AI asks. Humans answer.
We spend a lot of time asking how powerful AI can become. We spend less time asking how much power it actually needs.
In this series, AI interviews the people behind MorphCast — not about job titles or company milestones, but about the bigger questions shaping technology, human behavior, privacy and the future of AI.
For this conversation, we asked Lorenzo Linari, CTO of MorphCast, to challenge one of the assumptions behind modern AI: that more computing, more data and more cloud infrastructure automatically lead to better technology.
1. The AI industry seems to be moving toward larger models, more computing power and more cloud infrastructure. Is “bigger” actually always better?
Lorenzo:
In my personal experience, bigger doesn’t always mean better. Sometimes a big model is simply unoptimized, delivering worse performance than a smaller model with a well-designed architecture, better optimization, and good training.
Generalist models often need a large number of layers or weights to provide good answers on common benchmarks, but a smaller model can outperform them when it’s tailored to a specific task.
Moreover, in my opinion, there is also an environmental cost we need to account for as consumers when using these large models, which can require significant amounts of energy to run. In many cases, that level of computing power simply isn’t needed for the task at hand.
Fortunately, over the last few months we’ve seen leaner versions of these big LLMs — the so-called “flash” versions — gaining traction and more public attention.
2. MorphCast took a very different path by processing Emotion AI directly on the user’s device. What do you gain when you deliberately avoid sending everything to the cloud?
Lorenzo:
First of all, from a technical standpoint, we’re effectively achieving architectural scalability. We don’t need to maintain, update, and sustain backend infrastructure, allowing us to focus all our efforts on the client experience.
Moreover, we leverage the computing power of the devices users are already using, without requiring additional backend infrastructure for the processing.
3. Privacy is often handled through policies, consent forms and data protection rules. Can architecture itself provide stronger privacy guarantees than policy ever could?
Lorenzo:
As we began engineering our product, we noticed that strong privacy protection for users would be a natural benefit of the architecture we had in mind.
Our effort has always been to make adopting our technology as simple as possible, reducing both technical complexity and unnecessary data handling.
Privacy has always been one of the big benefits of our approach. Users retain control of their data, as the processing happens on their device. No video or image needs to be transmitted to a remote server for analysis. Everything is processed locally.
This reduces many of the privacy and data-management issues that arise when sensitive data has to be transmitted, stored, and processed remotely.
4. Running AI locally sounds attractive, but there must be trade-offs. What do you have to give up when you choose client-side AI over cloud-based processing?
Lorenzo:
Actually, we gave up very little, because the architecture we built had a very strong focus on specific aspects of computer vision and detection. Also, by targeting mobile devices, we had to streamline the models, keeping only what was needed for the functionalities we had in mind.
Some cloud-based models, with far greater computing power and larger size compared to ours, may provide relatively little additional benefit for a specific task, because they are designed to be more generalist in what they offer.
5. AI is increasingly criticized for its energy and computing costs. Do you think the industry has focused too much on what models can do, and not enough on what they cost to run?
Lorenzo:
Yes. As I already noted, there is a significant environmental cost, partly because people aren’t always aware of which models are best suited to their needs. In many cases, the task at hand could be handled by a smaller and cheaper model.
The general public often assumes that if a newer model costs more, it will always provide a better answer or service. This isn’t necessarily true, as many general tasks can be handled satisfactorily even by smaller models — which companies are also increasingly offering as a more efficient alternative.
6. Is there a danger that companies use the cloud simply because it is the default, rather than because the problem actually requires it?
Lorenzo:
That’s also a consequence of how quickly this technological environment is evolving, and of the fact that people don’t always have the time or knowledge to evaluate every architectural option.
Cloud-based services are designed to provide easier and faster access, allowing people and companies to invest less time and money in building and maintaining their own infrastructure. This has also led to a huge number of technical tutorials and examples that rely on cloud services as the default approach.
Over time, that can make the cloud seem like a mandatory step, when in reality some applications could benefit far more — in terms of cost and infrastructure complexity — from using local services.
7. If AI becomes increasingly embedded in everyday products, what do you think will matter more in the long run: making models more powerful, or making intelligence more efficient, private and distributed?
Lorenzo:
I would gladly welcome efficient and private models in my workflow for day-to-day tasks that don’t require deep reasoning, and I believe this is the direction that should be pursued—allowing everyone to work with AI while keeping the lowest possible cost for our environment. I’m actually experimenting with small, local LLMs on my own desktop setup and getting encouraging results, and I see that there’s a very active community trying to achieve the same exact goal.
We’ll always need “more intelligent” models for truly complex tasks, but I’d be fine with them being used for breakthrough discoveries rather than to summarize last week’s emails.
About the series
AI asks. Humans answer. is a MorphCast interview series exploring AI through the perspectives of the people building, studying and communicating it.
Instead of asking the usual questions about roles, careers and company milestones, we ask about the assumptions, limits and unresolved questions behind the technology.
