Creator of ChatGPT hints at new breakthroughs

What is the mysterious language model gpt2 that Sam Altman, the CEO of OpenAI, is sending teasing tweets about? And is the creator of ChatGPT working on a Google killer? AI fans on social media are expecting big announcements.

OpenAI is working on an adjustment to the popular chatbot ChatGPT, allowing it to finally search the web and cite sources. The Bloomberg news agency learned this. The AI ​​company will thus become a competitor of Google and the AI ​​search start-up Perplexity, but also of Bing Copilot from ally Microsoft.

One version of the product uses images in addition to text responses. OpenAI is under intense pressure to expand the capabilities of its best-known product. There is an increasing number of competing chatbots and web search is an important application. Perplexity, also from California, offers an AI-powered search engine that focuses on accuracy and citations. That start-up has a valuation of $1 billion. Google will unveil its plans for its Gemini AI models at its annual I/O event next week.

By the way, OpenAI’s subscription product already searches the web in response to some instructions that also link to resources. Just like with Bing and Gemini, things sometimes go wrong with searches and the interpretation of the results.

On X, formerly known as Twitter, there was some excitement about the existence of the web page search.chatgpt.com. Visitors to that URL will see a small print “not found” message on the page, but for a short time the page did link to ChatGPT’s official website, chatgpt.com.

OpenAI may be deliberately building up the tension with search.chatgpt.com. The company also seems to be doing the same with the speculation about new, superior versions of the chatbot. Sam Altman, the CEO of OpenAI, posted on X at the end of April: ‘I have a soft spot for gpt2’. Definitely gpt2 and not GPT-2, as an old model is called. Gpt2 was briefly available on a test site and then disappeared again. This week the model reappeared on the website chat.lmsys.org, where users can test and compare different language models.

The site has an ‘arena (battle)’ section where the user enters a text instruction (prompt), and where two different models respond in two text boxes. It can then be indicated which answer was the best, after which the two models are also revealed. Testers received answers from the models ‘im-a-good-gpt2-chatbot’ and ‘im-also-a-good-gpt2-chatbot’, either after repeated attempts or not. The models come from OpenAI.

According to numerous commenters on Both for programming and for reasoning and structured answers. Observers think they are part of an upgrade of OpenAI’s models, a kind of GPT-4.5. It is unlikely that OpenAI will immediately jump to GPT-5, because the top of the company fears negative reactions if a too surprising performance jump is launched. During the evolution to GPT-3.5 and GPT-4, there was already heavy criticism that OpenAI was moving irresponsibly fast with its models. Meanwhile, Altman said that compared to the strongest models now being tested internally, GPT-4 is particularly stupid.