Elon Musk unveils xAI's first LLM, Grok


VentureBeat presents: AI Unleashed – An exclusive executive event for enterprise data leaders. Network and learn with industry peers. Learn More


Following up on his proclamation last week that xAI would begin allowing selected users access to its first AI product, founder Elon Musk on Sunday morning revealed it to the world, and it is very much aligned with his sensibilities and often irreverent and immature sense of humor, while boasting access to realtime information and high efficiency.

The product is a Large Language Model (LLM) called “Grok,” named after the slang term that means “understanding,” and is built to compete with other leaders in the space such as OpenAI’s GPT and Anthropic’s Claude 2.

“Just released Grok,” Musk posted on his X social network at nearly 1 am Eastern on Sunday, November 5, 2023.

Musk’s post contained a link to the xAI website which states that Grok is currently available to “a limited number of users in the United States,” and that prospective users can join its waitlist to gain early access, though to do so requires an account on the X social network (formerly Twitter). There was no cost listed to use Grok.

Event

AI Unleashed

An exclusive invite-only evening of insights and networking, designed for senior enterprise executives overseeing data stacks and strategies.

 


Learn More

The xAI website goes on to provide many more details about how Grok was built and trained, including the facts that it started with a prototype model “Grok-0” trained on 33 billion parameters of data, compared to 70 billion for the new Meta LLama 2 and an apparent 20 billion for OpenAI’s older GPT-3.5 models.

Impressively, xAI claims on its site that Grok-0 “approaches LLaMA 2 (70B) capabilities on standard LM benchmarks but uses only half of its training resources.”

The xAI team is said to have “made significant improvements in reasoning and coding capabilities,” enough to create a new model, Grok-1, which is the “frontier LLM” powering the Grok chatbot client, similar to how OpenAI’s GPT model powers its ChatGPT consumer-facing experience.

xAI also posted a chart showing Grok’s performance in four categories of machine learning (ML) benchmarks and tasks, including middle school math (GSM8k), multiple choice questions (MMLU), Python code completion (HumanEval), and math problems written in LATEX (MATH).

Grok “surpass[es] all other models in its compute class, including ChatGPT-3.5 and Inflection-1,” xAI’s website states of the performance. “It is only surpassed by models that were trained with a significantly larger amount of training data and compute resources like GPT-4. This showcases the rapid progress we are making at xAI in training LLMs with exceptional efficiency.”

Chart from xAI on Grok-1 performance on machine learning benchmarks. Credit: xAI

A ‘humorous’ AI inspired by ‘Hitchhiker’s Guide’

On xAI’s website, Grok is described as “an AI modeled after the Hitchhiker’s Guide to the Galaxy,” the seminal 1970s radio drama and satirical sci-fi book series by UK author Douglas Adams (it was adapted into a major movie in 2005).

In the inaugural book, a telepathic alien organism called the Babel Fish is placed into the protagonist’s ear, allowing him to automatically translate and understand alien speech. The book is also famed for a supercomputer revealing the meaning of the universe to be the number 42, which aligns with Musk’s previously stated goal of making xAI’s product into a “maximum truth-seeking AI.”

The xAI webpage goes on to describe Grok as being “intended to answer almost anything and, far harder, even suggest what questions to ask! Grok is designed to answer questions with a bit of wit and has a rebellious streak, so please don’t use it if you hate humor!”

Earlier on the evening of Friday, November 3 ET, Musk shared screenshots on X of Grok’s responses to user prompts that showcased its humor, including a step-by-step response on “how to make cocaine” that included sarcasm and the warning “start cooking and hope you don’t blow yourself up or get arrested. Just kidding! Please don’t actually try to make cocaine.”

“It’s also based & loves sarcasm,” Musk added in another X post early in the morning ET on Nov. 4, “I have no idea who could have guided it this way…” followed by emoji, intimating he was the one who directed the product to have these qualities.

How is xAI using X (formerly Twitter) posts/data for its Grok LLM?

On a more serious note, the “unique and fundamental advantage of Grok” according to its creators “is that it has real-time knowledge of the world via the X platform,” formerly Twitter, the social network Musk acquired in October 2022 after a lengthy-back-and-forth publicly messy negotiation process, and which his tenure has since more than halved in valuation.

In a post on X made early Sunday morning ET, Musk included two screenshots demonstrating how Grok could return more recent information than a “typical GPT,” in this case, the Phind v7 model based on Meta’s Code Llama.

“Grok has current information, but other doesn’t,” Musk wrote, showing how a user was able to ask about “Elon’s last interview with Joe Rogan” and what the podcast host Rogan was wearing, and receive accurate results.

It wasn’t immediately disclosed how xAI was using the X social network or users’ posts to train Grok, but Musk previously announced he was cutting off OpenAI’s access to X/Twitter’s database for training purposes, a move rich with dramatic irony since Musk himself bankrolled and co-founded OpenAI in 2015, before exiting the company years later after a reported failed coup to take control from co-founder and current CEO Sam Altman.

Musk’s screenshots also included the user prompting Grok with the command “/web” demonstrating it has some web browsing or searching capabilities, which his rival OpenAI restored to ChatGPT in September, following a six-month-long hiatus due to people using it to bypass news publisher paywalls.

Plans to expand Grok’s availability

Earlier on Friday evening ET, Musk posted on X that Grok’s availability would be expanded to “all X Premium+ subscribers” when it’s out of early beta, but did not provide an exact nor approximate timing estimate on when this might occur.

Nonetheless, the move to begin sharing Grok screenshots and limited availability to a subset of users suggests Musk’s desire to move fast to compete with his former business partners at OpenAI as the latter prepares to announce a slew of new AI features on Monday, Nov. 6, at its first DevDay developer conference in San Francisco.

VentureBeat’s mission is to be a digital town square for technical decision-makers to gain knowledge about transformative enterprise technology and transact. Discover our Briefings.





Source link