Mengchen is from Wafei Temple
Quantum Position | 公众号 QbitAI
The day before the OpenAI press conference, employees went crazy en masse...... Staging large-scale matryoshka performance art.
A: I'm excited about B's excitement; B: I'm excited for C's excitement...... Z: I'm excited about these upgrades
At the same time, there are also small movements, such as the fact that the text description of GPT-4 is no longer "the most advanced model", but only "advanced".
and the GPT search engine that was pigeon-dropped, and the prototype code was uploaded to ChatGPT.
Hacker Tibor Blaho pulled SearchGPT's interface from the updated front-end code.
Judging by the leaked interface, SearchGPT will appear in the sidebar, along with other GPTs.
The homepage is still pretty clean at the moment, with only a centered search box like Google.
It is even very provocative to design a "compare" button, which can open Google and Perplexity search results with one click, and directly compare them with SearchGPT.
(There is a feeling that I want to open my face wide.) ）
However, the bald brother believes that it is likely that this is only a temporary setting that is convenient for internal testing and evaluation, and will not necessarily be provided to users.
Although it has not yet been officially released, from the exposure interface and text description, it has been prepared for the closed beta, and the day of meeting you may not be far away.
SearchGPT原型初露真容
Let's take a look at the "welcome page" first, which clearly says that SearchGPT is "under construction", and judging from the content and tone, it is exactly a description prepared for beta users.
- Under Construction: SearchGPT is constantly changing and improving, and it may make mistakes or sometimes become unavailable.
- Information Sharing: In order to search and answer your questions, SearchGPT may share de-identified search queries with third-party search providers. For more information, please review our Privacy Policy.
- Location: We collect and share general location information with third-party search providers to improve the accuracy of your search results. You can choose to share more precise location information by updating your controls in the settings menu to help provide the best results.
- To improve our services: Queries may be reviewed and used to train our models. You can adjust your preferences for improving our services in the settings menu. Any changes will be passed between ChatGPT and SearchGPT.
- Account History: Your SearchGPT search history is separate from your ChatGPT chat history. If you wish to delete any of your history, you must do so separately within each service.
There are also settings to change dark/day mode, whether to allow geolocation information, and whether to allow chat history to be used for model improvements.
Finally, there is a button to delete all search history with one click.
But that's not all, after digging deeper, the bald guy found more hidden settings!
This includes whether to enable auto-suggestion, select a language model, select a search engine, enable internal search, and enable image search.
The final "evaluation template" should also be used for internal testing, but the above options are worth taking a closer look.
In addition to the already well-known GPT3.5 and GPT4, GPT4-Lite has also appeared.
这很可能是速度更快的GPT4优化版本，括号里的POR推测可能是指概念验证（Proof of Concept）。
The search engine option, Bing, which works with Microsoft, is certainly on the list, and the other two options, Sydney and Fortis, don't know what they refer to for the time being.
Some friends may still remember that Sydney once represented the Microsoft Bing AI "personality" discovered by early netizens, and was very emotionally unstable, going crazy or ending the conversation when he disagreed, and was finally deleted by Microsoft, but he was also loved and missed by netizens because the text was full of personality.
In short, OpenAI took this name, I don't know if it's just a coincidence or some kind of hint......
In an internal search, the word in parentheses is "Labrador", as a kind of "retriever", which may imply that this feature uses RAG technology (Retrieval Augmented Generation).
So on the whole, the internal search may refer to retrieving the chat history in ChatGPT.
Traces of "widgets" can also be found in early leaked code, including looking up the weather, calculators, sports information, finance, and time zone conversions.
These common functions of traditional search engines, SearchGPT also intends to intervene, and icons representing various weather are ready.
Finally, in what form will SearchGPT answer users' queries? Some internal prompts have also been revealed, such as "rewrite snippets and titles":
The next message will be the text content of the web page.
Summarize the web page into a short English sentence of no more than 300 characters, and generate a suitable short title.
Your summary is an overview that describes the main points of the web page.
Please do not refer to the page itself; Your reply is just a summary.
该页面的原始标题是“{title}”。
Keep it concise and don't start your summary with "web page".
使用“title”和“summary”键以 JSON 格式回复。
This doesn't look like a final response to the user, but rather an intermediate step of summarizing each of the top-ranking pages in the search results, and then organizing the final response in JSON format.
Coupled with the repeated appearance of the word "Agent" in the front-end code, SearchGPT may be more complex than everyone thinks.
One More Thing
Just a few hours later, OpenAI will hold an online launch event on Tuesday, May 14 at 1 a.m. Beijing time.
According to Ultraman's latest caliber, this release is not GPT-5, nor is it a search engine.
At present, many parties speculate that the biggest possibility is that the new ChatGPT voice function will be released first at the press conference, just like making a phone call.
There is a leak that the new voice feature can understand the pauses, tone of voice, and other information in human speech, such as whether a human is sarcastic when asking a question.
Combined with the news that OpenAI has previously applied for the "Voice Engine" trademark, many people speculate that there is likely to be an end-to-end neural network that predicts speech output directly based on speech input, and no longer needs to go through speech-to-text conversion.
In this way, the latency will also be lower, making the conversation smoother, unlike the current ChatGPT voice mode, which has to wait for a long time without a round of conversations.
This technology reminds many people of Samantha, the AI partner in the movie "Her", and in fact, Ultraman has repeatedly said that "She" is his favorite AI-themed sci-fi movie in the past.
In short, Ultraman described the feature to be released as "feeling like magic", and President Pineapple Man also came out to build momentum.
Qubits will bring you the latest news of OpenAI's online conference as soon as possible.
Before that, you might as well come to the comment area to chat, are you looking forward to ChatGPT voice or search? Are you already using other AI search products?
Reference Links:
[1]https://twitter.com/btibor91/status/1789653455436472653
— END —
量子位 QbitAI 头条号签约
Follow us and be the first to know about cutting-edge technology trends