Alessandro Perilli

Vice President, Research, AI @ IDC. Leading the AI Economics research group.

[Alessandro Perilli @perilli]

A personal update about my role at @IDC: starting today, I lead the newly formed AI Economics research group, reporting to Bob Parker. Our data show that one of the top priorities for end-user organizations across the globe is to better understand the cost structure of their AI initiatives and how to quantify the business value they can expect to derive from them.

IDC has already produced a tremendous amount of research in this area, and we have analysts looking at it from every angle: infrastructure, applications, data, enterprise tooling, and workforce. In fact, I have over 250 slides on the subject that I regularly use in my keynotes.

We are now going to double down on this work and take a more structured approach to help our clients answer some of the most challenging questions about the cost and value of AI.

One of our first research engines will be a monthly pulse study to track month-over-month changes in key dimensions such as budget allocation and spending, cost structure and value realization, pricing models and contract terms, cost management tooling and requirements, and much more.

We will start by examining the economics of agentic AI in enterprises across industries, then expand to explore the economics of physical AI.

I'm very excited to see IDC investing in the research needed to turn the promise of AI into a net positive for the global economy. Our CEO, Lorenzo Larini, and our Chief of Research, Meredith Whalen, have been strong advocates for this initiative.

I want to thank Shari Lava for the tremendous support she has given me during my first seven months at IDC, and my Enterprise AI strategy team for their excellent research on tough topics such as sovereign AI, the penetration of open-weight models in the US economy, and the future of human-AI and AI-AI collaboration.

Oh, and I'm hiring. See a link to my posts on this topic in the first comment below.

[Alessandro Perilli @perilli]

A healthy exercise for company executives: on a business trip, visit a country that bars access to your AI of choice. Not a soft or hard cap. A complete block.

You know those movies that suddenly switch to ultra-slow motion? That’s how you’ll feel. And it will get progressively worse as models become more capable and we become more reliant on them.

And while you experience that slow-mo, remember how hard it is to calculate the business value of AI at the endpoint.

[Alessandro Perilli @perilli]

A remarkable training data set for those who get the assets. Synthetic media generation needs more high quality training data.

[Polymarket @Polymarket]

JUST IN: Getty Images files for Chapter 11 bankruptcy after struggling with mounting debt & growing competition from AI-generated images.

[Alessandro Perilli @perilli]

After my keynote at our @IDC AI & Autonomy Forum in Tokyo, I spent three days visiting some of Japan's largest organizations to show them our latest AI Economics data. Japan is the first place where I've spoken about the economic opportunity associated with sovereign AI, an opportunity that almost no one is really pursuing.

I did so partly because Japan is ahead of many of the other top 10 economic powers in executing its national AI strategy, and partly because we are finally about to publish a report on the subject after months of research (stay tuned, I'll say a lot more soon).

Yet, I didn't get the opportunity to tell the full story in any of these meetings. We have too much data. I'll have my chance during another client presentation next week, when I will be given 60 minutes to talk about AI budget governance strategies and spending trends, cost structures and approaches to value realization for AI initiatives, AI pricing model preferences and established contract terms, AI cost management tool selection and missing capabilities, and more.

OK. Perhaps 60 minutes won't be enough...

[Alessandro Perilli @perilli]

👀

[Doris Tsao @doristsao]

1/ We discovered something totally unexpected about the neural structure of consciousness that overturns decades of thinking on its biological basis and brings us significantly closer to the recipe for generating conscious awareness.
New preprint: https://t.co/K8BwiHXvXg (led by @HesseJanis, with @FrankLanfranchi, @_conor_mcgrory, @Yuelin_Shi, and @EngelTatiana)

[Alessandro Perilli @perilli]

I'm in Tokyo for a long business trip to meet face-to-face with our Japanese clients and support the IDC sales team. During the weekend, I explore the city as I always do when I'm here, but this trip is completely different from all the others. On this trip, I have access to real-time, state-of-the-art, bidirectional, human-level AI voice translation.

Tokyo has changed a lot over the last two decades, and there are now many more places where you can speak English. But, for the most part, the language barrier remains very strong. So whenever I'm in a museum, shop, or restaurant where people can't understand me, I place my phone between my interlocutor and me and say, "Please translate everything I say from English to Japanese and vice versa." And the AI starts acting like a professional interpreter.

Japanese people are bewildered when I do this, and since most of them are professionals who regularly interact with foreigners, you can tell they see new opportunities after a few sentences, as they realize how well this technology works. So, they start asking me what app I'm using, whether I am on a paid plan, which languages it can translate, etc.

