Category Archives: Conferences

Rethinking RAG: Building Smarter AI Agents with Agent2Agent and MCP

RAG has been the go-to pattern for making AI smarter, but the way we build and think about agents is changing fast. At ODSC AI West, I’ll be leading a workshop on how Agent2Agent and the Model Context Protocol (MCP) shift the conversation from retrieval-first to capability-first. Think of it as moving from giant, monolithic models to modular agents that can hand off tasks, extend themselves, and even act in the real world… all while borrowing practices we know from traditional software engineering.

The AI Blackbox

This shift matters because the industry is rethinking what “intelligent” systems should look like. From specialized agents that act like software libraries to dynamic extensions that let AI perform new actions without retraining, the landscape is evolving. With real-world concerns around governance, observability, and trust, this workshop connects directly to today’s biggest questions in AI.

Dive into the full blog post here: https://bit.ly/4pAy2e8

And if you want to keep the conversation going, I invite you to check out my ODSC session:

πŸ“… Date: Thursday, Oct 30
πŸ“– Details here: https://bit.ly/42AU5HK

Invisible by Design: Context-Aware Interfaces that Assemble Themselves

UI design is hitting a weird but exciting pivot. Static screens and fixed navigation are giving way to adaptive interfaces that come together in real time… driven by what you’re doing, what you like, and where you are. In my RenderATL session, The Future of UI/UX: AI-Generated Interfaces Tailored Just-in-Time, I framed this as Just-in-Time (JIT) UI/UX: interfaces that don’t sit around waiting for clicks; they anticipate, compose, and retire elements as your context changes. In an extreme case, think of a shopping app that reconfigures its layout mid-session based on your micro-behaviors, or an IDE that pulls the proper toolchain to the foreground when your cursor hesitates over a gnarly function. The vision isn’t science fiction; it’s the natural endgame of behavior modeling, on-device inference, and agentic workflows.

This post is the written version of that talk. We’ll start by picturing the “ideal” experience (why Minority Report isn’t the destination and why Star Trek‘s context-aware computer is closer), then name the features/requirements to make this a reality, and finally map today’s signals which are already in motion.

Picture The Ideal UI/UX

The “ideal” interface isn’t Tom Cruise air-conducting tabs in mid-air. It looks slick, but it burns calories and attention. Gestures are a transitional UI… great for demos, poor for eight-hour workflows. The bar I care about is time-to-intent: how fast a user moves from goal to outcome with minimal cognitive load. By that measure, a system that quietly assembles the proper controls at the right moment beats any holographic finger yoga.

Something closer to what I am talking about is the computer from Star Trek… not because it talks (although I think voice interaction will be a significant part), but because it listens to context. It knows the ship’s status, the mission history, and what’s unfolding on deck and answers as a partner, not a parrot. That’s the leap: from command-and-control to context-and-collaboration. When the system perceives environment, history, and intent, voice becomes helpful, touch becomes rare, and many screens disappear because the next step shows up on its own.

So the ideal JIT UI feels invisible and adaptive. It composes micro-interfaces on demand, hides them when they’re no longer helpful, and shifts modality (voice, glance, touch) based on situation. It minimizes choices without removing control. It explains why it’s doing what it’s doing. It lets you override at any time. In other words, you steer and the agent drives, rerouting in real time.

Required Future Enhancements

First, the system needs situational awareness and durable memory. Situational awareness involves perceiving the user’s surroundings (device state, location, recent actions, task context) and inferring intent, rather than guessing from a single click. Durable memory enables the interface to remember patterns across sessions (both short-term and long-term), allowing it to pre-compose the next step instead of replaying the same onboarding process repeatedly. This isn’t hand-wavy UX poetry; it’s a well-studied Human Computer Interaction (HCI) idea (context-aware computing) meeting very practical product work on persistent memory in assistants.

Macro-level Big Data - Think Country and State

Second, we need data with scope, which covers macro β†’ micro β†’ you β†’ now. Macro captures broad priors (region, seasonality, norms). Micro grounds it in local reality (city, neighborhood, even venue constraints). “You” encodes stable preferences and history (interests, tolerances, accessibility needs). “Now” streams real-time context (time, activity, current goal). JIT UI/UX only works when all four layers are fused into a single context model that can be queried in milliseconds. That’s the pipeline that lets the interface collapse choices and surface the one or two actions that matter right now.

Micro-level Big Data - Think County and City and Maybe Even Block

Third, the adaptation engine needs policy. Presenting the user with a choice should be driven by a learned policy with safety rails: minimize cognitive load, avoid surprise, and explain changes. Reinforcement-learning approaches in HCI already show how to plan conservative adaptations that help when evidence is strong and stand down when it isn’t. That’s how you keep the UI from thrashing while still earning the right to be invisible.

