Gradient Institute

Supporting Gradient Institute’s mission to build safety, ethics, accountability and transparency into AI systems: developing new algorithms, training organisations operating AI systems and providing technical guidance for AI policy development.

Principal AI Specialist

Honest AI

In an era where technology evolves unprecedentedly fast, staying educated and appreciating ethical boundaries is more crucial than ever. I desire to bring you "Honest AI" to make sense of the complex world of artificial intelligence, ethics, and the future of work. Understand what’s genuine—honest!— vs. the hype. 

Honest AI is a publication that includes short essays, blog posts, a podcast and a newsletter. 

By subscribing to Human Override (the weekly Honest AI newsletter), you become part of a movement towards responsible innovation and ethical awareness in the digital age.

Founder

Archimedes

The most exclusive members club for LinkedIn thought leaders.

Member

The Ethics of AI: Facts, Fictions and Forecasts

Author

Around

Interim Head of Data & Finance Analytics

Advisor, data science

ARK Venture Studio

Scope opportunities for addressing AI Risk and contributing to the Complex AI Risk Think Tank. AI is a great technology. The risk management industry has the historic opportunity to play a pivotal role in enabling its safe adoption.

Head of Opportunity Scoping

Tesla

Data Science, Tesla Insurance

FUTURELY

Advisor, Product Management and Strategy

Genade Inc.

Genade was a nice learning experience :)

We built digital products to help insurance agents, brokers, wholesalers to automate customer acquisition, lead qualification tasks, and the quoting process. Some examples were interactive web apps that run prospects' fact-finding, qualify leads, pre-populate ACORD forms, and organize information in a super simple CRM. Insurance agents and brokers who used our tools saw a significant increase in time savings and could manage twice as many customers.

We were honored to win the support of:
• 2018 - 19 NYU 300K Entrepreneurs Challenge (Semi-finalist)
• 2019 NYU Summer Launchpad
• 2019 NYU Summer Sprint
• Ignite Fellowship Beta

Unfortunately, the COVID-19 pandemic introduced team and sales dynamics that made our quest difficult to drive forward.

Genade's story ended in December 2020. Many challenges in the world of insurance agents' productivity, lead management, and customer acquisition remain unsolved, and we failed to create a sustainable platform for running an insurance agency on autopilot.

Founder & CEO

Miranda srl - Miranda Bot

Co-Founder, Advisor

AI can be both technically brilliant and used to build a deeply human future.

I've built AI systems from the ground up, written "The Ethics of AI," and advised hundreds of leaders navigating this transformation. I've learned most teams get lost chasing shiny tech while forgetting the humans who actually use it.

I see everywhere this pattern:

- AI projects that sound impressive but fail in practice
- Teams paralyzed by ethical concerns
- Leaders afraid they're falling behind
- Products that work on paper but frustrate real users

My focus is simple: helping you build or use AI that actually serves people. That works for you, and not against you.

Not AI for AI's sake. Not algorithms that impress other engineers. AI that solves real problems and creates genuine value.

I work with founders and leaders who want to:

☑ Build products users actually love
☑ Navigate ethical challenges with confidence
☑ Create sustainable competitive advantages
☑ Sleep well knowing their AI makes the world better

If you're tired of AI hype and want practical guidance on building technology that matters, let's connect.

Join 12,000 people for daily posts on ethical AI, entrepreneurship, future of work. And weekly drops on honest, practical ways to live with AI 
👉 https://honestai.substack.com/s/human-override

LOL (I can’t stop thinking about this!)
We’re calling it the next industrial revolution.
Hackers are calling it opportunity:

Most people I talk with tell me the MS Copilot suite of products sucks, but it’s the only one allowed because it's “secured.”

Is it?

At DEFCON, Copilot Studio agents got shredded.

A handful of prompts was all it took for researchers to:

– Dump full CRM records
– Expose private tools
– Trigger actions without approval

The “autonomous agents” everyone is rushing to deploy turned into data exfiltration machines, no guardrails, no brakes.

And the irony is several vendors keep pitching “no human in the loop” as a feature.

In reality, it’s a hacker’s dream: 
One exploit = complete system compromise.

This time it was Salesforce records, billing info, and internal comms.

Next time? Could be worse.

Here’s the memo for enterprises:

Autonomous AI without serious security is malpractice.

We’re widening the gap between capability and governance… and hackers are the only ones winning.