I, on the other hand, have the chance to engage in long conversations with people I could never have interacted with before, and learn so much about their culture and society.

Of all the things generative AI has given us, real-time synthetic voice is a gift whose value can't be measured.

[Alessandro Perilli @perilli]

👀

[Pedro Domingos @pmddomingos]

That didn’t take long: the model swallows the harness.
https://t.co/LtE1zQjboc

[Alessandro Perilli @perilli]

But agents can’t discern if a product is actually better. What buying agents
will be able to discern, in an agentic-led commerce world, is which selling agent is providing the most detailed information about the product being evaluated. Today we tend to do the opposite: minimize the cognitive load to avoid analysis paralysis because human attention is finite and fragile, and rather exploit human emotions to trigger the sale. In an agent-to-agent commerce world, there’s no human emotion to exploit and agents can process mountain of information. But richness of information is not a proxy for quality.

For product quality assessment, we’d have to assume that the buying agent is somehow capable of *testing* the product being evaluated (clearly impossible with most categories of products, including most types of enterprise software) and it follows evaluation criteria that resemble human judgement (equally improbable for a long series of reasons). I’m not convinced.

[Alessandro Perilli @perilli]

@wintonARK and yet, you would be surprised by how many end users I talk to have not come to this realization. The impact of agentic AI on competitive dynamics remains an ethereal concept if all your competitors do the same identical things you do.

[Alessandro Perilli @perilli]

What are we looking at here?

[Richard Sutton @RichardSSutton]

The PhD thesis of my 15th PhD student, Shibhansh Dohare (@ShibhanshD), is now available.

Title: Learning Forever using Artificial Neural Networks
Url: https://t.co/Zp2GxZ03aw

Abstract:
In this dissertation, I study the ability of artificial neural networks to learn new things. I refer to this ability as plasticity. Plasticity is a desirable property of intelligent systems in the real world. Systems that interact with the real world can become outdated if they do not learn from the new information revealed to them by the extremely large and partially observable real world. All naturally intelligent systems continually learn new things to keep up with the changes in the world. Many applications like predicting markets or weather, learning human preferences, controlling factories or prosthetics, etc require AI systems that continually learn new things.
I provide direct and thorough demonstrations of loss of plasticity in deep learning systems. The demonstrations range from simple feed-forward networks to deep residual networks and cover a wide range of algorithmic choices. While seeking an understanding of the phenomenon, I found that loss of plasticity occurred simultaneously with hidden units becoming dormant and becoming similar to each other, hampering the ability of the networks to learn new things.
Next, I describe the continual backpropagation algorithm. I designed this algorithm, in collaboration with others, to maintain plasticity in artificial neural networks. Continual backpropagation is a simple extension of the conventional backpropagation algorithm. The conventional backpropagation has two main parts: initialization with small random weights and gradient descent at each step. In addition to gradient descent, continual backpropagation reinitializes a small fraction of units at each step. Continual backpropagation indefinitely maintained plasticity in many continual supervised learning problems.
Finally, I study plasticity of deep reinforcement learning systems. First, I show that plasticity loss has an effect even in stationary reinforcement learning problems. It is unsurprising that plasticity loss affects stationary reinforcement learning because it requires continual learning due to changing policies and bootstrapping. This effect is visible under long-term learning. I found that some commonly used RL algorithms do not scale with data. Their performance drops as they continue interacting with the environment. Plasticity loss hampers the ability of these algorithms to recover and relearn when the performance drops, resulting in agents that do not get better with experience. Continual backpropagation allows these algorithms to recover when performance drops, leading to agents whose performance keeps improving with experience. Second, I show that loss of plasticity is a significant problem in continual reinforcement learning problems, and continual backpropagation maintains plasticity in these problems.

Shibhansh is now a research scientist at Keen Technologies.

[Alessandro Perilli @perilli]

Cognitive bankruptcy incoming. Fast and ultrafast inference for the latest AI models can transform certain processes in specific industries, from media and entertainment to healthcare. But when it comes to AI at the endpoint, it’s wonderful only as long as the execution asymmetry is in your favour.

You can reply to those 600 unread emails in a few hours and get back to inbox zero. Or produce that 20-page client report in 30 minutes instead of three days.

But as more co-workers, clients, and partners embrace the technology, that lovely asymmetry disappears. You suddenly realise that EVERY person you interact with for business can send you a dense 60-slide presentation to memorise before a meeting, a 50-page analysis to peer-review before a deadline, or a 100-dimensional model to test before launch. EVERY day.

