Unfortunately (for them?) Grok is quite poor quality compared to GPT-4o, Claude, Gemini, and numerous open source models.
It's worth noting that xAI is also focused on photorealistic image generation, which none of the others seem to be focused on. I believe that's because they're going to position themselves as the platform for generating pornography, which we all know will be a lucrative business, but none of the other major players want to touch. It's the only thing that justifies the enormous amount of funding.
There’s also another problem — their most loyal user base already had access to it. And still nobody uses it because of its inferiority. They’re way too late into the game, and to catch up, on top of the infrastructure, you need top notch talent. But talent is limited, and pulling them out from big players won’t be easy because they’re also pouring billions to keep them in the shop. And having even a bit of negative PR when it comes to your employer might be a deciding factor for some potential candidates.
Another X factor (heh) is that it's a lot harder to convince existing highly-paid, usually left-leaning individuals to go work for Elon Musk's AI lab. Of course it's more complicated than that, but the sheer number of workers in our industry who despise his politics and how he uses his money to advance them makes it a tall order for xAI to become a serious AI lab.
Worth nothing that the image generation part is not developed by xAI - they started with using Flux from BFL (Black Forest Labs), and now seem to have made a finetune of it (in partnership with BFL, I assume). So most of the image gen tech that Grok has isn't made by them.
Right, but I expect their extensive post-training they're doing will be something they carry over to their own pre-training initiative they've been gearing up for.
I remember before OpenAI no big company wanted to host a chatbot or provide a service around them after people made the all the bots say few racist lines. All the customer service bots were basic matching and answering questions from fixed database. Even if you even see 1-2 year back, companies were so focused on refusals, early Google's and Anthropic's model refused lot more things than now.
Now it's the same with image generation or voice model. They refuse everything but it takes one player to decide the market is worth capturing and everyone else will follow the money. As long as people can't generate nudes of any given person and there is a toggle(on by default) to block accidental nudes, I would say it's fine.
I don't think most major players will want to get into the pornography game. It's far, far too fraught with ethical concerns to make it worth it. Musk has no such concerns, though. And I believe that also extends to revenge porn, which will be one of the biggest use cases for this tech at scale.
There are different stages though. Just because revenge porn exists doesn't mean all sexual output is unethical. If a company proves that generating animated hentai is say worth $10b in revenue, I am sure others will follow suit in dialling down the restrictions slowly.
And AI just reduces the budget and time, it can't output anything other that what could be created by say Adobe suite.
I think you're looking at it the wrong way. Is there demand for AI-generated images in general? No, it's just that there is a lot of demand for middling, utilitarian artwork. And as it happens, AI can generate it more quickly and cheaply than human artists.
It will work the same for porn. It's just that mainstream models are sanitized not to generate it, so the cost to enter is higher.
I'll believe it when I see it. I think it's a gimmick until I see otherwise. People don't want to expend energy figuring out what they want, and generating porn with ai sounds like a lot of work for shitty results.
Hahahahahaha… yeah no. The demand for porn tailored to a person's fetishes is a cash cow for artists looking to triple their income from
their day job.
Ignoring the snarky analogy, yes, budgets have a ceiling, so it’s helpful to have services at multiple price points, even if the value is not a linear scale of price.
While I think your argument makes sense in the abstract, the person you’re replying to asked if it was cheap, not if it was within budget. Depending on one’s needs, if their budget allows them to afford something but that doesn’t adequately fulfil the need, it can still be considered expensive.
True - both OpenAI and Anthropic have free usage tiers, although they typically offer inferior models. Although IIRC OpenAI does offer some limited uses of 4o for free users, but Anthropic seems to have stopped offering 3.5 Sonnet for free.
It probably needed certain infrastructure scale to make sense before offering a freemium model on an large existing platform like Twitter. Running this stuff is very expensive.
Not entirely related, but what was all the hype about the xAI Colossus? It's claimed to be the most powerful cluster by some measures. I know that a cluster isn't comparable to another because it's all a matter of connectivity and use cases, but does it actually make sense have 100k H100s connected through Ethernet or was it something others didn't do because it is not actually useful?
And I need this in my twitter workflow for....? Generating share images instead of bringing them with my content to post? Searching something instead of the search box? It just seems like no incentive for me to go over there and use it for anything.
I have not used Grok or know of any comparisons to other LLMs - so did the Twitter datasets give them any edge in performance or other quality of benchmark?
It's a question, I suppose. Let's feed it in and see what happens. Grok in "fun mode":
<enormous response full of typical AI speak elided>
Let's whittle that output down a bit with bullshitremover:
Elon loves X. His first big thing was X.com. SpaceX has X for space.
Tesla Model X is part of the S3XY lineup. He even named his kid X.
Now Twitter is X. His AI company is xAI. X is Elon's thing - branding,
exploration, rebellion. Love it or hate it, can't ignore it.
He's had it for a while — he founded another company 20 years ago, x.com, which merged into PayPal. Following that, he wanted to rebrand PayPal into x.com, and was eventually fired for it
Partially. He basically kept ignoring the board and lying to the board -- and that was one such case.
Another example was he decided that all of PayPal needed to be Windows computers so every engineer has to stop what they were doing and convert. Board said stop, he said okay, then continued.
The iPhone case made sense. Apple had Mac OS X, which was the system after Mac OS 9. It had major architectural changes, so much so that every major release was always 10.N, where N was a number from 1 to 15, plus a minor version (10.16 became macOS 11). The iPhone X was released on the tenth anniversary of the iPhone.
In both cases the X stood for the roman numeral 10 and the correct pronunciations for the names were “Mac OS ten” and “iPhone ten”. In the iPhone’s case, they even skipped a version (there was never an iPhone 9).
It's worth noting that xAI is also focused on photorealistic image generation, which none of the others seem to be focused on. I believe that's because they're going to position themselves as the platform for generating pornography, which we all know will be a lucrative business, but none of the other major players want to touch. It's the only thing that justifies the enormous amount of funding.
reply