Autonomous is cool.

Until your agent is working for someone else.

(credit: Michael Bargury on X + Eduardo Ordax for making me discover this story!)

♻ Repost so your CTO do their homework when selecting which AI vendor to trust.

📩 I explore honest observations about staying human while everything turns robotic on my free newsletter. Get your Human Override: https://lnkd.in/eCuGB-SJ

Dr. Alberto Chierici

3x founder & AI specialist helping startups scale fast, build trust, and stay true from day one.

AI is easy to copy. 
These 4 things aren’t:

 Clear thinking.
 Data.
 Partnership.
 Distribution.
(That order matters.)

Everything else is just noise.

Agree or disagree—what do you think is the most under-valued moat today?

♻️ Pass this forward if you care about building real moats.

–––
📩 I write honest reflections about staying human in a world turning robotic.
 Want more? My newsletter Human Override drops weekly (link in comments).
➕ Follow Dr. Alberto Chierici for more like this.

AI is not the moat

The #1 reason 90% of founders fail? They lie to themselves.
How to avoid the 5 most common fatal mistakes:

I caught up with a friend. A 23-year-old founder who burned through $2.3M in 8 months.

Zero revenue.
Zero validation.
And I’ve seen this story play out far too often.

Here’s the brutal truth about why 90% of first-time founders fail (and how the 10% who succeed think differently):

❌ FATAL MISTAKE #1: Building in secret
“They’ll copy my idea”
✅ Successful founders: Talk to 100+ customers before writing code
❌ Failed founders: Spend 12 months chasing “the perfect product” alone
💡 Reality check: Your assumptions are always wrong

❌ FATAL MISTAKE #2: The feature trap
“Users will love options”
✅ Successful founders: Ship ugly MVPs that solve one problem perfectly
❌ Failed founders: Add 47 features nobody asked for
💡 Reality check: Complexity kills conversions

❌ FATAL MISTAKE #3: Hiring friends, firing slowly
“My roommate believes in the vision”
✅ Successful founders: Hire slow, fire fast, pay for results
❌ Failed founders: Hand equity to drinking buddies
💡 Reality check: Friendship ≠ competence

❌ FATAL MISTAKE #4: The validation delusion
“People told me it’s a great idea”
✅ Successful founders: Measure behavior, not opinions
❌ Failed founders: Count compliments as validation
💡 Reality check: Credit cards don’t lie

❌ FATAL MISTAKE #5: Playing CEO too early
“I need a title and business cards”
✅ Successful founders: Do the unglamorous work first
❌ Failed founders: Pretend to be Steve Jobs
💡 Reality check: You haven’t earned it yet

I’ve made every single one of these mistakes at different stages.
The times I succeeded? I did the opposite.

The patterns are clear:
 - Obsess over customers, not competitors
 - Measure everything tied to money
 - Fire fast, hire slow
 - Ship unsexy products to real users
 - Ask “what could kill us this week?” every week

🚀 Your 48-Hour Challenge:
 - Message 10 potential customers today
 - Ask about their biggest problem (not your solution)
 - Record their exact words
 - Build only what they’ll pay for

Stop planning. Start validating.

👉 Which mistake stings the most when you read it?
(If you can’t admit it publicly, you’re already losing.)

♻️ Repost if you want fewer founders to crash and burn

💾 Save this for when the road gets tough

👥 Tag a founder who needs to hear this brutal truth

Reading books won’t make you an expert.

This will:

The fastest progress I ever made, in machine learning, building startups, even writing a book, came from doing the opposite.

Later, I discovered there’s a word for it: Ultralearning.

Benjamin Franklin, Richard Feynman, Elon Musk… they all followed the same playbook.

Scott H. Young codified 9 principles in his book Ultralearning. I tested them in real life:

I learned sales by cold-calling 50 prospects.
I mastered recursion by drilling the same exercises over and over.
I became a writer by publishing consistently, not by studying textbooks.

⚡The principle: Don’t just consume knowledge. Confront reality.

I turned the 9 principles into a swipeable carousel.

👉 Which one resonates most with you? Drop it in the comments.

♻ Share this with someone trying to learn faster.

Learning to learn 

Small words. Big difference. 14 communication shifts that save relationships.
(They work for managers and parents.)

Parenting is hard.
( As well as educating, leading, managing, etc.) 

