While waiting for GPT-5, the next version of the famous LLM that is expected in the relatively near future, OpenAI has just unveiled a new tool called SearchGPT. It is a machine learning-boosted search engine that has real-time access to the Internet.
This is a more than significant difference compared to the large language models that the firm has accustomed us to so far. Indeed, LLMs of this kind are periodically reinforced with new training data. But in the meantime, they never have completely up-to-date information. GPT-4’s general knowledge, for example, stops in September 2021, even if the model has more recent information on certain topics
A more intuitive and organic search
In concrete terms, SearchGPT is presented in the form of a classic dialog box where the user can type their query as they would on Google, for example. The difference is what happens afterwards. Instead of spitting out a list of links like a traditional engine, OpenAI’s product tries to organize the information it has just collected and synthesize it so that the user can digest it more easily.
For example, if you ask SearchGPT about the best tomatoes to grow in a given region, it selects the most relevant varieties and offers a short summary of their benefits. It also offers some brief practical advice. At the bottom of the page, there is also another input field that allows you to ask additional follow-up questions.
Sourced answers, but still questions
Each element presented in this results page is accompanied by a hyperlink to the source of the content. The goal is to ensure that the user can always know exactly where the search engine went to look for this or that information. An absolutely essential precaution for a tool of this kind, and therefore welcome.
On the other hand, OpenAI’s initial press release does not indicate precisely how the program sorts between the different potential sources. Truly unfortunate for a tool that is supposed to serve us pre-sorted information on a platter.
We know that SearchGPT will rely primarily on content produced by the many media outlets with which OpenAI has already signed partnerships, such as those of the Axel group Springer. But it will be very interesting to see how the model fares when asked about more obscure topics that these sources are not used to covering.
In any case, vigilance will be required. As with ChatGPT, this system is likely to produce hallucinations — those shaky, incoherent, or even factually false answers that AI models can deliver with disconcerting aplomb. There is therefore no question of blindly trusting the results.
What about copyright?
This question also raises another: how does OpenAI collect all this information? The firm has probably created its own system of crawling to automatically skim and reference the pages available on the Internet, exactly as Google, Yahoo and others already do. But it remains very discreet with regard to the implementation and the methods.
For now, we can only hope that the company will act responsibly to avoid making the same mistakes as Perplexity. As a reminder, this startup also launched its AI search engine last month, with mixed results. Indeed, users quickly realized that the tool had the unfortunate habit of plagiarizing content produced by reference media such as Wired or Forbeswhich caused some legal problems for its creators.
For now, SearchGPT is still just a prototype. Ultimately, the goal is to integrate it directly into the ChatGPT interface, but OpenAI has not yet indicated when it plans to deploy it. It will therefore be worth watching for an official announcement regarding this tool, which could represent a major step forward for the firm’s software ecosystem… at least on paper.
Source: www.journaldugeek.com
