After attending MWC24 in Las Vegas, I’d like to share some highlights that captured my attention. Compared to MWC24 Barcelona, the Las Vegas edition was a more intimate experience, with smaller crowds and shorter walks between sessions. The convenience allowed me to attend more AI-focused events, which was perfect for my company Parser, our AI-driven software consulting firm. Here are my takeaways from this exciting event:

As one of the most memorable moments, I had the opportunity to talk with the author of two of my favourite books: “Outliers” and “David & Goliath”, Malcolm Gladwell. He was now presenting his latest book “Revenge of the Tipping Point”, which is in my todo list.

1. Livin’ On The Edge – Generative AI

“What happens on the edge stays on the edge.” –Dr. Vinesh Sukumar, Qualcomm Tech

It is not just about AI—it is about Generative AI. When people talk about AI transformation, they are often referring to Generative AI, which uses transformer architectures or other Deep Neural Networks as opposed to traditional Machine Learning techniques. This distinction is critical, especially in a network environment, as Generative AI demands high data capacity networks and computing power to perform inferences rapidly using GPUs or similar hardware.

“You should bring the model to the data, not the other way around.” –Bob Sacunas, UiPath

Imagine a future where your personal assistant is a 3D-rendered virtual human, interacting with you in real-time, responding to your questions while analysing your environment to provide guidance. This requires substantial multimodal data, such as live video streams, real-time audio responses, and immediate inference capabilities—all working together to interact with the physical world through task-oriented agents. To make this vision a reality, robust Quality of Service (QoS) networks are required to handle large data payloads, ensure near real-time responsiveness, and enable edge computing for faster processing. 

“Data should not be moved for AI use.” –Bryan Saftler, Databricks

The business implications are substantial. Moving computation to the edge not only reduces latency and improves the user experience but also creates new revenue streams for telecom providers. Telcos have an opportunity to capture a significant share of what was previously cloud-based inference revenue. By processing data on the edge, as James Kaplan from Meetkai put it: telcos can offer “inference credits” for Generative AI applications running closer to the device, while only transmitting metadata to the cloud. This can be summarised as: “5G networks and the desire for telcos to get a piece of the GenAI revenue generation cake” represent a huge shift in the value chain.

These insights underscore the importance of localising computation. We do not need to send 4K video to the cloud for analysis when the same inference can be done closer to the device, with only the relevant metadata being sent to the cloud. By keeping as much data local as possible, we reduce latency, enhance privacy, and make AI interactions more efficient and responsive.

2. RAN AI Run … – AI powered Software-Defined Networks and RAN

“Software Defined Networks (SDN) & Radio Access Networks (RAN) are evolving, driven by AI.” This was highlighted by Nvidia during a session that explored how the development of SDN parallels that of software-defined personal computers. In this analogy, GPUs are the new hardware backbone for large language models (LLMs), akin to CPUs in the early computing era. Just as compilers and operating systems evolved to create a rich software stack, SDNs are poised to develop sophisticated software layers that will enable diverse services and applications.

Nvidia introduced a compelling concept: the Software-Defined RAN, which integrates seamlessly with AI data centers. The combination is expected to produce numerous digital twins and optimise network functions in real-time. This convergence between AI and telecommunications creates an opportunity for smarter, more efficient systems, enhancing the operator experience.

To be more specific, SDN and RAN can leverage LLMs to elevate automation, intelligence, and ease of management in software-defined environments, including core network and radio components. In an era where 5G and beyond are increasingly complex, AI-driven, self-adapting networks will be crucial to improving network efficiency and simplifying operations. This marks a significant step forward in the journey towards intelligent, adaptive connectivity.

3. Day-0 of Generative AI

“We are on Day-0 of generative AI.” —Ami Badani, Arm

Generative AI is still in its infancy, a promising but largely unrealised potential. We are on “Day-0,” a moment to experiment, learn, and prepare. “AI is like TCP/IP or C++. It can be used in many ways” said Joel Brand from Marvell Technology. It is a general-purpose technology, much like the internet or mobile, and unlike blockchain or the metaverse, its mass adoption is not a question of “if,” but “when.” To be ready for that inevitable leap, we must get comfortable with the unknown—those “dark, turbulent waters” of GenAI adoption.

“AI is industrialising creativity […] It is like an army of interns to help me be creative” –Geoff Hollingworth, Rakuten Symphony.

