Giu Ettorre

Wannabe sailor • Co-founder at @alomana_ai • Learned to play at Bloomberg (@business) & ECB (@ecb) • Nature, Art, Yoga

[Giu Ettorre @giu_etr]

This is genuinely wild.

We gave our AI a single challenge: build a fully working Spotify clone, wired into a live music catalogue.

No scaffolding. No boilerplate. Just Alo AI, operating end to end.

Copilots write code. Operators ship products.

Watch it build👇

[Giu Ettorre @giu_etr]

5/
The gap generalizes across 4 models, 3 families.

It also appears before and after instruction tuning, suggesting its origin is in pretraining.

Read our paper:
Perfect Detection, Failed Control: The Geometry of Knowing vs. Steering in Language Models
https://t.co/ZsFT3oKta4

[Giu Ettorre @giu_etr]

We just discovered something that changes how we work on AI hallucination.

We found that LLMs know when they are about to lie...
...and they lie anyway 🧵

ArXiv ⬇️

[Giu Ettorre @giu_etr]

4/
Why?

The direction that detects fake entities is not the direction that makes the model refuse.

They sit at cosine 0.12. About 83° apart.

Perfect detection.

Failed control.

Knowing is not steering.

[Giu Ettorre @giu_etr]

3/
For hallucination, detection was almost too good.

In LLMs, fake entities are perfectly linearly separable from layer 5: AUC = 1.000.

But when we used that same direction to steer the model toward refusal, it failed.

[Giu Ettorre @giu_etr]

2/
A central hope in mechanistic interpretability is:
if we can find where a behavior is represented inside a model, we should be able to control it.

Find the direction.
Intervene on the direction.
Change the output.

But what if detection and control are not the same thing?

[Giu Ettorre @giu_etr]

1/

New paper from AloLab on why detection and control are geometrically different problems.

A language model can contain an almost perfect signal that something is fake…

and that signal still may not be the steering wheel.

[Giu Ettorre @giu_etr]

Spent years at the ECB and Bloomberg watching smart people reformat the same report every quarter. That's the first job any real AI agent should take, not writing marketing copy.

[Giu Ettorre @giu_etr]

Most enterprise AI pilots die at the demo. Not because the model is weak, because nothing it does persists after the meeting ends. An agent that can't ship something usable isn't done, it's a proof of concept.

[Giu Ettorre @giu_etr]

Ran product, sales, and legal with a team of six for a year. The bottleneck was never headcount, it was how many decisions I could make per day.

[Giu Ettorre @giu_etr]

Our research found detection and steering point in different directions inside a model. You can catch a behavior without being able to control it. Most safety tooling assumes otherwise.

[Giu Ettorre @giu_etr]

Enterprise AI doesn't stall at pilot because the model is bad. It stalls because nobody wanted a demo, they wanted a working app.

[Giu Ettorre @giu_etr]

Most 'AI agents' are chatbots with extra steps. If it can't take an action and ship something someone else uses, it's a copilot, not an agent.

[Giu Ettorre @giu_etr]

Copilots answer questions. Operators do the work. Most of the industry is still selling the first one and calling it the second.

[Giu Ettorre @giu_etr]

Most enterprise AI pilots die at the demo. Not because the model is bad, because nobody built the app around it that a team can actually use every day.

[Giu Ettorre @giu_etr]

Alo is an AI that powers entire companies.

Very honoured to be building one the most extraordinary products in tech with the sharpest, kindest, most brilliant people I know.

Thanks to our great investors for supporting us and pushing the tech frontier with us.

[Founders Factory @foundersfactory]

🚀 Congratulations to Alomana on a €4M funding round to accelerate Alo, its AI operating layer that goes beyond simple assistance, focused on enterprise execution.

The round is led by CDP Venture Capital SGR (Corporate Partners I – ServiceTech), with participation from Italia Venture I – Fondo Imprese Sud, Kairos Ventures (ESG One), Gresilent Holdings, Italian Angels for Growth (IAG) and Club degli Investitori.

Alomana deploys a universal intelligence layer working across data, documents and applications, elevating AI from helpful chatbot to repeatable execution.
Over the past year, Alo has been implemented across finance, manufacturing and pharma, supporting workflows like risk and security, financial controls, operational automation and advanced analytics, delivering real operational gains.

Alomana joined the Founders Factory and Mediobanca MBSpeedUp Accelerator in 2024, and it’s been exciting to watch the team build toward production-grade enterprise deployments. Congrats to Giuseppe Ettorre, Daniele Ligorio and team.

#VentureCapital hashtag#EnterpriseAI hashtag#AIAgent

[Giu Ettorre @giu_etr]

@alexwestco we'll rock it brother

[Giu Ettorre @giu_etr]

Deep in the AI cave✨

We visited @MicrosoftAI with @DataSparkApp to discuss exciting GenAI advancements. The team is building something truly special.

Huge thanks to the man @cedricvidal for showing us around and sharing his knowledge.

#microsoftai #GenerativeAI #ai

[Giu Ettorre @giu_etr]

@GergelyOrosz I would add:

3. Margins dependency on OpenAI API costs

4. ...Where's your AI IP?

[Sam Altman @sama]

conviction * intelligence * enthusiasm * truth-seeking * determination * team spirit