𝙣𝙘𝙚 across Europe. Thanks to @stefanweitz , Andrew Bolwell, and @jonathanweiner and the entire #HumanX team for such thoughtful curations they put together per usual!

The last three weeks have been a dizzying blur. Honestly, I’m struggling to find the right words, but I am so touched by an overwhelming number of private messages by so many people that I wanted to make sure I write about it publicly. If this is how we, as Iranian-Americans far from the war zone, feel—I cannot fathom the weight on millions of innocent people like my parents living in Iran. Nothing broke my heart more than my father telling me that during the Persian New Year’s spring solstice, he couldn't distinguish the sound of the traditional New Year cannon fire (the "ball drop") from the sound of the bombings in Tehran! I am finding myself losing hope in our collective humanity when we become comfortable treating human beings whom we are not related to as "collateral damage." The minute we dehumanize people as "they," we begin to view casualties as "statistics" rather than "lives" with "stories." In business settings—which are ultimately about human beings— whenever someone asks me for one piece of advice, I always say: "Have empathy." The past few weeks have been a grave reminder of how desperately our world seems to lack empathy as we enter endless wars, one after another. For over 3,000 years, the Persian New Year, Nowruz, has stood as a symbol of rebirth and the triumph of the softness of spring over the harshness of winter. It is a beautiful celebration of a rich heritage that has survived wars, empires, and tragedy year over year. May this spring bring us the collective empathy, and all the rejuvenation that Nowruz promises. We all desperately need it. 🕊️ p.s.: It felt almost impossible to celebrate Nowruz this year. Here is a throwback to happier years when Omid and I got to celebrate Nowruz at the The Metropolitan Museum of Art in NYC that supports the preservation and acquisition of ancient Iranian art and cultural artifacts.

I had a blast stopping by #WebSummit in Lisbon last week. I got to spend time with so many inspiring technologists and journalists, and my favorite moments were the times when we finally stopped talking about #AI, haha. I seriously would have never imagined that our niche field of science "Dialogue Agents" would one day become the topic of everyday conversations everywhere. I loved the title of our panel conversation "Agents of Change", which beautifully summarizes how AI Agents will be fundamentally changing the world (of Enterprise SaaS) as we know it. With the term "AI Agent" being slabbed on top of anything and everything these days, the main question I keep getting asked is what AI Agents even are... Listen to my conversations with Marcelo Gripa Timur Göreci on this and more to come.

Seeing all the reactions to Anthropic's new "computer use" functionality made me think about what we historically used to call *AI Agent* before it became this hyped up marketing term that people slab on top of anything these days. Back in 2007, James Allen and his colleagues at University of Rochester / IHMC, won the best paper award at Association for the Advancement of Artificial Intelligence (AAAI) for their first of its kind "Agent" which they called "PLOW Agent". PLOW would learn to control your browser and execute on tasks such as finding the nearest restaurants using tools, all during a realtime collaborative session with a human who demonstrates it for the first time. Back then, I was a high-schooler who had just started researching commonsense reasoning (which was touted the "holy grail of AI"), so I found the following demo of PLOW reasoning in action quite mesmerizing. Fast forward, both Omid and myself joined James' research lab called "Dialogue Systems", with the goal of building AI Agents that can execute tasks via a goal-directed dialogue -- which is very much what we still pursue at our company Verneek! I believe that one of the main reasons that the public finds AI scary, jumping from "it controls my browser" to "it will control the humanity", is that we lose sight of grounding our progress in the historical context of the the compounded decades of research in AI. The truth is that none of this tremendous progress has happened overnight. Behaviorally, there are so many parallels between what Anthropic's "computer use" does and what "PLOW" did more than 18 years ago. They both very much "control" a browser and perform textual and visual understanding and reason about the task at hand to achieve set goals by taking actions in an environment. What has been truly a game changer is the general nature of the technology that powers today's AI. The fact that we can capture multimodal understanding, knowledge representation, reasoning, dialogue management, planning, and much more, all through one (deep neural network) system is still very mind blowing to me and many scientists -- specially those of us that come from James' school of thought that started in symbolic reasoning before the deep learning revolution. There couldn't be a more exciting time be to alive to do what we do! Kudos to the Anthropic team. You can watch the rest of James' talk via this archived Microsoft Research link: microsoft.com/en-us/research…

