On the day before the OpenAI conference, employees collectively went crazy... staging a large-scale nesting doll performance.
A: I am excited for B's excitement; B: I am excited for C's excitement... Z: I am excited for these upgrades
At the same time, there were also continuous small actions, such as now GPT-4's text description is no longer 'state-of-the-art model,' but just 'advanced'.
And the pigeonholed GPT search engine, prototype code uploaded to ChatGPT.
Hacker 'Bald Brother' TiborBlaho, scoured the updated front-end code and extracted SearchGPT's interface.
From the leaked interface, it appears that SearchGPT will appear in the sidebar, along with other GPTs.
The homepage is currently very concise, like Google with just a centered search box. It even provocatively designed a 'compare' button, which can directly open Google and Perplexity search results for comparison with SearchGPT. (feeling like a face-off challenge)
However, Bald Brother believes that this is likely only a temporary setting for internal testing and evaluation, not necessarily provided to users.
Although not yet officially released, from the exposed interface and text descriptions, it is already preparing for internal testing, and the day of meeting everyone may not be far away.
Let's start with the 'Welcome page', currently clearly stating that SearchGPT is 'under construction,' which looks like an explanation prepared for internal test users from the content and tone.
Under construction: SearchGPT is constantly changing and improving, it may make mistakes or be unavailable at times.
Information sharing: To search and answer your questions, SearchGPT may share de-identified search queries with third-party search providers. For more information, please see 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.
Improving our service: Queries may be reviewed and used to train our models. You can adjust your preferences for improving our service in the settings menu. Any changes will be communicated between ChatGPT and SearchGPT.
Account history: Your SearchGPT search history is separate from your ChatGPT chat history. If you want to delete either history, you must do so separately in each service.
In the settings, there are options for changing dark/light mode, allowing geolocation retrieval, allowing chat records for model improvement.
Finally, there is a button to delete all search records with one click.
However, this is not all, after digging deeper, Bald Brother found more hidden settings! Including whether to enable auto suggestions, select language models, choose search engines, enable internal search, and enable image search.
The 'evaluation template' should probably be temporarily unused for internal testing, but the options above are worth looking into.
The language model options, apart from the known GPT3.5 and GPT4, there is also GPT4-Lite.
This is likely a faster version of GPT4 in optimization, with 'POR' in parentheses possibly meaning Proof of Concept.
The search engine options, Bing in partnership with Microsoft is definitely listed, while the other two options Sydney and Fortis are currently unknown.
Perhaps some friends still remember that Sydney once represented the early netizens' discovery of Microsoft Bing AI 'personalities,' with very unstable emotions, going crazy or ending conversations at a whim, and ultimately being tragically deleted by Microsoft. However, i t was also cherished and remembered by netizens for its text full of personality.
Overall, naming it OpenAI, whether it is just a coincidence or some kind of suggestion...
Internal search, with the word 'labrador' in parentheses, as a 'retriever,' possibly implies that this function uses RAG technology (Retrieval-Augmented Generation).
So, looking at it comprehensively, internal search may refer to retrieving chat records within ChatGPT.
In the early leaked code, traces of 'widgets' can be found, including weather queries, calculator, sports match information, finance, and time zone conversion.
These traditional search engine common features, SearchGPT also intends to get involved, with icons representing various weather conditions already prepared.
Finally, in what form will SearchGPT answer user queries? Some hint words internally have also been exposed, such as 'rewrite snippets and titles':
The next message will be the text content of the webpage.
Summarize the webpage into a short English sentence, no more than 300 characters, and generate a suitable short title.
Your summary is an overview that describes the key points of the webpage.
Please do not refer to the webpage itself; your response is just a summary.
The original title of the page is '{title}'.
Keep it concise, do not start the summary with 'webpage'.
Use 'title' and 'summary' keys to reply in JSON format.
It doesn't look like the final response presented to users, but rather an intermediate step, summarizing each page that ranks high in search results, concatenating the summaries in JSON format, and then organizing the final response.
And with the frequent appearance of the 'Agent' in the front-end code, SearchGPT may be more complex than people imagined.
Just a few hours later, OpenAI will hold an online press conference on early Tuesday morning, May 14th Beijing time.
According to the latest Altman interpretation, this release is not GPT-5, nor a search engine.
Currently, the most likely speculation is that the conference will first release new ChatGPT voice functions, just like making phone calls.
Leaked information suggests that the new voice function can understand pauses and tones in human speech, such as determining whether there is irony when humans ask questions.
Combined with OpenAI's previous application for the 'VoiceEngine' trademark, many people speculate that there may be an end-to-end neural network, directly predicting voice output based on voice input, without the need for voice-to-text conversion.
This way, the delay will be lower, making the conversation smoother, unlike the current ChatGPT voice mode requiring long waits every turn.
This technology reminds many people of the AI companion Samantha in the movie 'Her,' in fact, Altman has also expressed many times that 'Her' is his favorite AI-themed science fiction film.
Overall, Altman describes the upcoming feature as 'feeling like magic,' and CEO Pineapple is also out there stirring up support.