AI Industry - Academia Summit, Where Research Meets Reality

As large language models (LLMs) become more advanced, their ability to handle multiple languages has seen significant advances. However, challenges remain, in particular for languages that are significantly different from English in their structure and characteristics, and that are additionally substantially less resourced. This talk will explore how multilinguality is integrated into LLMs, including the various phases of LLM development pipelines including pre-training, post-training and evaluation. In particular I will present Eclektic, a novel benchmark that aims to uncover how well parametric knowledge is being transferred from one language to another. AdditionalIy, I will highlight recent benchmarks and leaderboards we developed towards assessing the capabilities of LLMs in Semitic Languages, to drive real-world applications and increase the technological reach for the languages of this region.

Add to Calendarcalendarcalendarcalendarcalendar

In the era of data proliferation, there are many domains, where collecting large datasets is simply not feasible. How, then, can we extract meaningful patterns and insights from limited data? This talk explores how machine learning offers powerful tools for analyzing "not-so-deep" datasets.
We will show how studying the geometry of data and comparing statistical distributions - whether derived from images or textual sources - can lead to compelling, statistically significant conclusions. Two case studies will illustrate this approach: the first analyzes ancient handwriting from around 600 BCE to investigate levels of literacy at the time of the Iron age; the second focuses on biblical text attribution, demonstrating how computational methods can uncover linguistic layers, infer authorship, and help disentangle the composite nature of sacred texts.
Join us as we explore how AI can illuminate the ancient world, even when working with sparse data - bridging the gap between cutting-edge technology and long-standing historical questions.

Add to Calendarcalendarcalendarcalendarcalendar

Models are already extremely intelligent and can solve PHD level math problems. Now they need more knowledge (data),  so the so-called math PHD will also know the updated price of the Nike shoe, to apply in the mathematical model. 

Add to Calendarcalendarcalendarcalendarcalendar

Ever wonder why GenAI breaks in production, even with 'proven' models? In this talk, we'll explore the fail points, where they come from, and the systems and organizational habits that enable them.

Despite dramatic gains in model capability, many GenAI initiatives fail as soon as they move beyond the prototype. It's not the model that breaks; it's the environment it’s deployed into. Fragile handoffs, missing feedback loops, and unclear team boundaries create the perfect storm for silent, cascading failures.

The rise of vibecoding, where AI-generated code ships without proper review or integration has made these issues harder to detect and easier to scale. Public misfires from tools like Base44 illustrate how quickly AI can introduce instability when organizational readiness is lacking.

This talk draws on real-world AutonomyAI deployments to surface the most common failure modes in GenAI production. It introduces the PACT model, a practical framework for aligning product intent, architecture, communication, and task clarity. The goal: help teams build systems where GenAI can deliver lasting value, not just short-lived demos.

Add to Calendarcalendarcalendarcalendarcalendar

In this lecture we will introduce and discuss the impact SLMs have on production and under what circumstances you should use them.
We will continue with revealing one of the biggest industry secrets: you can use even much smaller models (~200M params) for most of your use cases, and we will explain when and how.

Add to Calendarcalendarcalendarcalendarcalendar

MCP give superpowers to LLMs, but we can take it further. The story of GitMCP, mcp-ui, and how we shape the future of agentic experiences.

Liad is currently an AI lead in Monday.com, previously led AI in Shopify, and also leading the UI workgroup in the MCP standard committee.

Add to Calendarcalendarcalendarcalendarcalendar

This talk will present a practical roadmap for integrating AI across the physician’s entire career arc—from the MD classroom through residency and into lifelong practice—grounded in a project initiated at the Dina Recanati Medical School in Reichmann University. I outline three horizons: short term (foundational AI literacy in the MD curriculum, simulation based decision support sandboxes, early documentation pilots), mid term (residency workflows such as AI enabled rounds, triage and note generation, competency based assessment, and faculty development), and long term (credentialed physician AI co pilots embedded in a learning health system, continuous outcome monitoring, and maintenance of certification aligned upskilling). Cross-cutting themes include governance, safety, privacy, ethics, and equity by design.

Add to Calendarcalendarcalendarcalendarcalendar

Every engineer building AI agents has faced it: you tweak a prompt, swap a model, or change a RAG setting—only to realize it either made things worse or improved one thing while breaking another. Why does this happen? Because teams are testing just one version of the agent out of the million possible combinations—and hoping it works.

Today's evaluation tools are built for single-point assessment, not the kind of massive multi-dimensional comparison you actually need—sure, they might let you A/B test two prompts or pick between three models, but they can't help you explore hundreds of combinations across cost, latency, and accuracy all at once.

In this talk, we’ll show how a structured approach to testing alternatives can flip the script. Borrowing ideas from multi-objective optimization, we’ll show how our SDK and UI let engineers allocate their testing budget while Traigent intelligently explores high-potential configurations and surfaces the best tradeoffs. You’ll see how this approach leads to 4–7x quality gains and up to 90% cost savings—without blind guesses or manual trial-and-error.

Add to Calendarcalendarcalendarcalendarcalendar

The devil is in the details.
The rapid development of artificial intelligence presents a double challenge: on one hand, it offers unprecedented efficiency; on the other, it may reduce users’ cognitive engagement.

In the first part of the lecture, I will review the work of Dr. Natalia Kasmyna from MIT, who examined how groups of students wrote essays with the help of AI, compared to using a search engine and without any assistance. The temptation to use immediate solutions, combined with a natural human tendency toward laziness and habit preservation, may lead to decreased brain activity and even damage neural connectivity. In the MIT study, students who used ChatGPT produced quick, focused, and impressive results. But brain activity measured by EEG told a different story: direct use of AI reduced brain activity, connectivity between critical regions, and the emotional connection to the written material.

Recent findings indicate a link between unbalanced use of AI and a decline in neurocognitive measures. However, the challenge does not lie in the existence of the technology itself, but in how we integrate it into daily practices and learning-support systems. Thoughtful, challenging use can enhance intelligence, promote rehabilitation processes, and contribute to brain health. In the lecture, I will also review these opposite results, showing that integrating reward mechanisms, game-like elements, and strategies that encourage deep information processing may transform AI from a mere facilitator into a true generator of thought in education.

In the final, main part of the talk, I will show how such use can contribute to healing the brain in cases of degeneration or neurodegenerative diseases such as Alzheimer’s or Parkinson’s. It is possible to slow brain decline associated with aging - to slow it down in early stages, or even stop it in the future, by combining drug and digital interventions. Similar approaches can also improve longevity in healthy aging people.

At Remepy, we developed a cognitive-emotional training app based on scientific and clinical research. The app uses AI to properly adjust the challenge level, with the goal of gently strengthening the brain and supporting neural connectivity. In our studies at Remepy, we showed that within just three weeks, it is possible to conduct personalized brain training that leads not only to improved connectivity but also to enhanced cognitive function, essentially rejuvenating the “brain age.” Additionally, we observed dramatic changes in immune system markers directly related to accelerated aging and dementia.

We will demonstrate that if artificial intelligence is used correctly – it can contribute to brain health!


Related articles and further reading: 

https://www.nature.com/articles/s41746-025-01765-1 

https://www.themarker.com/magazine/2025-08-12/ty-article-magazine/.premium/00000198-6f38-d12c-a9b8-7f39780e0000 

Add to Calendarcalendarcalendarcalendarcalendar

This talk is a dive into how we designed, built, and shipped at Lemonade a production-grade voice agent using OpenAI’s Realtime APIs. I’ll walk through the general architecture and how we combined LLM-based reasoning with deterministic state machines — not just the “what,” but the “why” behind those choices.

More importantly, I’ll share the real-world challenges that don’t show up in diagrams: trade-offs between flexibility and control, managing latency at scale, designing for fast iteration while keeping quality, and deciding when to let the LLM lead vs. when to fall back on structure.

The talk blends technical insights with product and organizational lessons — useful for anyone building or planning to build with generative AI: engineers, product leads, researchers, and tech execs alike.

Add to Calendarcalendarcalendarcalendarcalendar

How do you integrate AI as an additional entity alongside warriors, how does AI reduce the cognitive load of warriors and directly affect what modern warfare with robots looks like, in the presentation I will share operational use cases.

Add to Calendarcalendarcalendarcalendarcalendar

This talk is a dive into how we designed, built, and shipped at Lemonade a production-grade voice agent using OpenAI’s Realtime APIs. I’ll walk through the general architecture and how we combined LLM-based reasoning with deterministic state machines — not just the “what,” but the “why” behind those choices.
More importantly, I’ll share the real-world challenges that don’t show up in diagrams: trade-offs between flexibility and control, managing latency at scale, designing for fast iteration while keeping quality, and deciding when to let the LLM lead vs. when to fall back on structure.
The talk blends technical insights with product and organizational lessons — useful for anyone building or planning to build with generative AI: engineers, product leads, researchers, and tech execs alike.

Add to Calendarcalendarcalendarcalendarcalendar

Five months ago, I left Microsoft’s healthcare team to join the AI Center at Sheba Tel Hashomer. My goal was to get closer to the heart of the healthcare system and experience firsthand what it means to integrate GenAI from within. 

What I quickly discovered is that bringing AI into a hospital from within is every bit as challenging—and often even more challenging—than doing it from the outside. In this talk, I’ll share key lessons learned: some passed down from colleagues who paved the way, and others I learned the hard way myself. 

We’ll explore how to design solutions that are generative yet flexible, how to engage doctors so they annotate data and provide meaningful feedback, and how to ensure AI systems remain explainable and responsible—without compromising user experience or demanding too much of clinicians’ time. 

If you’re curious about what real-world GenAI implementation looks like in one of the most complex environments imaginable, this talk is about as real as it gets.

Add to Calendarcalendarcalendarcalendarcalendar

Don’t Just Read It - Talk to It:
How I Transformed My Book into a GPT-Powered Learning Experience
What happens when a book stops sitting on the shelf – and starts talking to you?

In this talk, I’ll share the unconventional journey of how I took a professional book I wrote about personal branding and networking, and trained ChatGPT on it to turn it into an interactive experience.
Instead of buying and reading the book the “traditional” way, anyone can now ask it open-ended questions and receive personalized answers – including updated guides on branding in the age of AI. I’ll share what I learned from the process – from the technical aspects and copyright challenges to insights about the future of books and writing. This session is perfect for anyone who wants (and needs) to rethink what learning looks like in a world where books don’t have to stay silent.

Add to Calendarcalendarcalendarcalendarcalendar

What if we could move a big chunk of bug fixing and solving production issues to agentic AI? That would be so cool. In this talk we will go through the end to end process of setting up a background agentic workflow that detects production errors, finds their root causes, assesses the right solution and opens a PR - so you wake up in the morning to tasks almost fully completed for you by your loyal agent.

Together we will dive into the entire process of setting up this system that is currently running in real production environments - understanding the different tools, the infra challenges, the agentic accuracy spectrum, and more

Add to Calendarcalendarcalendarcalendarcalendar

This presentation explores the application of advanced Generative AI techniques to identify anomalies between medical summaries and billing records, enabling proactive detection of errors and cost-saving opportunities. By integrating GenAI-powered anomaly detection seamlessly into AS400 operational workflows, healthcare organizations can optimize processes, improve billing accuracy, and drive substantial financial efficiency.

Add to Calendarcalendarcalendarcalendarcalendar

Hadas Lorber is a Senior Lecturer in Public Policy at Reichman University, Director of the Institute for Applied Research in Artificial Intelligence at HIT, and Leads the Israel-US Relations Project at the Institute for National Security Studies (INSS).

With over 25 years of experience in senior positions in Israel's intelligence and national security system, she currently serves as Chair of the Working Group for Formulating an International Strategy on AI within the Expert Forum of the Ministry of Innovation, Science and Technology.

Add to Calendarcalendarcalendarcalendarcalendar

We built an AI system that ingests CVE advisories and patches, and produces tested exploits - all in under 15 minutes for about $1 each. This talk shares the key lessons learned: guiding LLMs through multi-step reasoning, splitting tasks across agents, and using refinement loops with validation to avoid false positives.

Add to Calendarcalendarcalendarcalendarcalendar

This session presents a hybrid cloud approach for developing Generative AI solutions in healthcare, combining access to cutting-edge cloud-based AI models with secure integration to on-premises medical systems. We will explore architectural strategies, connectivity patterns, and privacy-preserving techniques that ensure regulatory compliance while enabling innovation at scale.

Add to Calendarcalendarcalendarcalendarcalendar

The discovery of new drugs is a laborious, lengthy, and costly process, often hindered by low success rates and biological complexity. Recent AI advancements and growing biomedical data availability are driving innovation in drug development.
This talk explores key drug discovery challenges, including integrating multi-modal and multi-domain data. It reviews critical computational problems and highlights how IBM's open source foundation models address these challenges through multi-domain multi-task approaches.
The talk will cover practical AI applications accelerating drug discovery, reducing costs, and improving success rates, with real-world examples of AI transforming pharmaceutical research.

Add to Calendarcalendarcalendarcalendarcalendar

We present a new Hebrew medical language model designed to extract structured clinical timelines from electronic health records, enabling the construction of patient journeys. Our model is based on DictaBERT 2.0 and continually pre-trained on over five million de-identified hospital records. To evaluate its effectiveness, we introduce two new datasets—one from internal medicine and emergency departments, and another from oncology—annotated for event temporal relations. Our results show that our model achieves strong performance on both datasets. We also find that vocabulary adaptation improves token efficiency and that de-identification does not compromise downstream performance, supporting privacy-conscious model development. The model is made available for research use under ethical restrictions.

Add to Calendarcalendarcalendarcalendarcalendar