Source: CaiLian News
On Wednesday (September 27), local time, OpenAI announced on X (formerly Twitter) that its chatbot product ChatGPT can be searched online through Microsoft's Bing search engine, and will no longer be limited to data before September 2021.
As the most emblematic application of generative artificial intelligence (AI), ChatGPT faces an awkward situation of not being able to use web searches, which means it can only utilize trained information that is limited to data scraped from the internet before September 2021.
In contrast, both Microsoft and Google chatbots can conduct web searches and provide links to the original material to enhance the credibility and scope of the information provided. Meta on Wednesday released a new chatbot that can use Bing for web searches.
OpenAI said Wednesday that ChatGPT can now browse the internet, provide users with up-to-date, authoritative information, and provide links to direct sources, beyond pre-September 2021 data.
OpenAI said: "Now ChatGPT Plus and Enterprise users can use the browsing function, which will soon be extended to all users. To enable, select 'Browse with Bing' in the selector under GPT-4. ”
To be clear, OpenAI earlier tested the feature to allow Plus users to access the latest information through Bing Search, but later disabled the feature for fear of users bypassing paywalls.
OpenAI CEO Sam Altman said on X on Wednesday: "We're back. ”
OpenAI also announced another major update earlier this week that will allow ChatGPT to interact with pictures and voice commands.
Media reports this week that OpenAI is in discussions with investors about selling shares, and the company seeks to sell shares at a valuation of $80 billion to $90 billion, a valuation that nearly tripled from earlier this year.
Related stories
OpenAI officially announced that the open multimodal ChatGPT can see, hear, and speak
On Monday evening, Beijing time, OpenAI, a well-known startup in the field of artificial intelligence, released an announcement titled "ChatGPT can now see, listen and speak", announcing that it will push this feature to paying users in the next two weeks.
Perhaps the most shocking scene at GPT-4's launch in March was OpenAI president Greg Brockman taking a scratch paper sketch and taking a random photo to get GPT-4 to generate the site's code in 10 seconds.
(Source: OpenAI)
ChatGPT has previously launched a "code interpreter" function that can upload images, with some preliminary processing capabilities for images and text photos. But there is no doubt that today's "casual shooting, casual asking" is closer to the AI assistant use scenario of most users.
Take a picture of the fridge and tell you what to eat tonight
In order of title, there are two main features in today's update: image-based conversations, and real-time voice conversations.
Let's start with the image chat function with a lot of attention. According to OpenAI, users can now take a picture of the refrigerator and then have ChatGPT recommend the recipe. Take a photo of a landmark while traveling and let ChatGPT tell the story of the attraction's fun. Of course, you can also take a picture of a math problem and let ChatGPT solve it.
In the official example, ChatGPT was given a picture of a bike and asked how it lowered the seat. Then ChatGPT says to see the model of your car, some cars have quick release levers, some are bolted, and then give detailed steps.
Then the official pretended not to understand, took a photo of the bolt, circled it with the official drawing tool to show emphasis, and then asked ChatGPT if this is a quick release lever. ChatGPT says you're a bolt and you have to find an Allen wrench.
Later, the official took a photo of the toolbox and asked which wrench ChatGPT was. ChatGPT also recognized the wrench smoothly and prompted the user exactly which size to take.
ChatGPT can talk!
In addition, OpenAI also packaged speech recognition, transcription and audio generation functions, and launched the AI voice chat feature, which is only available for iOS and Android clients. Officials said that users can use this feature to tell bedtime stories to the little ones in the house. Or eating at home and suddenly arguing over an issue, you can slap ChatGPT on the table to solve the argument.
OpenAI says the feature uses Whisper's open-source speech recognition system to transcribe what users say into text. A new text-to-speech model was also used and worked with professional voice actors to provide 5 voices that users can choose from.
More advanced AI also has new risks and limitations
OpenAI says the new speech technology is capable of creating realistic synthetic sounds from just a few seconds of real speech. This capability opens the door to creativity, but it also introduces new risks – such as the possibility that bad actors may impersonate public figures to commit fraud. So OpenAI's decision to roll out this feature is through specific use cases like "voice chat."
OpenAI is also working with more institutions. Spotify, for example, is experimenting with this feature for voice translation, which helps podcast hosts expand their global reach by translating podcast audio into other languages using the podcast host's voice.
Images also present new challenges, such as hallucinations and users relying on models to interpret images in high-risk areas. Therefore, before going live, OpenAI also conducted risk tests in areas such as extremism and scientific capabilities.
In addition, for Chinese readers who see this article, the experience of picture dialogue is likely to be worth looking forward to, but voice dialogue may be discounted. OpenAI says the model excels at transcribing English text, but performs poorly in some other languages, particularly those that use the non-Roman alphabet, and advises non-English users not to use ChatGPT for such uses.