This has already started happening without fast or ultrafast inference.

This leads to cognitive bankruptcy because the ability to produce more does not automatically translate into the ability to absorb more.

Unless companies start enforcing very strict rules on the size of their output (emails, documents, presentations, etc.), and the etiquette of output delivery, AI may paradoxically lead to less efficient execution and faster burnout, not the opposite.

[Alessandro Perilli @perilli]

So. We are now at a point where AI agents can continuously observe our computer screens and be directed to make changes with voice commands alone. It will take some time for people to get used to it, but the technology already works. I use it every day, and it speeds up my workflow tremendously. So what's the next possible step? Well, eye tracking, of course.

The ultimate friction remover is a device (eventually embedded in our monitors) that can see exactly where we are looking and use that information to provide much more precise context for our requests. Try to imagine that scenario.

What happens at that point?

If we get to eye-tracking-assisted (+ voice-directed) computer use, productivity will be tied to agentic AI inference speed first and human thinking speed second.

How many windows can you glance at and comment on in one minute?

In this hypothetical scenario, inference speed will be more important than ever. Optimizing GPU speed, model architecture, and many other components of the AI stack will become critical.

This is perhaps part of why we are seeing AI providers remove the ability to explicitly define model classes and reasoning levels in agents that orchestrate our tasks (e.g., OpenAI Dots).

[Alessandro Perilli @perilli]

https://t.co/3sI8A68tte

[Alessandro Perilli @perilli]

I’m hiring for two positions focused on AI Economics: one with an emphasis on financial modelling, the other on industry expertise at the business process level.

For both, I’m looking for candidates who already use agentic AI in their current jobs or daily lives. These positions are not just about writing traditional analyst reports. We are building cool stuff at IDC !

I can promise unparalleled exposure to top decision-makers across industries and geographies.

If you want to understand how the economy works and how it’s being reshaped by AI, reach out. DMs are open.

(Colleagues and friends: if you know someone who could be a good fit, reshare this post or reach out to me in private. Ty!)

[Alessandro Perilli @perilli]

Lots of people are asking what I do at IDC, so let me tell you the plan.

The research program I lead focuses on two opposite ends of a spectrum.

On one side: the newest, most innovative ideas in AI. I won’t look only at the technology layer, but also at how that layer affects business and organizational models, professions, welfare strategies, the economy, and more. Think of this as an incubator program for bold new concepts. We want to capture what’s next in AI before it becomes obvious to everyone.

If and when these ideas become mature enough, we’ll create new vertical research programs to fully explore how the market absorbs the technologies and the models behind them.

On the other side of the spectrum: how large end-user organizations adopt AI, starting with CIOs, CISOs, CEOs, and boards. I will look at how they make decisions, how they perceive the AI opportunity and the market landscape, who is doing something differently, and how those decisions make them successful. I will also look at how they rethink the role of their organizations to address the challenges AI creates.

Think of this as an attempt to capture the playbook of AI winners, rather than focusing only on the challenges posed by new, disruptive technologies.

To move between these two extremes, my team and I will use the invaluable data IDC allows us to collect at a global scale. We can offer a macro view of who is doing what, where, and at what scale. And that view is not limited to North America. My team is in Europe, the Middle East, India, Singapore, and South Korea.

What I described is a very ambitious, monumental mission. Obviously, I can’t do this alone. Beyond my team, I can count on an amazing group of IDC analysts all around the world. I’ll work with all of them to understand how the world is changing and what our clients must do to win in the age of AI.

Isn’t this the best job in the world?

[Alessandro Perilli @perilli]

A personal update about my role at @IDC: starting today, I lead the newly formed AI Economics research group, reporting to Bob Parker. Our data show that one of the top priorities for end-user organizations across the globe is to better understand the cost structure of their AI initiatives and how to quantify the business value they can expect to derive from them.

IDC has already produced a tremendous amount of research in this area, and we have analysts looking at it from every angle: infrastructure, applications, data, enterprise tooling, and workforce. In fact, I have over 250 slides on the subject that I regularly use in my keynotes.

We are now going to double down on this work and take a more structured approach to help our clients answer some of the most challenging questions about the cost and value of AI.

One of our first research engines will be a monthly pulse study to track month-over-month changes in key dimensions such as budget allocation and spending, cost structure and value realization, pricing models and contract terms, cost management tooling and requirements, and much more.

We will start by examining the economics of agentic AI in enterprises across industries, then expand to explore the economics of physical AI.

