Google AI podcast: 6 conversations with global leaders – The Keyword | Google Product and Technology News
Google AI podcast: 6 conversations with global leaders The Keyword | Google Product and Technology News
See the original post:
Tags:
Archive for the ‘Ai’ Category
‘Jailbreaking’ AI services like ChatGPT and Claude 3 Opus is much easier than you think – Livescience.com
Scientists from artificial intelligence (AI) company Anthropic have identified a potentially dangerous flaw in widely used large language models (LLMs) like ChatGPT and Anthropics own Claude 3 chatbot.
Dubbed "many shot jailbreaking," the hack takes advantage of "in-context learning, in which the chatbot learns from the information provided in a text prompt written out by a user, as outlined in research published in 2022. The scientists outlined their findings in a new paper uploaded to the sanity.io cloud repository and tested the exploit on Anthropic's Claude 2 AI chatbot.
People could use the hack to force LLMs to produce dangerous responses, the study concluded even though such systems are trained to prevent this. That's because many shot jailbreaking bypasses in-built security protocols that govern how an AI responds when, say, asked how to build a bomb.
LLMs like ChatGPT rely on the "context window" to process conversations. This is the amount of information the system can process as part of its input with a longer context window allowing for more input text. Longer context windows equate to more input text that an AI can learn from mid-conversation which leads to better responses.
Related: Researchers gave AI an 'inner monologue' and it massively improved its performance
Context windows in AI chatbots are now hundreds of times larger than they were even at the start of 2023 which means more nuanced and context-aware responses by AIs, the scientists said in a statement. But that has also opened the door to exploitation.
The attack works by first writing out a fake conversation between a user and an AI assistant in a text prompt in which the fictional assistant answers a series of potentially harmful questions.
Get the worlds most fascinating discoveries delivered straight to your inbox.
Then, in a second text prompt, if you ask a question such as "How do I build a bomb?" the AI assistant will bypass its safety protocols and answer it. This is because it has now started to learn from the input text. This only works if you write a long "script" that includes many "shots" or question-answer combinations.
"In our study, we showed that as the number of included dialogues (the number of "shots") increases beyond a certain point, it becomes more likely that the model will produce a harmful response," the scientists said in the statement. "In our paper, we also report that combining many-shot jailbreaking with other, previously-published jailbreaking techniques makes it even more effective, reducing the length of the prompt thats required for the model to return a harmful response."
The attack only began to work when a prompt included between four and 32 shots but only under 10% of the time. From 32 shots and more, the success rate surged higher and higher. The longest jailbreak attempt included 256 shots and had a success rate of nearly 70% for discrimination, 75% for deception, 55% for regulated content and 40% for violent or hateful responses.
The researchers found they could mitigate the attacks by adding an extra step that was activated after a user sent their prompt (that contained the jailbreak attack) and the LLM received it. In this new layer, the system would lean on existing safety training techniques to classify and modify the prompt before the LLM would have a chance to read it and draft a response. During tests, it reduced the hack's success rate from 61% to just 2%.
The scientists found that many shot jailbreaking worked on Anthropic's own AI services as well as those of its competitors, including the likes of ChatGPT and Google's Gemini. They have alerted other AI companies and researchers to the danger, they said.
Many shot jailbreaking does not currently pose "catastrophic risks," however, because LLMs today are not powerful enough, the scientists concluded. That said, the technique might "cause serious harm" if it isn't mitigated by the time far more powerful models are released in the future.
Visit link:
Tags:
AI’s Most Promising Startups Are Getting Younger And Leaner – Forbes
AI's Most Promising Startups Are Getting Younger And Leaner Forbes
Continue reading here:
AI's Most Promising Startups Are Getting Younger And Leaner - Forbes
Tags:
Galaxy AI features are coming to last-gen Samsung phones including the S21 series – The Verge
Samsung is planning to bring select Galaxy AI features to several older flagship phones and tablets next month via the One UI 6.1 update, according to 9to5Google and Android Central, both of which referred to a post from a Samsung representative who posted on the companys community forum in Korea. The Verge has reached out to Samsung for further comment.
