Google consistently builds cool AI products, while OpenAI seems to wait for these launches to steal the spotlight.
“Be OpenAI, wait for Google to drop a cute math model, then launch a competing search engine that could potentially blow up Google’s $2T internet search monopoly and send Google execs into existential dread,” quipped Aidan McLau, the chief executive of Topology Invest.
A few hours after Google launched its new model capable of solving International Math Olympiad problems, OpenAI grabbed attention with its new Google alternative, SearchGPT. It combines OpenAI’s AI models with real-time web information to provide fast and relevant answers to user queries.
Currently in the prototype phase, it is available to a limited group of 10,000 test users.
“We believe there is room to make search much better than it is today. We are launching a new prototype called SearchGPT. We will learn from this prototype, improve it, and then integrate the technology into ChatGPT to make it real-time and maximally helpful,” said OpenAI chief Sam Altman.
This isn’t the first time OpenAI has done something like this. Earlier this year, when Google released Gemini 1.5, OpenAI announced Sora on the same day. Then, just a day before Google I/O 2024, OpenAI hosted its Spring Update and released GPT-4o.
However, OpenAI has yet to make Sora publicly available, and the voice features for ChatGPT are still not accessible.
Who knows, OpenAI might have achieved AGI internally. “OpenAI has probably already achieved gold in the Math Olympiad—something even the most optimistic AI researchers would not have expected before 2025,” posted a user on X who goes by the name Chubby.
This might actually be true, as Altman responded to Google DeepMind’s IMO score with a simple ‘lol’.
Math all the Way
Solving problems at the IMO Olympiad is any day a greater achievement than launching an AI-based web search.
“AlphaProof is one of the most exciting applications of LLMs combined with RL. The Gemini model automatically translates natural language problem statements into formal statements (i.e., formalizer network),” said Elvis Saravia, the co-founder of DAIR.
“LLMs are alien beasts. It is deeply troubling that our frontier models can both achieve a silver medal in the Math Olympiad and fail to answer. “Which number is bigger, 9.11 or 9.9?’” said Jim Fan, the lead of Embodied AI (GEAR Lab) at NVIDIA.
He further explained that AlphaProof and AlphaGeometry-2 are trained on formal proofs and domain-specific symbolic engines. “In a way, they are highly specialized towards solving Olympiads, even though they build on a general-purpose LLM base,” he said.
Meanwhile, OpenAI is quietly developing Project Strawberry to significantly boost the reasoning capabilities of its AI models. While details about the project remain undisclosed, it focuses on a novel approach that enables AI to plan ahead and autonomously navigate the internet for in-depth research.
Internally, OpenAI has tested AI that scored over 90% on the MATH dataset, which benchmarks championship-level math problems. This progress aims to tackle current AI reasoning limitations, such as common sense issues and logical errors that often lead to inaccurate outputs.
Previously known as Project Q*, which was leaked last year and could solve new math problems, Project Strawberry is now working to enhance long-horizon tasks (LHT). This involves a specialised “post-training” phase, adapting base models for better performance, similar to Stanford’s 2022 Self-Taught Reasoner (STaR), which enables AI to generate its own training data for improved intelligence.
Do we Really Need SearchGPT?
While SearchGPT is a nice feature to have in ChatGPT, it appears that Perplexity AI has already taken the lead in this segment.
“By the way, Perplexity is awesome. They made me rethink what search & AI integration could be. It’s often the first place I go to now when I need to start researching a new topic,” posted Lex Fridman on X.
“SearchGPT is like a nice feature at this point. Nobody can take the crown from Perplexity after they introduced agentic search – it’s simply too good and OpenAI cannot just steamroll them,” posted a user on X.
Building a tool like Perplexity AI isn’t particularly difficult today, so it’s puzzling why it took OpenAI so long. “This is cool, but the name sounds like something a high schooler would put on their resume as their first solo project,” joked a user on X.
On a related note, Bishal Saha, a dropout from Lovely Professional University, created Omniplex, an open-source alternative to Perplexity AI, over a single weekend.
It’s unclear how concerned Perplexity AI chief Aravind Srinivas is right now, but Google’s stock did plummet after the OpenAI SearchGPT demo. In response to competitors like Perplexity AI and ChatGPT, the search giant introduced ‘AI Overviews’ at Google I/O 2024. This feature generates summaries for user queries.