 But I learned the problem often isn’t the child
 …it’s how I communicate in heated moments.

Here are 14 swaps that change everything:

❌ “Well done!”
 ✅ “I like how you solved that.”
 → Specific feedback builds motivation instead of praise dependency.

❌ “Get up for school or you’ll be late!”
 ✅ “Good morning! It’s time to wake up—today we’re going to…”
 → Exciting starts make transitions smoother.

❌ “Go brush your teeth!”
 ✅ “It’s time to clean our teeth—so they stay strong.”
 → Explaining why increases cooperation.

❌ “You still haven’t tied your shoes!?”
 ✅ “Please tie your shoes. We leave in 5 minutes.”
 → Clear instructions beat sarcasm.

❌ “Don’t you dare talk to me that way!”
 ✅ “I don’t like when you…”
 → Calm expression models emotional regulation.

(…and 9 more examples below 👇)

Each tiny shift builds better relationships.
 At home. At work. With anyone.

Want the full 14 examples in a fridge-friendly printable PDF?
 Comment “LIST” and I’ll send it over.

♻️ Repost if you know someone who needs clearer conversations.

And if you want more: join 2,500+ readers for Human Override, a weekly framework to keep your humanity sharp while tech gets sharper.
 📩 https://lnkd.in/eCuGB-SJ

That difference changes everything:

Traditional security protects data.
AI security protects decisions.

Some alarming statistics I just learned:

—
64% of organizations worry about data integrity attacks.
Adversaries can poison AI models, injecting bias or false information.
→ The danger it’s corrupted decision-making across an entire organization (on top of new, creative ways of stealin records).

—
57% of companies question whether they can trust their AI systems.
When AI makes autonomous decisions with sensitive data, this risk acquires a whole new weight.
→ It’s a business risk that can erode customer confidence and regulatory compliance.

—
Organizations are racing to deploy AI, but:

73% invest in AI-specific security tools AFTER deployment.
Security is bolted on, not built in.
→ Competitive pressure is outpacing security readiness.

—
Most organizations rely on five or more tools for data discovery and classification while 57% juggle five or more encryption key managers.
Fragmentation creates gaps, misconfigurations, and higher risk of exposing sensitive AI training data.

—
My take: AI security can’t be treated as “just another cyber layer.”
It requires a foundation built from day one — or organizations risk embedding vulnerabilities directly into their AI systems.

♻ Share for your team, approaching AI security differently from traditional cybersecurity.

Source (Kiteworks, I’m an affiliate, but only for platforms and tools I trust and learn from): https://lnkd.in/eXdJTdmM

Let me take you back to where this started: 
our work Slack channel called #ai-news was blowing up. 

Too many links. Too many hot takes. 
So we created a quieter space: #ai-discourse-australia. 

Lately, there’s been a lot of talk in Australia about AI approaches and local companies launching “sovereign AI” projects. 

The government is trying to decide: should we make strict rules for AI, or let it grow freely so it helps productivity? 

I posted some thoughts here a few weeks ago, and it kicked off great conversations. 

Why does this matter? If you’re leading a company, shaping policy, or just trying to use AI responsibly, this debate affects you. 

Should we build our own AI systems? Or keep using tools built elsewhere? It’s a hard question. I wrote a few considerations in my newsletter issue👇

On the owning vs. renting tradeoffs SIGNAL Australia’s Computer Society wants the country to invest $2–4 billion in sovereign AI. Meanwhile, Switzerla

Open article: Shall you homebrew your AI? by Dr. Alberto Chierici • 4 min read

Stop engineering prompts. Start defining problems.
I call it “vibe prompting:”

Few weeks ago, I watched a friend spend 47 minutes building a “mega-prompt.”
14 frameworks nested inside each other.
The bot was so confused it literally apologized 😆

It’s been since June I’ve had this hunch that the latest models don’t need much prompting.

So I asked him to try something radical:
“Help me reduce customer churn. Here’s our data. What patterns do you see?”

8 seconds later → clear analysis. Actionable insights.

Then ask the bot to ask you clarifying questions, follow the vibe, see where the conversation leads you.
I often add this:

“[your prompt]
Answer as a well-informed 1st‑person deep search of the web. Then critique as a 3rd‑person analyst (expert on the matter). Finally, merge both to share your conclusion, and ask me any clarifying questions as needed.”