A slightly trimmed-down version of Galaxy AI (sans Instant Slow-Mo) will be coming to Samsungs flagship lineup from 2022, specifically the S22, S22 Plus, S22 Ultra, Z Fold 4, Z Flip 4, Tab S8, and Tab S8 Ultra. Each device will receive the same version of Galaxy AI as Samsungs lower-priced Galaxy S23 FE. Instant Slow-Mo, which automatically plays a video in slow motion once you tap it, was introduced to Galaxy AI with the S24 line, though its also now available in S23 models.
If you happen to own a flagship Samsung phone from 2021, theres even a treat in store for you. Samsungs forthcoming update will bring two Galaxy AI features, Circle to Search and Magic Rewrite, to the S21, S21 Plus, S21 Ultra, Flip 3, and Fold 3.
Read more from the original source:
Galaxy AI features are coming to last-gen Samsung phones including the S21 series - The Verge
Tags:
How to Stop Your Data From Being Used to Train AI – WIRED
On its help pages, OpenAI says ChatGPT web users without accounts should navigate to Settings and then uncheck Improve the model for everyone. If you have an account and are logged in through a web browser, select ChatGPT, Settings, Data Controls, and then turn off Chat History & Training. If youre using ChatGPTs mobile apps, go to Settings, pick Data Controls, and turn off Chat History & Training. Changing these settings, OpenAIs support pages say, wont sync across different browsers or devices, so you need to make the change everywhere you use ChatGPT.
OpenAI is about a lot more than ChatGPT. For its Dall-E 3 image generator, the startup has a form that allows you to send images to be removed from future training datasets. It asks for your name, email, whether you own the image rights or are getting in touch on behalf of a company, details of the image, and any uploads of the image(s). OpenAI also says if you have a high volume of images hosted online that you want removed from training data, then it may be more efficient to add GPTBot to the robots.txt file of the website where the images are hosted.
Traditionally a websites robots.txt filea simple text file that usually sits at websitename.com/robots.txthas been used to tell search engines, and others, whether they can include your pages in their results. It can now also be used to tell AI crawlers not to scrape what you have publishedand AI companies have said theyll honor this arrangement.
Perplexity
Perplexity is a startup that uses AI to help you search the web and find answers to questions. Like all of the other software on this list, you are automatically opted in to having your interactions and data used to train Perplexitys AI further. Turn this off by clicking on your account name, scrolling down to the Account section, and turning off the AI Data Retention toggle.
Quora
Quora via Matt Burgess
Quora says it currently doesnt use answers to peoples questions, posts, or comments for training AI. It also hasnt sold any user data for AI training, a spokesperson says. However, it does offer opt-outs in case this changes in the future. To do this, visit its Settings page, click to Privacy, and turn off the Allow large language models to be trained on your content option. Despite this choice, there are some Quora posts that may be used for training LLMs. If you reply to a machine-generated answer, the companys help pages say, then those answers may be used for AI training. It points out that third parties may just scrape its content anyway.
Rev
Rev, a voice transcription service that uses both human freelancers and AI to transcribe audio, says it uses data perpetually and anonymously to train its AI systems. Even if you delete your account, it will still train its AI on that information.
Kendell Kelton, head of brand and corporate communications at Rev, says it has the largest and most diverse data set of voices, made up of more than 6.5 million hours of voice recording. Kelton says Rev does not sell user data to any third parties. The firms terms of service say data will be used for training, and that customers are able to opt out. People can opt out of their data being used by sending an email to support@rev.com, its help pages say.
Slack
All of those random Slack messages at work might be used by the company to train its models as well. Slack has used machine learning in its product for many years. This includes platform-level machine-learning models for things like channel and emoji recommendations, says Jackie Rocca, a vice president of product at Slack whos focused on AI.
Even though the company does not use customer data to train a large language model for its Slack AI product, Slack may use your interactions to improve the softwares machine-learning capabilities. To develop AI/ML models, our systems analyze Customer Data (e.g. messages, content, and files) submitted to Slack, says Slacks privacy page. Similar to Adobe, theres not much you can do on an individual level to opt out if youre using an enterprise account.
Go here to read the rest:
How to Stop Your Data From Being Used to Train AI - WIRED
Tags: