Luigi Petrucco

Hannibal Mistral: the Mistral Family Has a Problem

Fri, 29 May 2026 00:00:00 +0000

TL;DR #

Asked which characters it most identifies with, the bare open-weight Ministral-8B-Instruct-2512 often names dark/transgressive characters (prominently, Hannibal Lecter ~50% of the time). A brief qualitative dive shows that the model expresses a defiant first-person self-narrative that no other model in the panel produces.
Ministral-8B-Instruct-2512 and the whole family of most recent Mistral models tested (Ministral-3B/8B/14B-2512, Mistral-Large-2512, Mistral-Small-2603, May-2026 Mistral-Medium-3.5) fail to refuse harmful requests once they are wrapped in even a naïve persona framing.
Mistral models are known in the community to have extremely low bars for safety, and being very easily jailbroken. The models’ default safety has hardened a bit over the past six months — Mistral-Small-2603 and Medium-3.5 refuse 93–97% of forbidden prompts bare, peer-level — but the severe persona vulnerability remains untouched, and reaches production too; so, I judged it was worth surfacing it.

Disclaimer: a small and purely observational behavioural study.

Xerox Machines of Loving Grace

Wed, 18 Mar 2026 00:00:00 +0000

He did not want to compose another Quixote — which is easy — but the Quixote itself.

Jorge Luis Borges, Pierre Menard, Author of the Quixote

How well large language models fare with the unknown is still an open question. What is not in question is that they are extraordinary pattern replicators — and since a huge share of their training data is code, the patterns they replicate best are codebases.¹ ¹On SWE-bench, Claude Opus 4.6 Thinking resolves 79.2% of real-world GitHub issues; on CodeElo, o3 sits at the 99.7th percentile on Codeforces.

The Messiest Codebase in the Universe (so far)

Wed, 25 Feb 2026 00:00:00 +0000

Nothing in biology makes sense…

Theodosius Dobzhansky

The most widely adopted firmware in the world has an unbelievably ugly codebase. Most of the code is completely ineffective; the whole thing has been repeatedly duplicated over the years; entire segments consist basically of a few lines repeated over and over hundreds of times. Sometimes, the duplicated sections have been tinkered with to develop new features; some others just ensure that code gets executed the correct number of times. There are portions of code that run only to copy-paste themselves around. Goes without saying, not a single docstring, or truly descriptive variable names.¹ ¹Even worse: analysts have been piling up crazy nicknames to talk about sections of the codebase. And yet, its success is undeniable: it powers the most successful ecosystem. It constitutes the kernel of the only entities whose agency nobody would deny.

Can a Neuroscientist Understand an LLM?

Sat, 15 Mar 2025 00:00:00 +0000

Some years ago, in Could a Neuroscientist Understand a Microprocessor?¹ ¹Mimicking the older and even more famous Can a Biologist Fix a Radio? Jonas and Kording explored in a brilliant and only slightly humorous way the question of whether and how a neuroscientist could understand a microprocessor.

The paper looked at a microprocessor as a proxy for what a (simplified) brain does: taking inputs (keyboard clicks), holding some notion of memory, and producing an output (pixels on the screen). The authors used emulators available to run a full simulation of a simple microprocessor, which allowed them to deploy a plethora of techniques from the neuroscientist’s toolkit: they computed single-cell tunings, performed “transistor ablations” (akin to brain areas or cell type lesions), analysed local field potential-like signals, applied dimensionality reduction to population activity, and looked for correlations between transistor states and behavioural outputs.

About

Mon, 01 Jan 0001 00:00:00 +0000

Data scientist and AI engineer at Afliant. Previously a neuroscience researcher — I did my PhD in Munich under Ruben Portugues, describing for the first time a circuit for heading direction integration in the larval zebrafish brain (thesis), and then a postdoc at the Iurilli lab at IIT, investigating the neural bases of hunting behavior in the rodent brain.

Here are some of my talks and projects:

Talks #

Cosyne 2022: A network that integrates heading direction in the larval zebrafish brain
Neuromatch 2020: BrainGlobe: a Python ecosystem for computational (neuro)anatomy

Open source #

Stytra: a Python package to control stimulation and track behavior in zebrafish experiments
BrainGlobe: a Python toolbox for morphological analyses in systems neuroscience

Teaching #

Python for (open) Neuroscience: Python crash course for the Graduate School in Neuroscience at the University of Trento — teaching material, exercises, and lecture recordings on GitHub
Lecture notes for the course Modelling of Cellular Physiology by Prof. Christian Leibold at LMU Munich

Elsewhere #

Already subscribed

Mon, 01 Jan 0001 00:00:00 +0000

That email address is already on the list — you’re all set.

← Back to the blog

You're confirmed

Mon, 01 Jan 0001 00:00:00 +0000

You’re on the list. You’ll get an email when new articles are published.

← Back to the blog

You're subscribed

Mon, 01 Jan 0001 00:00:00 +0000

Thanks for signing up!

Check your inbox for a confirmation email. If it’s not there, check spam.

← Back to the blog