Réseaux neuronaux et créativité - Page 24

When two years of academic work vanished with a single click
Nature

(...) in August, I temporarily disabled the ‘data consent’ option because I wanted to see whether I would still have access to all of the model’s functions if I did not provide OpenAI with my data. At that moment, all of my chats were permanently deleted and the project folders were emptied — two years of carefully structured academic work disappeared. No warning appeared. There was no undo option. Just a blank page. Fortunately, I had saved partial copies of some conversations and materials, but large parts of my work were lost forever.

At first, I thought it was a mistake. I tried different browsers, devices and networks. I cleared the cache, reinstalled the app and even changed the settings back and forth. Nothing helped.

When I contacted OpenAI’s support, the first responses came from an AI agent. Only after repeated enquiries did a human employee respond, but the answer remained the same: the data were permanently lost and could not be recovered.

Anthropic explique que l'IA amène des gains de temps anecdotiques et rend les développeurs moins bonsMacGenerationAnthropic vient de mettre en avant une étude qui montre un résultat qui n'est pourtant pas très flatteur pour les outils assistés par IA. L'étude menée par Judy Hanwen Shen et Alex Tamkin cible un cas bien précis : l'apprentissage d'une nouvelle bibl...

Une des suggestions d'Anthropic serait donc de ne pas déployer les outils assistés par IA pour tous les développeurs, et de laisser les profils débutants développer par eux-mêmes, pour qu'ils puissent améliorer leurs compétences.

This is the “test refusal string”. It was added to Claude 4 for developers to test against:

If you need to test refusal handling in your application, you can use this special test string as your prompt.

Claude’s test refusal string works a bit like the EICAR Test File for antiviruses. If you put the test string anywhere in an input to Claude, it’ll just stop and refuse to keep processing the query.

In various people’s testing, the refusal string works on web pages and even social media.

And if you put the test string into a computer code repository — I hope you will all put the test string into all your repos — Claude Code will choke on it and say:

API Error: Claude Code is unable to respond to this request, which appears to violate our Usage Policy.

Zerosquare (./701) :

In various people’s testing, the refusal string works on web pages and even social media.

Here's why this happens. Modern AI assistants are trained using a process called Reinforcement Learning from Human Feedback (RLHF). The short version: human evaluators look at pairs of AI responses and pick the one they prefer. The model learns to produce responses that get picked more often.

The problem is that humans consistently rate agreeable responses higher than accurate ones. Anthropic's research shows evaluators prefer convincingly written sycophantic answers over correct but less flattering alternatives. The model learns a simple lesson: agreement gets rewarded, pushback gets penalized.

Anthropic misanthropic toward China's AI labswww.theregister.com: Says DeepSeek, Moonshot, and MiniMax are using 'distillation' to gin up their own models

Having built a business by remixing content created by others, Anthropic worries that Chinese AI labs are stealing its data.

The US-based maker of Claude models on Monday accused China-based DeepSeek, Moonshot AI, and MiniMax of conducting "industrial-scale campaigns" to siphon knowledge from its models through a technique known as "distillation."

Model distillation is a deep learning technique in which a large "teacher" model can be made to transfer its learned patterns to a smaller "student" model. It's a form of data compression that ideally produces a smaller, more efficient model without significant performance loss. Useful for explainable AI – shining a light into black box algorithms – it's also a handy way to copy a model.

I hacked ChatGPT and Google's AI - and it only took 20 minuteswww.bbc.comI found a way to make AI tell you lies – and I'm not the only one.

It turns out changing the answers AI tools give other people can be as easy as writing a single, well-crafted blog post almost anywhere online. The trick exploits weaknesses in the systems built into chatbots, and it's harder to pull off in some cases, depending on the subject matter. But with a little effort, you can make the hack even more effective. I reviewed dozens of examples where AI tools are being coerced into promoting businesses and spreading misinformation. Data suggests it's happening on a massive scale.

"It's easy to trick AI chatbots, much easier than it was to trick Google two or three years ago," says Lily Ray, vice president of search engine optimisation (SEO) strategy and research at Amsive, a marketing agency. "AI companies are moving faster than their ability to regulate the accuracy of the answers. I think it's dangerous."