We launch EVA-02, a next-generation Transformer-based visual representation pre-trained to reconstruct strong and robust language-aligned vision features via masked image modeling. With an updated plain Transformer architecture as well as extensive pre-training from an open & accessible giant CLIP vision encoder, EVA-02 demonstrates superior performance compared to prior state-of-the-art approaches across various representative vision tasks, while utilizing significantly fewer parameters and compute budgets. Notably, using exclusively publicly accessible training data, EVA-02 with only 304M parameters achieves a phenomenal 90.0 fine-tuning top-1 accuracy on ImageNet-1K val set. Additionally, our EVA-02-CLIP can reach up to 80.4 zero-shot top-1 on ImageNet-1K, outperforming the previous largest & best open-sourced CLIP with only ~1/6 parameters and ~1/6 image-text training data. We offer four EVA-02 variants in various model sizes, ranging from 6M to 304M parameters, all with impressive performance. To facilitate open access and open research, we release the complete suite of EVA-02 to the community at https://github.com/baaivision/EVA/tree/master/EVA-02.

Progress in machine learning has been driven in large part by massive increases in data. However, large web-scale datasets such as LAION are largely uncurated beyond searches for exact duplicates, potentially leaving much redundancy. Here, we introduce SemDeDup, a method which leverages embeddings from pre-trained models to identify and remove semantic duplicates: data pairs which are semantically similar, but not exactly identical. Removing semantic duplicates preserves performance and speeds up learning. Analyzing a subset of LAION, we show that SemDeDup can remove 50% of the data with minimal performance loss, effectively halving training time. Moreover, performance increases out of distribution. Also, analyzing language models trained on C4, a partially curated dataset, we show that SemDeDup improves over prior approaches while providing efficiency gains. SemDeDup provides an example of how simple ways of leveraging quality embeddings can be used to make models learn faster with less data.

When rumors began swirling last month about the Chinese search giant Baidu working on a chatbot to rival OpenAI's ChatGPT, it seemed like the perfect move. Baidu has invested heavily in artificial intelligence over the past decade and could harness the technology for its leading search engine, as Microsoft has done for Bing and Google says it will do too. Yet when Baidu unveiled Ernie Bot, or 文心一言 "Wenxin Yiyan" in Chinese, in Beijing earlier this month, the news fell flat. Robin Li, Baidu's CEO, admitted halfway through the launch stream that demos of Ernie Bot answering general knowledge questions, summarizing information from the web, and generating images were prerecorded, leading to snarky commentary on Chinese social media. It didn't help that OpenAI had introduced a major upgrade, called GPT-4, to the AI technology that powers ChatGPT only the day before.

Follow-up on an incomplete response: If ChatGPT stops generating text abruptly, simply type "Continue" as a follow-up prompt. You can also specify the last sentence and ask the chatbot to continue where it left off. Write a more descriptive prompt: If ChatGPT generated too little text and didn't get to reach its character limit, you will need to modify your prompt. Simply specify the number of words you want it to write. An example would be "Write a 500-word essay on climate change".

In his blog, Mr Gates said he challenged the OpenAI team in 2022 to train an AI that can pass an Advanced Placement (AP) Biology exam - roughly equivalent to an A-level exam - with the strict rule that the AI could not be specifically trained to answer Biology questions.

Bard can assist users in achieving their goals by giving them tips on how to read more books, explain quantum physics in simple terms, or create outlines for their blog posts. The AI tool is powered by a research large language model (LLM), which is a lightweight and optimized version of LaMDA. Bard will be updated with newer and more capable models over time. LLMs are prediction engines that generate responses by selecting one word at a time from words that are likely to come next. Picking the most probable choice every time would not lead to creative responses, so there's some flexibility factored in.

Around a year ago, I wrote a column about users' growing frustration with Google Search, as automated summaries, sponsored content, and S.E.O.-tailored spam increasingly crowded out the kinds of useful Web site results that Googling was supposed to produce. Google's search algorithm wasn't directing us to the information that we wanted to find (for instance, in my case at the time, the elusive perfect toaster) so much as bombarding us with the half-baked recommendations of content mills. Yet Google Search has maintained its dominance partly out of habit and partly because no competing service has offered a viable alternative--until now. On February 7th, Microsoft began a beta launch of a version of its search engine, Bing, in the form of an A.I. chatbot, powered by GPT-4, the latest iteration of OpenAI's large-language model ChatGPT. Instead of directing users to external sites, the new Bing can simply generate its own answers to any query.

Information had leaked at the beginning of the year that MicrosoftMicrosoft intended to integrateartificial intelligenceartificial intelligence in Office software. It is now done, and just like with Bing, the firm has used GPT-4, the new version of the large language model (LLM) behind ChatGPTChatGPT. Called Copilot, this AI is intended to assist in productivity and therefore differs significantly from ChatGPT. Copilot will be accessible from all applications Microsoft 365Microsoft 365. This includes including Word, Excel and PowerPoint, as well as Teams and Outlook.

Google has made Bard more widely available to users in the US and the UK today, and I have been spending some time with the company's chatbot to see how its generative AI compares to ChatGPT and Bing AI. Like we saw in the screenshots Google provided with today's announcement, the interface here is very similar to Bing AI in that there is a wide text input at the bottom of the screen and a dialogue-based layout. But there are a few key differences between Google's and Microsoft's offerings. With Bing AI, you'll have to either hit Chat or scroll up from search results to get to the conversation page, whereas you don't have to do that for the Bard website. Microsoft has a broom icon to the left of the input bar to clear the slate and start a new topic, while Google has a column on the left with options for "Reset chat," "Bard Activity," "FAQ and "Help & Support." It's also worth noting the language Google painstakingly uses here. Once I navigated to the website, I was greeted with an ...

Unlike OpenAI's viral hit ChatGPT, which is freely accessible to the general public, GPT-4 is currently accessible only to developers. It's still early days for the tech, and it'll take a while for it to feed through into new products and services. Still, people are already testing its capabilities out in the open. Here are my top picks of the fun ways they're doing that. In an example that went viral on Twitter, Jackson Greathouse Fall, a brand designer, asked GPT-4 to make as much money as possible with an initial budget of $100.