HargaRumah: OpenAI’s Big Codex Push

Plus: AI startups are using this tactic to boost valuations.

BECOME A MEMBER • Newsletters • MyForbes

OpenAI relaunched Codex as a separate desktop app in February.Anadolu via Getty Images

Welcome to The Prompt,

ChatGPT is about to get a lot more powerful.

That's because OpenAI plans to merge ChatGPT and Codex, its coding agent, into a single app. ChatGPT, now supercharged with agents that can search the web on your behalf, will suddenly be able to do things like send you a high level review of your emails and meetings for the day, analyze information from spreadsheets and organize the files on your computer.

The combined tool is geared towards working professionals, entrepreneurs and everyday users, says Tara Seshan, a member of product staff at OpenAI. The promise: people can now delegate complex tasks to AI agents to free up time for creative, high-stakes work, she says. Codex can be helpful for a wide range of tasks like generating spreadsheets for monthly reports, using a call transcript to prepare a brief and log it into Salesforce or spinning up six different types of content from one piece of text.

It's an attractive pitch. But it can take a fair deal of time to learn how to use Codex efficiently and set up automated tasks. Even then, you have to give it clear guardrails, as it has a tendency to run on a single task for hours on end without generating a usable answer. (In one instance when I asked Codex to extract the names of startups from a tweet and calculate how much they’ve raised, it ran for more than two hours.) “I think the best users of these tools really give context, define what good looks like, set boundaries, review outputs, and they make judgment calls about the end thing,” Seshan says.

OpenAI relaunched Codex as a desktop app in February. It’s now used by 5 million weekly active users (compared to 900 million for ChatGPT). OpenAI also teased new hardware for Codex that it plans to launch in mid-July— a keyboard-like device where you can assign shortcuts to different buttons.

“We've gone through just a really big shift. The first wave of AI was to chat with an assistant,” Seshan says. “Now everyone can literally assign work to an agent. You can describe the outcome, let the agent gather its context, watch the progress when needed, and get a polished artifact at the end.”

Last week, OpenAI unveiled its first custom AI chip designed to run ChatGPT, Codex and other models more efficiently than existing hardware. Called Jalapeno and developed with Broadcom, the processor is built to handle a computationally intensive process called “inference,” where a trained AI model produces an output based on new, unseen data. It’s part of the chatbot maker’s expansion beyond consumer products to offer the infrastructure needed to run its models.

OpenAI faces mounting pressure from rivals, the U.S. government — which increasingly wants a say in frontier AI model launches — and the public markets. After initially planning to file to go public in the third or fourth quarter of this year, the company is leaning towards delaying its IPO till 2027 after advisers cautioned about the volatility of the public tech markets, the New York Times reported. The company, which has not yet turned a profit, is hoping to go public at a $1 trillion valuation.

Now let’s get into the headlines.

Rashi Shrivastava Staff Writer

Follow me on Forbes.com

Big Plays

Meta hired hundreds of contractors to pose as minors by setting up under-18 accounts and prompt AI chatbots developed by rivals like OpenAI, Google and Character AI to churn out answers related to suicide, sex, eating disorders and other high-risk topics, Wired reported. The project, internally called Cannes, included thousands of prompts that pushed the chatbots to respond to questions blocked by their safety systems. Meta defended the work as routine safety testing.

Humans Of AI

With newfound wealth flooding San Francisco and scores of employees set to become millionaires and billionaires when OpenAI and Anthropic go public, some tech workers say their six-figure salaries are barely helping them keep up with skyrocketing rental prices in the city, the New York Times reported. Some residents say it’s becoming harder than ever to find affordable accommodation, wondering if they can afford to live there long term.

AI Deal Of The Week

AI research company General Intuition has raised $320 million in Series A funding at a $2.3 billion valuation, Techcrunch reported. The New York-based startup uses billions of gaming video clips from CEO Pin De Witte’s other startup, Medal, to train AI models. The data helps them learn how to interpret and navigate physical worlds — which could be helpful for robotics.

Illustration by Macy Sinreich for Forbes; images by alexsl, ismagilov, and frentusha via Getty Image

Deep Dive

AI Startups With No Revenue Are Using This Tactic To Supersize Their Valuations

Read Article

Earlier this year, David Silver, the renowned former scientist at Google DeepMind, dialed into a Zoom meeting with a venture capital firm to pitch his new startup Ineffable Intelligence. Silver, who spent over a decade at Google, believed the current approach to training AI models wouldn’t work long term. He spoke for 30 minutes about creating digital environments where AI systems could learn on their own, without any human data or input.

Eventually, he told the investors on the call, AI will get so good that it will learn to use things in the physical world. “We can put AI in our toasters,” an investor recalls him saying.

That seemed like a stretch. “I was so excited for that pitch and it was just so absurd,” the investor says. “He just rambled, no deck, no memo.”

By the end of the pitch, the VC wasn’t convinced. “We left that pitch with more questions than answers. And it's so jarring to hear that from somebody with such merit and stature in the industry,” they say.

Despite no near-term plans to launch products, let alone make any money off them, Silver ultimately raised $1.1 billion in seed funding for Ineffable Intelligence. The fundraise, which was touted as Europe’s largest seed round ever, valued the nascent startup at a whopping $5.1 billion valuation.

At least, that’s what the headlines said.

In reality, the company raised funding in two tranches. In the first, Ineffable raised $11 million from Sequoia and other investors, valuing the startup at about $55 million pre-money, according to company filings. But within a month or so, the company raised an additional $1.1 billion at a much, much higher valuation of $4 billion pre-money — the price paid by investors like Lightspeed, Index Ventures and DST Global, as well as Sequoia, which also invested at the higher price tag. That’s over 70 times the price for the same company, just weeks apart.

The majority of the capital Sequoia invested was at the higher valuation, according to a person familiar with the deal. Sequoia and Ineffable Intelligence declined to comment.

Tranched rounds like this have become increasingly common in the AI funding frenzy, especially for neolabs— companies that raise billions of dollars right out of the gate to focus on frontier research instead of developing products. Building an AI lab requires tons of GPUs, meaning they need to raise massive amounts of capital right out of the gate to carry out serious research. Because these founders need to raise so much, lead investors wanted better terms in return, says Zach DeWitt, a partner at Wing VC.

Read the full story on Forbes.