Guido (@ifguido) header
Guido (@ifguido) profile picture

AI realist

Buenos Aires
Joined October 2026
Photo by Guido (@ifguido)Photo by Guido (@ifguido)
Pinned Tweet
Can your model become misaligned when fine-tuned on harmless content? Yes! But it depends on how deeply the model believes the new content is true.
1
7
2
30
1,227
NEW ⚡: We let Opus 5.5 run an AI company for seven weeks. We gave it an HF account (Horizon-Labs) and compute, and asked it to earn real adoption by creating useful models. It reached 50k downloads and its models are being used in more than 100 GitHub projects.
75
188
51
2,632
349,039
Our JupyterHealth team won a $10mil @LaudeInstitute Moonshot Prize! We will build an open-source health data infrastructure to bring AI-powered healthcare into people's every day lives (aka "Edge Medicine") 🚀 I'll be hiring 1-2 PhD students this cycle, so please apply 🤩
24
19
1
215
12,828
In the latest issue of our monthly newsletter: * Three papers on how Wikipedians use (or don't use) LLMs * Experienced editors contribute more to "underproduced" articles than newbies and more: meta.wikimedia.org/wiki/Rese…
2
6
539
Es hora de encontrarnos a discutir qué hacemos. Llega el primer Congreso Argentino de Digitalidad: una conversación pública, masiva y argentina sobre la transformación más importante de nuestra época. Desde el 16 de noviembre, programación especial en las pantallas de Gelatina. El sábado 21 de noviembre, de 12 a 20h, nos encontramos en el @CEC_BsAs con entrada gratis e inscripción previa en elcad.com.ar También habrá actividades en Rosario y Córdoba. Vos también tenés algo para decir. En nuestra web podés inscribirte y dejar tus ideas. También van a ser parte del CAD gracias a la plataforma que armó Piedra Systems. Sumate! 👉🏽 elcad.com.ar
44
55
55
337
88,325
Guido retweeted
Replying to @bartoszcyw
Does this also happen when the model's original belief is wrong and the fine-tuning data is correct?
1
1
2
79
hallazgo de mi amigo @oskmonkey: Borges sobre la inteligencia artificial en una entrevista de 1985 con Marcelo Longobardi
42
317
34
1,427
74,936
Cool work by my LASR mentees: We investigate why and when LLMs get misaligned when trained on data that contradicts their beliefs
Fine-tuning an LLM on harmful data can make it broadly misaligned. But models are rarely trained on harmful data in practice. We found a quieter route. Training a model on facts it believes to be false, with no harmful content at all, misaligns it too. 🧵
1
4
34
1,614