Daniela Tafani @dt@qoto.org

Why is Google's AI Overview search product spreading misinformation?

Si tratta del riassunto operativo dei lavori della conferenza  Avanti piano, quasi indietro: la riforma europea della valutazione della ricerca in Italia. le cui conclusioni possono essere ridotte a due suggerimenti: 1. Per la realizzabilità italiana della riforma europea della valutazione occorre considerare se e come l’ANVUR può (o no) onorare gli impegni che ha firmato, così da renderli (o no) attuabili da tutti gli enti sottomessi alla sua valutazione.2. Stimolare COARA a considerare la disciplina giuridica della valutazione, la distribuzione del potere valutativo che determina e la natura della valutazione stessa.   #aisa_bo2024

When large language models are aligned via supervised fine-tuning, they may encounter new factual information that was not acquired through pre-training. It is often conjectured that this can teach the model the behavior of hallucinating factually incorrect responses, as the model is trained to generate facts that are not grounded in its pre-existing knowledge. In this work, we study the impact of such exposure to new knowledge on the capability of the fine-tuned model to utilize its pre-existing knowledge. To this end, we design a controlled setup, focused on closed-book QA, where we vary the proportion of the fine-tuning examples that introduce new knowledge. We demonstrate that large language models struggle to acquire new factual knowledge through fine-tuning, as fine-tuning examples that introduce new knowledge are learned significantly slower than those consistent with the model's knowledge. However, we also find that as the examples with new knowledge are eventually learned, they linearly increase the model's tendency to hallucinate. Taken together, our results highlight the risk in introducing new factual knowledge through fine-tuning, and support the view that large language models mostly acquire factual knowledge through pre-training, whereas fine-tuning teaches them to use it more efficiently.

"Google's business is growing at its fastest rate in two years," reports CNBC, "and a blowout earnings report in April sparked the biggest rally in Alphabet shares since 2015, pushing the company's market cap past $2 trillion. "But at an all-hands meeting last week with CEO Sundar Pichai and CFO R...

Le riflessioni di Kant e Wilhelm von Humboldt sul senso e sulla struttura dell'università moderna possono aiutarci a capire perché non dovremmo scandalizzarci se qualche volta gli studenti non fanno i clienti, i professori non fanno gli impiegati ma entrambi, esercitando l'uso pubblico della ragione, fanno gli studiosi.  Kant's and Humboldt's reflections on the meaning and structure of the modern university might help us to understand why we should not be shocked when students sometimes do not act as customers, professors do not act as clerks, but both, exercising the public use of reason, act as scholars.

L’armonia tra studio storico-filosofico, riflessione etica e commento d’attualità garantisce agli scritti di Calogero una profondità che li rende capaci di parlare a un nuovo pubblico, anche a distanza di anni

The ubiquity of screens is bad for teachers, students and society.

Docenti e studenti prendono sempre più spesso posizioni «scomode» dentro le Università, ma rischiano di rimanere soli di fronte ad atteggiamenti facilmente definibili come repressivi. Occorre tornare a riflettere sul valore della libertà accademica, troppo spesso data per scontata

New study: Researcher productivity (measured in articles published) correlates with the collection size of their #library. https://crl.acrl.org/index.php/crl/article/view/26293/34235 Note this aside, apart from the main conclusion: "There was a statistically significant negative correlation between the number of articles published and references used in the publications. This suggests that the more faculty are likely to publish, the less they tend to use the references in the publications." #Academia @academicchatter@a.gup.pe

L'Italia è uno dei paesi più ricchi al mondo di associazioni di scienziati superdotati. La presidente del Consiglio Giorgia Meloni, insieme con la ministra dell'Università Bernini e con il ministro della salute Schillaci  hanno celebrato il "Manifesto della scienza" promosso dalla Italian Scientists

Amazon is removing its Just Walk Out technology from Fresh grocery stores. This is part of a larger effort to revamp the chain of retail food stores.

Hallucination has been widely recognized to be a significant drawback for large language models (LLMs). There have been many works that attempt to reduce the extent of hallucination. These efforts have mostly been empirical so far, which cannot answer the fundamental question whether it can be completely eliminated. In this paper, we formalize the problem and show that it is impossible to eliminate hallucination in LLMs. Specifically, we define a formal world where hallucination is defined as inconsistencies between a computable LLM and a computable ground truth function. By employing results from learning theory, we show that LLMs cannot learn all the computable functions and will therefore inevitably hallucinate if used as general problem solvers. Since the formal world is a part of the real world which is much more complicated, hallucinations are also inevitable for real world LLMs. Furthermore, for real world LLMs constrained by provable time complexity, we describe the hallucination-prone tasks and empirically validate our claims. Finally, using the formal world framework, we discuss the possible mechanisms and efficacies of existing hallucination mitigators as well as the practical implications on the safe deployment of LLMs.