I'm very excited to see IDC investing in the research needed to turn the promise of AI into a net positive for the global economy. Our CEO, Lorenzo Larini, and our Chief of Research, Meredith Whalen, have been strong advocates for this initiative.

I want to thank Shari Lava for the tremendous support she has given me during my first seven months at IDC, and my Enterprise AI strategy team for their excellent research on tough topics such as sovereign AI, the penetration of open-weight models in the US economy, and the future of human-AI and AI-AI collaboration.

Oh, and I'm hiring. See a link to my posts on this topic in the first comment below.

[Alessandro Perilli @perilli]

I think it's time to start modeling the droid economy, @wintonARK.

[Alessandro Perilli @perilli]

Are Team Collaboration Platforms the Last Stand for Human Interfaces in the Agentic Future?
https://t.co/gmKuqmE9QI

[Alessandro Perilli @perilli]

.@OpenAI launched ChatGPT Space and @blocks launched Buzz. Many others will follow. So, one of the questions on the table for many technology vendors is: going forward, what will be the best venue for human-agent and agent-agent interactions? Will it be team collaboration platforms? Traditional enterprise software? These new agentic AI platforms? Or something else?

And what will that venue do?

To seed the discussion, Daeil Chun and I published a research report titled “Are Team Collaboration Platforms the Last Stand for Human Interfaces in the Agentic Future?” in which we explore five possible scenarios, the forces that could drive change, and the data that could give us some clues. Link in the first comment.

Of course, there’s so much more to ponder here. Vendors across different market segments suddenly have to compete with one another to demonstrate that their offerings can deliver more value than the alternatives. How will they do that?

[Alessandro Perilli @perilli]

The problem with the “bring your own AI subscription” model is that this is not the end of 2023 anymore. People are now paranoid about token consumption.

When you allow a third-party application to use your existing AI allowance, you’re trusting that the developer has done everything possible to use your tokens efficiently. For users to extend that kind of trust, they must be convinced that the value the product will deliver is far greater than the unknown cost of the tokens it will consume. That’s a lot to ask, and a lot to demonstrate.

On the other side of the equation, if developers no longer bear the burden of paying for tokens up front, will they still have the same incentive to write the most cost efficient AI scaffolding possible?

[Alessandro Perilli @perilli]

“Hey ChatGPT/Claude/Gemini/Grok/Muse/whatever, check how many people on GitHub are trying to build an alternative to Photoshop/AutoCAD/MATLAB/whatever, then create a shared repository that combines their work into one solution. Invite each repository owner to join the project.”

Instead of talking about the SaaSpocalypse, we should be asking what the future of software looks like. In particular, where are the opportunities to create new value through commercial software?

Here’s a potential scenario:

In my 26 years in the IT industry, I’ve learned that very, very few technology vendors develop advanced domain expertise and specialize in addressing the needs of vertical industries. The vast majority develop general-purpose software that can later be customized through consulting and integration services to meet the needs of large enterprises.

But what if the software community, working collectively with agentic AI, could suddenly deliver the same value as those general-purpose features at negligible cost? What if, at that point, the next best opportunity to create value would lie in developing features and capabilities that depend on domain knowledge that is exceptionally hard to acquire and maintain?

Just one of the many possible scenarios we should be thinking about.

[Alessandro Perilli @perilli]

Original picture:

[Alessandro Perilli @perilli]

This weekend, I finally decided to build the very first Star Trek set ever produced by LEGO, which is, of course, the USS Enterprise NCC-1701-D from The Next Generation series (the only real Star Trek). The building techniques used in this set are mind-blowing: the designers had to go out of their way to recreate the many curves of the Enterprise-D with LEGO bricks. It really opens your mind to the possibilities still to be explored with this medium and makes you wonder how much more sophisticated LEGO sets will become as AI assists with their development.

To show you the outcome of two people working for eight hours, I took some pictures, but there’s a twist.

When I shoot on my iPhone, I usually use Project Indigo from Adobe Labs. When I opened the app this weekend, I discovered I’d been selected to try new AI-assisted editing capabilities.

Sounded fun. So I went through the pictures I’d just taken, one by one, and asked it to apply a Cinematic effect. I expected a clever filter, similar to the ML Enhance feature in Pixelmator Pro. But no.

Project Indigo changed the camera angle in each picture, completely relit the scene, and added lighting features to the LEGO set. The results are the two pictures you see above. To appreciate how far the AI went without hallucinating too many details, you have to see the original pictures, which I’ve posted in the comments.

Very impressive in terms of both quality and speed.

Now, time for Agentic Sunday.

[Alessandro Perilli @perilli]

Original picture: