Releasing ChatGPT made OpenAI the poster company of the AI race. But winning it is proving really hard.

• ChatGPT put OpenAI at the forefront of the drive to build human-like AI.
• But the momentum of that hugely impactful launch has been tough to maintain.

"Try talking with it here," Sam Altman wrote on X, linking to an "early demo" of ChatGPT.

He gave some thoughts on "language interfaces" and how people would soon "talk to the computer" with voice or text to get what they want, "for increasingly complex definitions of want!"

That thread marked what turned out to be one of the most impactful launches in Silicon Valley history.

ChatGPT prompted a surge of interest in generative AI. Tech giants like Microsoft, Apple, Google, and Meta poured billions of dollars into the technology, and venture capitalists pumped even more into startups focusing on it.

It made OpenAI the poster child of the race to create human-like artificial general intelligence, which has become Altman's stated mission.

But getting there from the ChatGPT demo he debuted in November 2022 is a long road. Few would've mistaken the chatbot — with a tendency to hallucinate and get basic facts wrong — for a human. The company is struggling under the weight of expectations the initial launch created — and following up with launches as smooth and impactful is getting tougher.

Meanwhile, rivals like Google, Meta, and smaller companies have tried to catch up by unveiling their own AI models, such as Gemini Ultra and Llama 3.

OpenAI's new products have had issues

On Tuesday, OpenAI said it was delaying a highly anticipated upgrade to its chatbot's "voice mode" feature, which had been due for imminent release.

The existing voice feature, first introduced in September 2023, had suffered delays in responding to users' attempts to have a back-and-forth conversation and struggled to "directly observe tone," as well as multiple speakers and background noises, the company said.

The upgrade was set to roll out to a small group of ChatGPT Plus users in late June until the company announced it needed to work on "the model's ability to detect and refuse certain content" and to prepare "infrastructure to scale to millions while maintaining real-time responses."

The hiccup came after OpenAI made headlines when Scarlett Johansson accused the company of touting a voice — "Sky" — that, she claimed, was "eerily similar" to hers.

Though Altman later claimed it "was never intended to resemble" Johansson's, the company decided to pause using the voice.

The voice feature isn't the only one having a hard time.

OpenAI still has no release date for Sora, the model that generates videos based on users' text prompts and was first demonstrated publicly in February. In May, Google boss Sundar Pichai said OpenAI's training of Sora may have violated YouTube's terms and conditions.

Open AI released GPT-4o, its most advanced multimodal AI model and the successor to GPT-4, in May but rival AI lab Anthropic showed data on Claude 3.5 Sonnet, which it released last week, that suggest its model beats GPT-4o on "graduate-level reasoning," "code," and "reasoning over text" benchmarks.

According to Chatbot Arena, a widely followed performance table, Claude 3.5 Sonnet beat GPT-4o in coding. GPT-4o maintains the overall top spot on the table but only by a narrow margin.

This month did give OpenAI something to celebrate; Apple announced that it was partnering with the company to bring ChatGPT to iPhones, iPads, and Macs. This gives OpenAI access to Apple's huge install base of more than 2 billion active devices.

But OpenAI isn't getting paid for the integration, and Apple didn't guarantee exclusivity, as it has signaled it's open to partnering with other companies on AI.

ChatGPT gave OpenAI a headstart in the race to make AGI. But in a race that long, anyone could win.