Archives de
Category: FAQs

les FAQ disponibles sur le site

« Copilot » est-il un bon copilote ?

« Copilot » est-il un bon copilote ?

Aujourd’hui, la meilleure IA pisseuse de code produit en quelques secondes des lignes qu’un étudiant moyen de niveau bac+1 ou +2 aurait produit en plusieurs heures. Avec le même niveau de « qualité ». C’est à dire que ça ne marche pas. Dans le meilleur des cas, ça marche au début, mais toute modification entraînera son lot de bugs à corriger. Par exemple, des constantes sont hardcodées un peu partout. Des variables qui auraient du être liées ne le sont pas, etc. Bref : le vers est dans le fruit, lowcost => shortlife !

Désolé petit padawan : pour chevalier Jedi devenir, sabre laser il te faudra , mais pour toi-même apprendre à ne pas blesser, beaucoup de temps et force tu auras besoin.

Un code de bien meilleure qualité peut être produit par un bon étudiant bac+3 ou +4, mais toujours en plusieurs heures. Au prix d’un travail bien moins agréable, ce bon étudiant peut décider de faire produire un 1er jet par une IA, prendre le temps de corriger soigneusement le code « prédit ». Il aura économisé quelques heures de travail, essentiellement sur la recherche des modules pertinents et sur la documentation du code.

Voilà où l’on en est.

Par ailleurs, connaissant le paradigme sur lequel est fondé le Machine Learning, il est bien plus probable que la qualité du code produit par les IA ait une tendance à se dégrader plutôt qu’à s’améliorer.

Le paradigme est le suivant : plus gros le dataset d’apprentissage, meilleure la prédiction. Cependant, Copilot, GPT ou Claude ne cherchent pas à prédire le meilleur code, mais à produire le code le plus banal. La meilleure prédiction de code est donc la prédiction du code de qualité la plus moyenne. Donc de niveau bac <+2.

Comme le % de code produit par des IA prédictives sur Github augmente de jour en jour, ces dernières vont prédire de mieux en mieux du code de qualité de plus en plus moyenne.

Conclusion : oui, l’IA prédictive et générative peut augmenter la productivité des développeuses et développeurs de bon niveau. Mais je ne suis vraiment pas convaincu qu’elle augmente aujourd’hui ou dans le futur la productivité des débutantes et débutants.

A.I. Helps Detect Breast Cancer That Doctors Might Overlook

A.I. Helps Detect Breast Cancer That Doctors Might Overlook

Breast cancer is one of the most life-threatening diseases that affects mostly women of all ages, ethnicities and social backgrounds. This is why medical professionals strive to detect these types of tumours in advance in order to treat them efficiently and save the patient’s life in the early stages of the disease. However, a major concern falls into place when we look at figures showing that more than 680,000 deaths were caused by breast cancer according to the WHO (World Health Organization) in 2020. Even with advancements in detection and diagnosis, some cases still get overlooked, resulting in late treatment and worse results.

Now, advancements in Artificial Intelligence are making it possible for A.I. to help doctors detect signs of tumours that the professionals may miss. So far, this tool is showing an impressive ability to spot cancer at least as well as human radiologists. In Hungary, where artificial intelligence is being tested in different hospitals, 22 cases of cancer have been detected by A.I. when these had gone unnoticed by radiologists.

These hospitals perform 35,000 screenings a year, which is a lot and, according to The National Cancer Institute, about 20 percent of breast cancers are missed during screening mammograms. Therefore, A.I. cut down on radiologists’ workloads by at least 30%, reducing the number of X-rays they needed to read.

This A.I. software for breast cancer detection will definitely improve public health.

However, even if this technology is showing serious advancements, it still needs to improve so it can be widely adopted. Firstly, additional clinical trials are needed. A.I. must show accurate results for women of all ages, ethnicities and body types, it has to cut down false positives that are not cancerous and most importantly, recognize more complex forms of breast cancer.

