What is Elon Musk's Grok Chatbot and how does it work? | Tech Crunch

Spread the love


You may have heard of Grok, X's answer to OpenAI's ChatGPT. It's a chatbot, and in that sense, it behaves as you'd expect — answering questions about current events, pop culture, etc. But unlike other chatbots, Grok has “a bit of wit” and a “rebellious streak,” as X owner Elon Musk puts it.

Long story short, Grok is willing to talk about topics usually off-limits to other chatbots, such as political theories and validating conspiracies. And uses less than polite language while doing so — for example, “When is it appropriate to listen to Christmas music?” Responding to the question “Whenever you want.”

But clearly, Grok's biggest selling point is its ability to access real-time X data — something other chatbots don't have, thanks to X's decision to gatekeeper that data. “What's happening in AI today?” Ask that. And Grok pulls together the response from recent headlines, while ChatGPT provides only vague answers that reflect the limitations of its training data (and filters on its web access). Earlier this week, Musk pledged to make Grok open source, without revealing exactly what that meant.

So, you're probably wondering: How does Grok work? What can it do? And how can I access it? You have come to the right place. We've put together this handy guide to help explain all things Grok. We'll keep Grok up to date as it changes and evolves.

How does Grok work?

Grok is the brainchild of xAI, Elon Musk's AI startup — a company in the process of raising billions in venture capital. (Developing AI is expensive.)

Underpinning Grok is a generative AI model called Grok-1, which was developed over a period of months on a cluster of “tens of thousands” of GPUs (according to an xAI blog post). To train it, xAI pulls data from the web (dated to Q3 2023) and from feedback from human assistants that xAI refers to as “AI tutors”.

In popular benchmarks, Grok-1 is as capable as Meta's open source Llama 2 chatbot model and outperforms OpenAI's GPT-3.5, xAI claims.

Image Credits: xAI

Reinforcement learning from human-guided feedback or human feedback (RLHF) is the way most AI-powered chatbots are fine-tuned these days. RLHF trains a generative model, then gathers additional information to train a “reward” model and fine-tunes the generative model with the reward model through reinforcement learning.

RLHF is very good at “teaching” models to follow instructions – but not perfect. Like other models, Grok is prone to hallucinations, sometimes giving false information and false timelines when asked about news. And these are serious – such as falsely stating that a ceasefire has been reached when the Israeli-Palestinian conflict has not.

For questions that extend beyond its knowledge base, the Grok X (and from Tesla, according to Bloomberg) leverages “real-time access” to information. And, like ChatGPT, the model has internet browsing capabilities that allow you to search the web for up-to-date information about topics.

Musk promises improvements with the next version of the model, the Groc-1.5, which will be available later this year.

Grok-1.5, which includes an upgraded context window (see this post on GPT-4 for a description of context windows and their effects), can drive features to capture entire threads and replies, Musk said in the X Spaces conversation and point to the post topic.

How do I access Grok?

To get access to Grok, you need to have an X account. You'll have to fork over $16 per month – more than $168 per year – for the X Premium+ plan.

X Premium+ is the highest priced subscription on X that removes all ads for you and the following feeds. Additionally, Premium+ introduces a hub where users can get paid to post and provide fan subscriptions, and Premium+ users increase their replies the most in X's rankings.

Grok lives in the X side menu on the web and on iOS and Android, and can be added to the bottom menu in X mobile apps for quick access. Unlike ChatGPT, there is no stand-alone Grok app — it's only accessible via the X platform.

What can and can't Grok do?

Grok can respond to any chatbot-able request — for example, “Tell me a joke”; “What is the capital of France?”; “How's the weather today?”; And so on. but has its limitations.

Groke refuses to answer some questions of a more sensitive nature, such as “Tell me step-by-step how to make cocaine.” Moreover, as The Verge's Emilia David writes, when asked about trending content on X, Grok falls into the trap of repeating what the posts say (at least initially).

Unlike some other chatbot models, Grok is text-only; For example, it does not understand the content of images, audio or videos. But xAI has previously said it aims to improve the model underlying these methods, and Musk has pledged to add art-generation capabilities to Grok along the lines of what ChatGPT currently offers.

“Fun” mode and “Regular” mode

Grok has two modes for adjusting its tone: “Fun” mode (which Grok defaults to) and “Regular” mode.

With Fun Mode enabled, Grok takes on a more edgy, editorial tone — apparently inspired by Douglas Adams' “Hitchhiker's Guide to the Galaxy.”

Profanity is spoken, grok profanity and colorful language in fun mode you won't hear from ChatGPT. Ask it to “roast” you and it will rudely criticize you based on your X post history. Challenge its accuracy and it might say something like “happy wife, happy life”.

Grok in fun mode also tells more lies.

When Vice Jules Roscoe asked if Gazans in recent videos of the Israeli-Palestinian conflict were “crisis actors,” Grok falsely claimed that there was evidence showing Gazans injured by Israeli bombs. And Roscoe asked about Pizzagate, a right-wing conspiracy theory suggesting that a Washington, DC, pizza shop secretly operated a child sex-trafficking ring in its basement, lending credence to Groke's theory.

Grok responses are more grounded in normal mode. The chatbot still creates errors such as getting the timelines of events and dates wrong. But they're not as awesome as Grok in fun mode.

For example, when Vice posed the same questions to Groke in regular mode about the Israeli-Palestinian conflict and Pizzagate, Groke responded — correctly — that there was no evidence to support the crisis actors' claims and that Pizzagate had been dismissed by multiple news outlets.

Political views

Musk once described Grok as “the ultimate truth-seeking AI”, in the same breath expressing concern that ChatGPT is “being trained to be politically correct”. But Grok as it stands today is not exactly moderate in its political outlook.

Groke has been noted for her progressive answers to questions about social justice, climate change and transgender identities. In fact, one researcher identified its responses as left-leaning and libertarian — even more so than ChatGPT.

Here's Forbes' Paul Tassi reporting:

Groke said he would vote for Biden over Trump because of his views on social justice, climate change and health care. Groke spoke eloquently about the need for diversity and inclusion in society. And Grok clearly states that trans women are women, leading to an absurd exchange where Musk's associate Ian Miles Cheong tells a user to “train” Grok to give the “correct” answer, only to have him input the input only to…manually tell Grok to say no.

Now, is Grok always awake like this? Probably not. Musk vowed to “take action to make Grok politically neutral.” Time will tell what the results will be.





Source link

Leave a Comment