Palo Alto Networks, Applied Compute and Private AI
With Applied Compute, Palo Alto Networks is betting $4.5 billion on private AI and on open models adapted to enterprise data.

Imprenditore, ricercatore e docente
Fondatore di Algoretico e di altre aziende che progettano modelli e sistemi di intelligenza artificiale, per uso proprio e per chi vuole metterla al lavoro. Fa ricerca su architetture e AI privata, insegna all'università e ha scritto libri per raccontare l'intelligenza artificiale a tutti.
Michele Laurelli is the founder and CEO of Algoretico S.r.l., where he designs AI models and systems with a focus on private infrastructure. He is also co-founder and Head of AI at Scriba AI, working on both research and the development of products and systems for other organisations.
He teaches Artificial Intelligence as an adjunct lecturer, is an Expert at Institut EuropIA and coordinates Mensa Italia’s Artificial Intelligence interest group. His work spans architectures, private AI, scientific research and teaching.
Teaching: TODO — add current universities and courses.

With Applied Compute, Palo Alto Networks is betting $4.5 billion on private AI and on open models adapted to enterprise data.

Nvidia owes its dominance to CUDA, the software platform that since 2006 has turned GPUs into a standard that is hard to walk away from, more than to the quality of its chips.

What Weight-Space Programming, presented by MIT, actually is, how to read its results, and why the idea of fixing a model by operating surgically on its representations, without retraining it, was at the heart of my 2024 paper Brain Surgery.

Skills promise expertise for AI agents, but they remain fragile, insecure prompts that are hard to test, and they are impoverishing the people who are learning to use AI.

Between September and October 2026, AI-written clones of Photoshop and Word appeared, along with tools that connect coding agents to disassemblers and near-daily ports of decompiled video games.

Seven hundred proofs released in a single day by OpenAI, and the harsh reply from the AHM. What does it mean to solve a problem if no one understands the proof?

AI works better than we understand it. Like the steam engine before thermodynamics, it needs to become a science: laws, measurements and verifiable predictions.

Why the conversion of legacy applications is not a matter of linguistic translation but of semantic reconstruction, and how we addressed the issue at Scriba.

The largest qualitative study ever conducted: Anthropic interviews half the world to understand real hopes, fears, and expectations about AI.

At Algoretico, we have developed a framework that does something no Transformer does today: it develops a cognitive personality that evolves with experience.

Artificial intelligence approached with scientific rigor, engineering precision, and human depth.

The Global AI Report by NTT DATA reveals a clear picture: experimentation gives way to the creation of measurable value.

Continuous Latent Reasoning: 16x-128x Semantic Compression and End-to-End Optimization for Next-Generation RAG Systems


Today, the pre-built management system is no longer just an operational limit, but a cognitive constraint that prevents the company from understanding, governing, and evolving its real complexity.

A technical analysis of why the self-attention architecture makes modern LLMs much more than mere \\\\\\\"stochastic parrots.\\\\\\\"

Draghi's warning about European stagnation concerns not only the number of artificial intelligence models developed but also the very structure of the economy and how we use (or do not use) AI in production processes.

Why the real productive revolution does not come from LLMs.

Exploring the critical importance of private AI infrastructure for organizations requiring absolute control, performance, and intellectual property ownership.

How orchestrated AI agents are transforming complex problem-solving through coordinated autonomy and specialized capabilities.

Most organizations don't have labeled datasets. They have processes, constraints, and domain expertise. Here's how to build AI systems that learn from structure, not just examples.

Beyond pattern matching and statistical correlation—exploring what distinguishes true intelligence from sophisticated computation, and why the question matters for how we build AI.

Why my students implement backpropagation by hand, build neural networks from NumPy, and learn to architect systems instead of calling APIs.

Why industrial automation demands AI that runs on-premise, operates without internet connectivity, and makes millisecond decisions in environments where downtime costs millions.

When AI systems trained on AI-generated content degrade over time, losing diversity and capability. Understanding the mechanics of model collapse and architectural solutions that preserve knowledge.

Building Retrieval-Augmented Generation systems that actually understand your organization's knowledge, not just find semantically similar text snippets.

Creativity doesn't emerge from unlimited freedom—it emerges from intelligent navigation of constraints. What this means for building AI systems that generate novel solutions.

The loss landscape of deep networks is high-dimensional, non-convex, and full of local minima. Yet gradient descent finds good solutions anyway. Understanding why reveals fundamental insights about deep learning.
The blog only uses technical cookies. With your permission I also turn on Google Analytics, which counts visits and articles read in aggregate. No ads, no profiling. Your choice also applies to michelelaurelli.it and the book page. The details are in the cookie policy.
You decide what to enable. You can change your mind at any time from “Cookie preferences” at the bottom of the page.
A single cookie, ml_consent, remembers your choice here, on michelelaurelli.it and on leggi.michelelaurelli.it. It does nothing else and is never shared with anyone.
Google Analytics 4 by Google Ireland Ltd. It counts visits and articles read in aggregate, with no advertising signals. Cookies _ga and _ga_D429LPHJH7, lasting up to 2 years.