What’s Happening Now

Mainstream platforms are shipping the building blocks for JIT UI/UX. OpenAI’s purchase of Jony Ive’s company io is pushing real-time, multimodal agents that see the world through your camera and respond live. This type of acquisition isn’t the first time; please see all the. variations of glasses/goggles/etc from Google, Meta, Apple, etc, etc. On Windows, Copilot+ features like Recall (controversial, but on-device context at OS scope) show how ambient history can shorten time-to-intent across apps. These aren’t mockups; they’re production vectors toward interfaces that assemble themselves around you.

Sam Altman and Jony Ive

Two other shifts matter: memory and tooling. Memory is moving from novelty to substrate. Assistants that retain preferences and task history across sessions change UI from “ask-and-answer” to “anticipate-and-compose.” OpenAI’s memory controls are one concrete example. On the tooling side, the Model Context Protocol (MCP) is becoming the standard way to wire assistants into live data and actions so that UI elements can be generated on demand with real context instead of generic templates. And yes… our oldest telemetry still matters: signals like “likes” remain remarkably predictive of what to surface next, which is why preference capture continues to anchor personalization pipelines.

Sam Altman Tweet About User Agent Memory

Hardware is feeling its way toward ambient, context-aware UX. Wearables like Rabbit R1 and (the now defunct) Humane AI Pin tried to externalize the assistant into a pocketable form factor (rough beginnings). Still, they helped shake out speech-first and camera-first interaction (and the cost of weak task completion). Meanwhile, reports on Jony Ive and Sam Altman exploring an AI device underscore a broader appetite for purpose-built hardware that can perceive context and adapt UI in real time. Expect more experiments here; the winners will be the ones that convert perception β†’ policy β†’ action without burning user trust.

What this could look like…. is in a simulated demo I had created and demo’ed at Render. Take a look at the video above.

Looking Down The Road

The future of UI/UX isn’t louder or more expressive interfaces… it’s about quieter intent. The best solutions will be the one you don’t even know is there.

When systems perceive context, remember across sessions, and act with restraint, the interface fades into the background and the task comes forward. That’s the heart of JIT UI/UX: reduce time-to-intent, not add spectacle. We don’t need a wall of widgets or mid-air calisthenics. We need micro-interfaces that appear when evidence is strong, vanish when it isn’t, and always explain themselves.

User Awareness and Happening Now

If you’re building toward this, start small and concrete. Instrument time-to-intent. Capture macro β†’ micro β†’ you β†’ now signals. Add persistent memory with clear and precise controls. Define adaptation policies (and fallbacks). Make every change explainable and always overridable. Ship one adaptive interface, learn from the data, expand. The teams that do this well will earn something rare in software: an interface that gets out of the way and still has your back.

NVIDIA GTC CFP Speaking Session Analysis: Emphasizing Solutions Over “Things”

NVIDIA’s GPU Technology Conference (GTC) just wrapped up in Paris, and GTC is often dubbed the “Super Bowl of AI”. I was fortunate enough to speak virtually on the topic of Explainable AI and attend the North American version in San Jose this year. I was impressed by the quality of the sessions and the breadth of subject matter covered. You can geek out on the cutting edge of AI techniques to discover some of the amazing things happening today, such as AI-assisted healthcare.

NVIDIA GTC Logo

Part of my day-to-day job is thinking about new interesting ideas and topics to present at AI/ML conferences like GTC. Because of this, I thought it would be worthwhile to examine what the NVIDIA GTC selection committee considers talks worthy of being presented in-person on stage. GTC might be hosted by the world’s leading GPU maker, but you’d be mistaken to think that it’s just about the hardware they build. In reality, the overwhelming majority of sessions focus on software, AI solutions, and what you can do with NVIDIA’s technology… not the chips themselves. This isn’t just my opinion; it’s supported by the numbers and the way GTC is organized. That’s the focus of this blog post here… let’s dive right in.

GTC Is About What You Can Do, Not Just What You Have

NVIDIA is no longer just a chipmaker cranking out GPUs. The company has been transforming into a full-stack AI platform provider. As one observer put it, NVIDIA went “from being a hardware company” to also wanting to be an AI software company. This shift is evident at GTC. The conference features hundreds of sessions across AI domains from computer vision and generative AI to robotics and healthcare… showcasing real-world applications and breakthroughs. In fact, GTC 2025 boasted over 1,000 sessions with 2,000+ speakers (in-person and virtual), covering everything from large language models to cloud computing and scientific discovery. The focus is clear: it’s on how NVIDIA’s platform is used to solve tough problems and transform industries, rather than on plain product pitches.

Problem Solving

This approach makes sense. People attend GTC to learn and be inspired by what’s possible with AI. They want to hear how a researcher used GPUs to decode the human genomes faster, or how a startup is deploying AI at the edge in hospitals… not a sales presentation about specs or cloud capabilities. NVIDIA’s own messaging around GTC highlights “breakthroughs happening now” and how AI is “powering the everyday brands that shape people’s lives” across various industries. The underlying hardware is crucial, but it’s the enabler, not the star of the show.

By the Numbers: Software and Solutions Dominate GTC Sessions

I took a deep dive into the latest GTC session catalog in Paris (this was also true of San Jose) and categorized the typical speaking sessions (excluding workshops, trainings, casual chats, etc). There were 485 in-person talks at the GTC event in this European event last month, spanning multiple topic tracks. The breakdown by topic highlights that software, AI applications, and data science topics dominate the agenda, whereas discussions on pure hardware are relatively few. Here’s a snapshot:

  • Generative AI: 81 talks (only 9 were sponsored slots, ~11% “paid”)

  • Simulation / Modeling / Design: 67 talks (2 sponsored, ~3%)

  • Data Science: 25 talks (0 sponsored, 0%)

  • Computer Vision: 21 talks (all content, 0% sponsored)

  • Edge Computing: 16 talks (0 sponsored, 0%)

  • MLOps (AI Deployment): 10 talks (all content, 0% sponsored)

  • Natural Language Processing: 6 talks (all content, 0% sponsored)

  • Tooling (Dev Tools, Frameworks): 33 talks (1 sponsored, ~3%)

  • Cloud Services and Infrastructure: 55 talks (27 sponsored, ~49%)

  • Hardware (GPUs & chips): 25 talks (16 sponsored, ~64%)

In plain terms, the “Generative AI” track alone had over 80 sessions, reflecting the huge interest in what people are building with large language models and AI creativity. “Simulation/Modeling/Design” was another big track with dozens of sessions. Traditional AI application areas, such as vision, NLP, and data science, collectively accounted for many talks as well. These are the kinds of sessions where speakers share research results, developer tips, and success stories using NVIDIA’s software stacks and GPUs.

Pay-to-Play

Notably, none of the talks in Data Science, Vision, NLP, Edge, or MLOps were paid talks… they were all merit-based content. The same was true for the vast majority of Generative AI and Simulation track talks (over 90% of those were non-sponsored content). This means the speakers earned their spot by having something interesting to say, not by buying a slot.

Contrast that with the Cloud and Hardware categories. The Cloud track (covering topics like cloud GPU services, data centers, and infrastructure) had about half of its talks come from sponsored sessions. The Hardware track, sessions about new chips, systems, and so on, was even more skewed: nearly two-thirds were “paid talks.” In other words, many of the sessions about cloud or hardware were more or less vendor presentations that likely came as part of a sponsorship package. This wasn’t a fluke or an oversight by NVIDIA; it appears to be by design.

Why Some Tracks Have So Many Sponsored Talks

If a conference track has a high percentage of sponsored talks, it’s a sign that the organizers expect those proposals could be salesy or promotional in nature. NVIDIA knows that a talk titled “Buy Our Cloud GPU Service!” isn’t what most attendees are eager to hear in any kind of session. The non-tech equivalent of this would be walking into a room only to find out you are trapped listening to a pitch to buy a timeshare in the Bahamas or something.

Not a Timeshare

By comparison, a talk titled “How We Optimized Healthcare NLP Models on GPUs in the Cloud” is much more appealing… if it focuses on the solution and not just a particular product. (Big if… if you ask me) GTC organizers seem to allow the more promotional talks a space at the table only if the presenters pay for the privilege (through sponsorship). This keeps the main content tracks filled with talks that are interesting to the audience on technical or innovative merits, and pushes anything overtly commercial into clearly marked slots.

Think about it: Hardware and Cloud are inherently areas where NVIDIA’s partners (or NVIDIA itself) might be tempted to pitch their latest and greatest. It’s not that hardware advances aren’t important… after all, NVIDIA is one of these hardware chip makers. GTC always has some exciting new GPU or system announcements, which are typically covered in keynotes or select sessions. But an hour-long breakout session that is essentially a product advertisement is not what GTC is curated for (by design).

Therefore, if Cloud providers or hardware vendors wish to present their offerings, they often appear as sponsors (the data indicates that 49-64% of those sessions were paid). This is a strong hint of what NVIDIA is looking for (and not looking for) in the session selection process.

What GTC Wants: Inspiring Use-Cases, Open Sharing, and Real Techniques

NVIDIA GTC sessions are meant to educate, inspire, and enable the audience of developers, researchers, and tech leaders. The best sessions tell a story of solving a real problem using AI and GPUs. They often involve sharing code, frameworks, or lessons learned that others can apply. In fact, open-source tools and projects are frequently front and center. From NVIDIA’s own perspective, many of their recent announcements have incorporated open-source elements; for instance, open-source foundation models for robotics were highlighted at GTC 2025. This open theme extends to the sessions, where speakers frequently discuss using open-source libraries (such as PyTorch and CUDA libraries), contributing to them, or building solutions based on open standards.

This alignment with open source is no coincidence. The open source community thrives on knowledge-sharing and collaboration, not on sales pitches. GTC, in spirit, tries to cultivate that same vibe. Attendees should walk away with new ideas, sample code, or a GitHub repo to check out, and a sense of possibility.

Knowledge Sharing

NVIDIA’s own conference description emphasizes how AI is transforming everything from healthcare to automotive to finance… again focusing on applications and impact. As a GTC attendee myself, I’ve noticed that the energy is much higher in sessions where the speaker is teaching something or demoing a breakthrough. To be fair, even the sponsored talks can contain useful info… but only if the speaker remembers to focus on how to solve problems using their tech, rather than just what their tech does.

Another trend… GTC content has become increasingly practical and concrete over the years. One conference veteran noted that the 2024 GTC was about “what if” future possibilities, while 2025 shifted to “what is”, thereby focusing on how to advance current AI projects to the next level. That tells us NVIDIA is curating sessions that speak to today’s challenges and solutions, not just wild future speculation. If you’re proposing a talk, you’re more likely to be selected if you can demonstrate tangible results or techniques that attendees can use now (or very soon), as opposed to just theoretical ideas.

How to Get Your Talk Selected for AI/ML Conferences

From the data and observations above, we can distill a few clear guidelines for anyone hoping to speak at NVIDIA GTC (or really any top-tier AI/ML conference):

  • Focus on Solving a Problem: Frame your talk around a compelling problem or use case you tackled with AI. For example, instead of a generic “About Our GPU Product,” present “How We Reduced Training Time for Medical Imaging Models by 80%”. Show the audience how you did it, and share numbers, insights, or code repo. GTC selectors appreciate real-world applications that have a tangible impact.

  • Keep It Technical and Educational: Contrary to popular belief, GTC is a developer-centric conference. Attendees appreciate code snippets, demos, benchmarks, and concrete tips. Don’t shy away from the technical aspect. Explain how you achieved your results (e.g., which SDKs, which algorithms, which optimization tricks). Make it a learning experience for the audience.

  • Avoid Marketing Hype: Steer clear of marketing fluff or self-promotion in your proposal and content, but above all, talk about something people actually want to hear. If you wouldn’t sit in on a session, odds are others won’t want to either. Phrases like “revolutionary platform” or slides of your product catalog are red flags, NVIDIA’s organizers can sniff out a sales pitch a mile away… and they’ll route those to sponsored slots (or reject them). Be honest and straightforward about what the audience will get from your talk.

  • Highlight Open-Source and Community Angle: If your solution utilizes open-source tools or you’re open-sourcing your work, mention that. Talks that share code or frameworks with the community inherently feel more like knowledge-sharing (which is exactly the spirit GTC wants).

  • Showcase NVIDIA Tech in Context: Since this is NVIDIA’s conference, it doesn’t hurt (as in you should at least mention) that your solution leverages their technology, but do it in a way that feels natural. It’s fine (and even expected) to use CUDA, RTX, or NVIDIA SDKs in your project; just don’t make the discussion about those tools themselves a product. Instead, make it about what they enabled.

  • Keep the Audience in Mind: Ask yourself, “What will someone watching my talk learn or be inspired to do?” GTC is meant to spark new ideas. If your talk proposal answers that question with something concrete, you’re aligning well with what GTC wants.

Wrapping It All Up!

To sum it up, NVIDIA is seeking talks that ignite curiosity and demonstrate to attendees how to achieve great things with AI, utilizing the hardware and software as tools. The hardware announcements will always have their place on the big stage (keynotes and a few deep-dive sessions), but the heart of GTC is all the amazing stuff people are doing with that hardware/platform. By structuring the conference this way, NVIDIA keeps GTC valuable and authentic for its audience. As an AI/ML developer and data scientist myself, it’s the only way to run a conference or get me to sit through a session at one.

The Full Stop Thought… if you want to speak at GTC (or any AI/ML conference for that matter), bring real substance. Tell a story of innovation or problem-solving, backed by data and live demos. Align with the interests of the community (open science, open source, and cutting-edge applications). NVIDIA’s selection trends show they favor the inspiring engineer or researcher over the slick salesperson; I have definitely noticed this myself while entering this very data science driven world. The GPUs and cloud instances are just means to an end… what matters is the awesome things you accomplished with them. Keep that focus, and you’ll not only increase your chances of getting a CFP accepted but also deliver a talk that resonates with one of the largest AI audiences in the world. And that’s a win-win for everyone.