“You can be a generalist and use generative AI to be a specialist”, said Ami. It gives individuals and teams the capacity to generate ideas, solve problems, and develop solutions at an unprecedented scale. But with all this potential comes responsibility. “Day-0” isn’t just about technological beginnings; it’s also about setting the right foundations, much like cybersecurity’s early days. As Abel Sanchez from MIT once pointed out, “AI and data must adhere to similar principles much as cybersecurity“: controlled access, IP protection, and clear guardrails.

“We must destroy the magic thinking, or bad decisions will be made“, said Geoff. The journey into GenAI is as much about learning as it is about innovating. To truly harness AI, we need to demystify it, train our people, and ensure we’re making informed, rational decisions—not relying on blind faith. After all, “people once prayed for the sun to rise“, said Geoff. Now, we must train our teams to understand and leverage AI effectively, not just hope for good outcomes.

4. Small Models, Big Impact: How SLMs and Model Merging Redefine AI

“Why rent your AI models when you can own them?” –Julien Simon, Arcee.ai. 

Julien referred to Open Source-based small language models (SLMs) and large language models (LLMs), emphasising their potential for companies seeking independence and specialisation. By training these models on your own environment with your proprietary data—the true IP of your company—you can unlock valuable AI applications. In most cases, SLMs can fulfil tasks like language understanding. They can also generate natural language responses without requiring the scale of larger models.

“In the future, AI applications will make decisions on where to run their calculations and inferences, whether on the device, at the edge, or in the cloud.” – said Julien when mentioning that orchestrations of multiple specialised agents will take over more complex tasks, leveraging the unique capabilities of multiple SLMs.

He also introduced an intriguing concept: model merging. This technique allows models to combine their expertise without the need for expensive fine-tuning or pre-training—by averaging the hyper-parameters of two similarly sized and architectural models, you can create a new one that retains the strengths of both. It may sound overly simplistic, but there is significant research and documentation supporting its effectiveness.

“It is not about performance but cost performance. No need to infer in milliseconds to respond to a chat question.” –Julien Simon,  Arcee.ai and former Hugging Face

Julien highlighted a critical point for many use cases: focusing on cost-effective AI solutions that balance performance and expense. His emphasis on cost performance is exemplified by Arcee.ai’s models, which ranks first on Hugging Face’s OpenLLM leaderboard benchmarks, outperforming notable well known proprietary LLMs.  These benchmarks include Qwen 2.1 (5B), Llama 3.1 (8B), and Llama 3.1 (70B), showcasing that small, specialised, cost-efficient models are indeed the future. I’m also eager to dive into the capabilities of the new Supernova.arcee.ai.

5. ROI vs AI: The Shift from Cost-Cutting to Revenue Generation

“When data is huge and too much for a human, GenAI goes in.” –Abhishek Sandhir, Sand Tech.

Nabil rightly pointed out that there are two types of automation: the kind that replaces human work and the kind that does what humans simply can’t. The former is great for cost savings but tends to be a one-off gain. The latter, however, has the potential to unlock ongoing revenue by enabling AI to handle complex, high-impact tasks such as preventing fraud, hyper-personalising campaigns, and analysing cybersecurity logs. These aren’t just about efficiency—they’re about driving new business value.

“Replacing or automating human activities falls into the category of cost savings and has a once in a lifetime, not recurrent, implication.” –Nabil Bukhari, Extreme Networks.

ROI focus upfront: Start AI projects with a clear ROI definition defining KPIs to measure progress. Also, prioritise opportunities for new revenue streams rather than one-time efficiency gains. AI’s value lies in tackling challenges beyond human capacity, creating new revenue opportunities.

6. Beyond Copilots: The rise of Agentic AI

“Most LLMs with RAG are used as glorified search” –Chetan Dube, Amelia & Quant.ai

Chetan added depth to the conversation on the limitations of traditional large language models (LLMs). He emphasised how many speak of AI copilots, but the real future lies in “real pilots.” With this perspective, he introduced a new term: Agentic AI, which is more than just generative AI—it integrates deterministic action, making it a promising approach for serious use cases.

“Using GenAI to transfer $100K to a destination account “most likely number” is not feasible for serious use cases” –Chetan Dube

Agentic AI aims to evolve virtual agents into real digital employees by requiring more precise control compared to generative AI, which bases predictions on the most probable next token. He illustrated this with the prior example about transferring money where deterministic action is crucial. A demo of the Quant Virtual Agent showcased this concept, retrieving information from APIs, navigating documentation, and offering human-like suggestions. While the agent did not actually commit actions such as signing up for a mobile plan, it demonstrated an understanding of the tasks in a compelling way.