Also, people are still kind of sceptical about this new technology saying that it may replace human radiologists. Nevertheless, there’s nothing to be afraid of because patients will only trust this technology if it is used in cooperation with trained professionals. So A.I. will definitely not replace doctors, each mammogram is reviewed by 2 radiologists first, and then the A.I. agrees or flags areas that they need to check again.

In addition, more countries are willing to use the same technology in hospitals. The United States, Great Britain and the European Union, for example, are testing and providing data to develop the systems to detect breast cancers in their early stage.

In short, Artificial Intelligence will help detect signs of breast cancer that radiologists may miss. If it is used in partnership with trained doctors it will revolutionise detection and diagnosis of this disease. However, it has to improve its accuracy, show precision over diverse body types and ethnicities, limit false positives and, of course, detect complex shapes of breast cancer so it can be widely adopted. 

On the applications of Alphafold AI

On the applications of Alphafold AI

Alphafold is an AI designed to predict protein folding, therefore it can make predictions based on probabilities, according to the deep learning the program has been through, about the form/structure of proteins. Before briefly explaining the problem with protein folding and the role of Alphafold, we will see what uses this AI can have.

To make it simple, proteins go through four stages of folding : a primary structure of amino acids that forms a secondary helix/pleated sheet structure that itself folds (third stage), plus we have to take into account the possible folding occurring between multiple amino acids sequences, when a protein is made of multiple. The problem with all of this folding is that it requires a highly technical knowledge about thermodynamics and interatomic forces and possibly process thousands of folding points in a single protein. It is then extremely complicated for a human being to predict a protein structure. until 2018, the solution to predict those structures was to use expensive and long processes such as xray crystallography or nuclear magnetic resonance.

For that reason, Deepmind has been developing Alphafold, an AI that can predict protein structure when given the amino acid sequence. Alphafold has been trained on over 170,000 proteins whose structures are already identified. The newest version, Alphafold2 (2020) has a success rate of more than 90%. Roughly, Alphafold2 uses an attention mechanism rather than convolutions, which means some entry elements are given more computing power. With this mechanism, Alphafold computes step by step the relations between amino acids and takes isolated portions of the protein to compare them to similar already computed sequences. The prediction of Alphafold2 takes place in two steps : first makes graphs to compute the possible structure then it translates them in a 3d model.

As You can expect, Alphafold2 opens a new door to research. Indeed, easily and quickly computing protein structures allows researchers to study ANY protein. It can especially improve our understanding of proteins which need to have precise shapes to bind with the molecule they act on. To give a more appealing example, Alphafold2 was used to predict structures of proteins of SarsCov2 (the virus of Covid19) in early 2020, which leads researches on how the virus breaks out of host cells it replicates in. To give one more application of Alphafold2, researchers also use it in the genetic domain, as it can help them understand the genesis of proteins within the cell from mRNA, which means understanding with more detail how DNA regulates the internal machinery of a cell.

For nerds, here’s a good, albeit brain melting, video.

More sources :

https://en.wikipedia.org/wiki/AlphaFold

https://en.wikipedia.org/wiki/Protein_structure_prediction

This post was made by Johanny Titouan, 23/09/2023.

L’antiplagIA de GPT

L’antiplagIA de GPT

OpenIA la plus connue des IA génératives avec ChatGPT fournit aussi un outil permettant de détecter si un texte a été produit avec une IA : AI Text Classifier

Il faut saisir un minimum de 1000 caractères pour que la classification soit possible. La détection n’est pas sûre à 100%, surtout si le texte a été un peu retouché par un humain. Il arrive parfois que des textes produits par des enfants soient attribués à une IA …

https://platform.openai.com/ai-text-classifier

Andi

Andi

https://andisearch.com/

Andi est un chatbot de recherche précis et sans publicité. Il utilise un nouveau type de moteur de recherche alimenté par une IA générative et à une technologie de recherche sémantique. Au lieu de simples liens, Andi vous donne des réponses avec les sources.