Here’s what I noticed:
* Simple prompts: 90% usable first try
* Complex prompts: 30% needed rewrites
* Output quality: asking to iterate and review the first answer too much makes things worse

If you’re coding AI applications using LLMs as an engine, you still need to structure good prompts. But for that, there is the Claude Console tool I’m baffled nobody uses as much as they should!

Go here: https://lnkd.in/edZc84xM → Click on "Generate a prompt" →Thank me later.
(Not an affiliate, but hey, Anthropic, you should pay me for that!)

Have you ever found adding complexity to an AI request actually made it better? Or does simplicity win for you too?


Let the era of #VibePrompting begin 📣

Everyone expected AI to replace developers. 
Here’s what’s really happening:

A new study reveals where the gaps are, and Cal Newport just broke it down brilliantly in his latest Deep Questions episode.

The big takeaways:
 - Why AI-generated code still struggles in real-world scenarios
 - The mismatch between expectations and reality
 - Cal’s perspective on where we go from here

It’s a fascinating read if you care about the future of work and technology.

👉 Have you tried AI coding tools yet? How close (or far) are they from replacing human developers?

Why AI assisted work is failing us, and how to fix that SIGNAL The AI boost that wasn’t METR recently published a study on AI coding assistants (tools

Open article: 🧠 Deeper Than Code by Dr. Alberto Chierici • 5 min read

This submission is truly great!

Non-standard technology calls for non-standard approaches in scientific understanding, oversight, leadership, and in building citizen literacy and skills. 

Embracing proactive strategies prevents unwanted externalities and bridges competitive gaps (that already exist and are probably widening in Australia).

Read on 👇

Gradient Institute has responded to the Productivity Commission's interim report on 'Harnessing data and digital technology', arguing that its proposed regulatory approach to AI applies conventional governance principles to a fundamentally unconventional technology that may lead to a paradigm shift in how society operates.

The Commission's approach can work well for traditional technologies – including some forms of AI. However, drawing from the International AI Safety Report 2025, we argue that it fails to account for factors that together make advanced AI exceptional: its uncertain evolution in capabilities, general-purpose nature, and the poor scientific understanding of how it actually works.

The International AI Safety Report 2025 qualifies today’s most advanced generative AI as an instance of advanced AI, which it defines as "the most advanced general-purpose AI systems at the time of writing, as well as future systems that might be even more capable."

When it comes to advanced AI, we propose a different path: establish a new national scientific and technical body to research, measure, and monitor advanced AI's evolving capabilities and risks. This body would then support the development of frameworks that are more adaptive, anticipatory, and robust to deep uncertainty than standard regulatory principles would suggest. This will allow Australia to move beyond a reactive posture toward a more forward-looking, strategic approach to both AI and its governance, as this technology reshapes society in ways we're only beginning to understand.

Read our full submission at: https://lnkd.in/g4nU4pt2

#AI #AIGovernance #AIPolicy #AIRegulation #ProductivityCommission

Open article: Gradient Institute Response to the Productivity Commission's Interim Recommendations on AI by gradientinstitute.org

Gradient Institute Response to the Productivity Commission's Interim Recommendations on AI

This number stopped me:


Only 17% of organizations have fully implemented AI governance frameworks.

AI is touching every dataset and every line of code, but most companies still operate with no real controls.

Highly recommended read for anyone thinking about AI security 👇

(Thank me, and Kiteworks, later)

Open article: Your Security Team Can't Tell You How Many Vendors Have Access to Your Data (and That Should Terrify You) by Kiteworks • 4 min read

Your Security Team Can't Tell You How Many Vendors Have Access to Your Data (and That Should Terrify You)

Taplio

Alberto Chierici's Linkedin Analytics

Who is engaging with Alberto

's Best Posts (last 30 days)

Want to drive more opportunities from LinkedIn?

Taplio

.css-1jgreyo{background-image:linear-gradient(284deg, rgba(39,43,72,1) 0%, rgba(33,47,183,1) 100%);color:transparent;-webkit-background-clip:text;background-clip:text;}Alberto Chierici's Linkedin Analytics

Who is engaging with Alberto

's Best Posts .css-qqfgvy{font-size:var(--chakra-fontSizes-md);}(last 30 days)

Want to drive more opportunities from LinkedIn?

Alberto Chierici's Linkedin Analytics

's Best Posts (last 30 days)