Chetan also presented the foundational principles of Agentic AI, acknowledging that grasping these concepts might not be immediate. Here are those principles, accompanied by my software-related interpretations:

• Functional orthogonalisation of lambdas in disjoint sets to mitigate hallucinations: This concept suggests modularising services to keep responsibilities distinct, similar to microservices architecture. Each lambda—representing a function—is responsible for a specific task, like retrieving account information or calculating savings. By minimising overlap, this approach reduces the chances of conflicting outputs or hallucinations.
• Cognitive determinism -> analytic actionable insights: Unlike the probabilistic nature of generative AI, Agentic AI relies on deterministic logic. Ensure consistent output by using deterministic logic, similar to pure functions that always yield the same results for given inputs.
• Asynchronous I/O: This resembles how our brains work—responses often emerge after some time, akin to searching your memory. In Agentic AI, it means using interruption-based responses as I have used to program a z80 8-bit zilog microprocessor.
• Agents correlation, [causation, disambiguation], and arbitration: This principle involves planning tasks, breaking them into subtasks, and coordinating execution to achieve goals. It moves beyond classification to true comprehension, coordinating multiple services like an orchestrator selecting the best path to success.
• Stateful context switching: Like human memory, Agentics AI manages data between short-term (“LLM context” during conversation) and long-term memory (stored in databases like graph or vector DBs). The goal is to build a dynamic mind-map—tracking past and current knowledge for future interactions.
• Sentiment analytics: Going beyond text analysis, this involves recognising facial expressions to infer emotional states, helping the agent better understand user intent (e.g. multimodal LLMs).
• Predictive analytics: This complements sentiment analytics to drive proactive actions, helping prevent issues before they escalate, specially on sensitive user interactions.
• Learning mechanisms: Inductive, Deductive, Empirical, Generative: Chetan explained that Agentics AI learns similarly to humans. Agents develop a mind map of knowledge using inductive learning (training models), deductive learning (business rules), empirical learning (A/B testing), and generative learning (creating new outputs), all feeding into a growing knowledge graph.

Ultimately, if an agent cannot solve a problem by breaking it down into functional components, it will redirect the issue to a human instead of fabricating a response—a crucial feature for trustworthiness and safety.

7. Open Banking for Networks: Will CAMARA Stick?

While Open Banking focuses on data openness to foster financial innovation, Open Gateway/CAMARA is about network capability openness, enabling a new wave of applications that leverage advanced telecom features to deliver superior user experiences in a connected world.

The MWC24 LV discussion about OpenAPIs, Open Gateways, and CAMARA felt like a deja vu with Open Banking, but for telecom networks.

The goal in both cases is to enable a more open and collaborative ecosystem where third-party developers can easily integrate with the core services of traditional providers. However, Open Gateway and CAMARA emphasise networks-as-a-service (NaaS), allowing developers to leverage standardised telecom APIs to facilitate interoperability between service providers. Imagine being able to customise telecom experiences: low-latency gaming, personalised network settings, or enhanced media streaming—this is the promise of these open telecom capabilities. Unlike Open Banking, which shares financial data securely, CAMARA APIs allow developers access to network capabilities like routing data to the nearest edge node, network slicing, using real-time location data, and adjusting the Quality of Service (QoS).

“In essence, CAMARA and Open Gateway create the connective tissue between 5G networks globally and AI applications.” 

This means a huge opportunity for those wanting to create global AI applications that utilise 5G networks. By linking 5G and AI seamlessly, it becomes easier to deploy sophisticated AI solutions on a global scale, with the network reliability and speed needed to unlock the full potential of both 5G and generative AI. The framework allows developers to build applications that adapt to network conditions dynamically, making them ideal for real-time data processing, such as those required by LLM-based services.

Reflecting on the deja vu moment, Matija Razem from Infobip commented “Open Banking arose out of a clear existing demand for secure data sharing. In contrast, Open Gateway/CAMARA APIs are being developed proactively to create an ecosystem where we still need to see if there is a genuine monetisable need”. The future will tell, but the proactive approach suggests a belief in creating the market rather than waiting for it to form.

Wrapping Up

On a personal note, one of the most memorable highlights of my trip was attending the T-Mobile Unconventional Awards celebration. There, I had the opportunity to meet Malcolm Gladwell, one of my favourite authors, whose books like “Outliers” and “David & Goliath” have been highly influential for me. I even managed to get an autograph on his latest book, “Revenge of the Tipping Point,” which I’m eager to read.

MWC24 Las Vegas highlighted the challenges and opportunities facing Generative AI, edge computing, Agentic AI, SLMs and telecom innovation. From integrating AI into telecom networks to leveraging specialised models and embracing Agentic AI, the industry is poised for transformation.