So much of the fear about AI has been fueled by the chatter on how AI has these so-called "emergent behaviors" that no one anticipated or built it for. As someone who spent my academic life on Commonsense Reasoning, which is by many accounts the holy grail of language understanding, I take issues with such characterizations. I keep reminding people that decade(s) before LMs became this Large, the best techniques we had for world/causal modeling and script learning were actually language models. All my years of research showed that predicting what happens next in a sequence of events, using LMs trained on the web-scale text corpora, was indeed the best way to predict complicated knowledge structures about the world! This is also why at Verneek we look at LMs as our intermediary reasoning engines grounded on top of environments, i.e., #ContextualRetrievalGeneration. So not much is quite "shocking" on how LLMs are so good at building representations/structures of language. It is so easy to feed on the hyperbolic takes on "AI getting out of control" if we don't look at the historical context. All the doomsday fear-mongering merely distracts all of us from both the real harms and the real benefits of AI, as #AugmentedIntelligence. Listen to some of my conversations at #CollisionConf on the same topics.

I'll be at #CollisionConf in Toronto for a hot minute! Reach out if you want to meet up there.

As someone who spent my academic life on Commonsense Reasoning, which is by many accounts the holy grail of language understanding, I take issues with the clickbaity headline that *LLMs do jaw-dropping things but nobody knows exactly why*! I keep reminding people that decade(s) before LMs became this Large, the best techniques we had for world/causal modeling and script learning were language models. Turns out predicting what happens next in a sequence of events, using LMs trained on the web-scale corpora, was the best way to predict complicated knowledge structures. So not much is quite "shocking" on how LLMs are so good at building representations/structures of language, as @chrmanning also says. @techreview @strwbilly headlines such as this get spun out of proportion for easily feeding the hyperbolic takes on "AI getting out of control", which distracts society from both the real harms and the real benefits of AI.

When we started Verneek three years ago, #AI was still the niche academic bubble we lived in, where we could see that the technology was finally working! We've had tectonic shifts in the public AI landscape ever since, but our founding thesis has stayed the same: Unlike the mainstream, we believed in deploying “Domain-general Intelligence, One Domain At a Time!”. We believe that "The best path forward for building the most helpful AI for the benefit of society is to build AIs with deep domain expertise". We have now helped millions of consumers, through partnerships with various enterprises, all served through our helpful AI expert instances developed using our #ContextualRetrievalGeneration technology. Three years ago, we believed that the time was ripe for human language to become the new holistic interface for interacting with a lot of our everyday environments in the consumer space, and we couldn't be more right about what the deep technology requirements for enterprise adoption should be! We are growing and #hiring across AI research and engineering roles. Join us! Apply here verneek.com

I'm looking forward to delivering the keynote at this year's @WhartonAnlytcs at @Wharton. I'll be following the lead of the trailblazing businesswomen @michelleapeluso, Kate Johnson, and @goldenaimee1 who delivered the last three keynotes, and this year's will be all about AI: "From Pattern Recognition to Cognition: What AI Means for the Society".

The news of @nytimes suing Microsoft and OpenAI over copyright infringement is no surprise and well-warranted. The discussions around the role of content creators in the generative AI ecosystem have been heating up, specially with using LLMs as consumer-facing search engines that do not cite their sources. Since @verneekai's inception, we have always believed that the main breakthrough of LLMs is how they give us a basis for everyday commonsense reasoning, where they should sit on top of (proprietary) data with its provenance intact, as opposed to acting as a large database memorizing everything, through what we call #ContextualRetrievalGeneration. With our business model of delivering purpose-built AI with deep domain expertise to consumer-facing businesses, we have found our AI Platform is a perfect fit for serving media companies that are getting left out of the generative AI equation. As Jill Koziol, the founder of Motherly® serving 40M users said, through Verneek AI any media company gets to "very purposefully launch our own AI, ensuring we own our expert content IP, protecting it and our data from being scraped by Large Language Models (LLM), while delivering greater value to our audience”. I am excited for us to get to play a small role in the monetization and distribution of media content for our media partners in 2024, without whom the entire AI ecosystem would not exist. cnbc.com/2023/12/27/new-york…

So, 7 years ago, at a meeting I had with @satyanadella discussing AI, he asked me “why is that your generation does not want to work for @Microsoft? Be candid”, I replied “well, Microsoft is not a cool innovative place!” I deeply admire what Satya has done in the past few years for completely overhauling Microsoft’s image, ranging from acquiring GitHub to now OpenAI, what a victory lap! what a legendary CEO!

Our society LOVES to elevate individuals to heroes, give them so much power, and then bring them all the way down…A lot of people who went crazy over the weekend for OpenAI’s governance debacle fail to realize that the recent progress in AI is neither made by one company nor by one person. Instead of wasting hours following the OpenAI shenanigans blow by blow, go rewatch the Imitation Game, to get inspired by the true hero scientists of our field!

I've been opening every single one of my AI talks for the past 12 years with this super simple yet unsolved coreference resolution problem. This year, finally, GPT-4 (and not ChatGPT!) could correctly answer this basic commonsense reasoning question, so to celebrate, this was my costume this Halloween... I'm free now!

Today in what could be a historic move in tech policy, the White House issued an executive order on "Safe, Secure, and Trustworthy AI". As an AI company with "#helpful and #trustworthy" right next to all of our AI solutions, I have talked about #AI regulation and all the pros and cons of it for a thriving tech ecosystem that works towards the benefit of society. I always said that the government has a lot of catching up to do for building its own expertise on AI, without outsourcing it to the existing tech monopolies, so I'm glad to see further involvement of @NIST in setting standards for rigorous testing, which can establish a deeper understanding of AI research and development inside the government. To get a better sense of what NIST can do, go read more about @EllaTabassi's work. Let's also be clear that the handful of companies that have lobbied to get AI regulations in place are the ones that will financially benefit the most from it; namely the ones seeking "AGI"!! I have always believed that the the way to bring domain-general AI to the masses, safely and in an actually helpful way, is to deploy it one domain at a time; which is why at Verneek we purpose-build AI models with deep domain expertise, with inherent guardrails in place! The #AI regulations should *not* apply to inherently guardrailed technologies. On perhaps what could be the most positive outcome for the longevity of the future tech industry, I hope that the government will truly facilitate the visa process for having highly skilled AI talent to study and stay in the US, as promised in the executive order. More than 65% of the AI founders in the US are immigrants, with a majority of us having wasted a ton of time and resources away from our research and instead spent it on jumping through the most brutal immigration hoops to get where we are. The US immigration policy for AI talent needs to change drastically. whitehouse.gov/briefing-room…

It brings me so much joy to welcome @MotherlyMedia to our Verneek AI Platform's ecosystem. I always talked about leaving academia for startups to make a real impact by building something from scratch that can actually one day help millions of people in their day-to-day lives. I'm thrilled that our AI can now help all the 40M engaged moms of Motherly make better and faster decisions! I was incredibly impressed by @JillKoziol's vision for bringing a purpose-built AI to their millions of users, that will seamlessly interact with the moms based on the curated content from Motherly’s trustworthy and expert-driven content -- which is exactly what our Verneek AI Platform enables building for any businesses. This is just the beginning of our partnership and I couldn't be more excited about all the upcoming features that we will be rolling out to make #QuinAI more and more helpful every day.