I know it’s not the main issue but this Grok stuff (some responses now deliberately skewed) should make it unviable as a commercial proposition. The entire point of LLMs is that they faithfully reproduce on a probabilistic basis. Take that away and you’re dealing with a fancy Daily Mail.
Fuck all AI. I dont want any part of any of it. It is a cancer that needs to be eradicated. I'm still on that cesspool occasionally and I blocked it and I block every single person who asks it a question.
I’ve wondered if, for this reason, he’s only Nazify the one hooked up on twitter and leave the one you get through the API alone.
I'm grateful to Elon for proving that it matters who runs AI
I am shocked, shocked, at the suggestion that supposedly objective 'facts' could possibly be manipulated to suit any individual's agenda.
Also for his input on the "what if billionnaires are a-ok, ackshually" debate.
Even decent people behind AI wouldn’t be enough. For example, the Russians have been seeding the web with propaganda just in order to train the LLMs that crawl their websites.
Well. Quite.
Granted he’s a youtuber, but he spent a bit of time coming up with likely / interesting user tests. Grok does better than I expected. Perplexity coming last i did expect tho. youtu.be/cMuif_hJGPI?...
I thought that was the intent already? Ai says what "we" (the owning class) wants it to parrot back as truth, and the peons shall consume it
Though the chatbots all have a long prompt not visible to the user, giving guidance on how they should answer – it's the only way of skewing the answer, whether away from bias inherent in the training, or towards bias attractive to the owner.
I dunno man, you could say it’s a good demonstration of capability: “So if we, instead of automatically appending to every prompt ‘and respond in the style and manner of a white supremacist’ you append ‘and respond in the Barclay’s banking group house style, following corporate guidelines’. We get”
Grok is clearly demonstrating an ability to produce outputs with a preselected response style using an artificially restricted set of house approved tokens.
It will be entertaining to watch Grok morph into a more 'realistic' version of Musk as he tweaks its algorithms to agree with him.
They were always trained on selected data giving weights and so on I think. This is just taking it to the extreme
Well. Yes.
"Faithfully reproduce" isn't possible using LLMs' probabilistic approaches - that's why the darn things present bullshit so often (and as fact). It's all unviable commercially - look at the money they lose. And it's not why they're being rammed down society's throat :(
(Some v helpful clarifications in the responses)
On a related note, there’s a nice irony in the people who see their work being most helped by AI being the people whose jobs are most likely to be replaced by AI.
Putting it a little differently, if ‘you’ think what ‘you’ do would be enormously helped by AI (the summarising 100 books thing is a classic) then what’s the point of you (in a differentiated sense)?
I work in a part of publishing which can – in principle – be easily commodified but in practice depends on thoughtful creativity for its best output. So the jobs are vulnerable to AI, but so is the quality, for now and for a while. Midterm survival depends on commodifying the right bits of the work.
AI will (already does) produce plausibly adequate slop; but which is in fact not adequate. If that slop gets slightly better, though, then it will replace higher quality publications if those are still orders of magnitude more expensive.
Yep slop. That is the problem people seem willing to use AI to produce something that ultimately will have no value. A lack of understanding of what is and isn’t valuable and the end product is often only part of that.
I say this with support and love. I just find the industry to be moribund, non-sensical in terms of business practices (not just about me, btw), and strangely antiquated in its striving for new paradigms. But this does not reflect on you. Cheers, mate.
Sucks that yall are being flooded by aislop. Made it exceedingly difficult to get my professionally well reviewed book even looked at. Doesnt help that there are all these unmentioned standards that are never mentioned in the guidelines and are antithetical to what was expectex of previous eras.
Very reminiscent of the old Woody Allen tale about going on a speed reading course which taught him how to read War & Peace in 20 minutes. When asked what the book is about he answers succinctly - Russia!
I'm a teacher. My job has become telling truth from fiction, as students lean on AI for their answers. Anything from "A clockwork Orange is about mechanical fruit" to "chuck steak is the most tender steak, inedible in stews, it is called chuck because you throw it" has come across my desk this year.
So you’re saying A Clockwork Orange isn’t about mechanical fruit? 😂 Someone could have just made that random guess on their own - no AI required.
Trust me when I say I could tell he didn't write it, even before he told me where he got the information. It had the written equivalent of a double thumbed hand with two wrists.
Is this disheartening, as a teacher, or do you feel resigned?
Mostly, I guess, I feel needed. Without me, they are going to cook a meal or follow a road rule some AI told them, they will calculate a distance with an AI or whatever, and have something truly terrible happen to them. My job is still needed, to help people differentiate between hallucinations.
That’s a positive way to look at it. I wonder if, in a few years, everyone will recognize how worthless AI is and abandon it. 🤞🏻
I don't think so, but they might start laying down some laws, hopefully. Would be nice if Google wasn't allowed to push AI before search results, for example.
I mean it should be absolutely obvious to everyone that using llms for information finding is just as open to enshittification and sponsored answers as Google websearch, Amazon search results, Facebook newsfeed, etc etc
Grok is horribly augmented in patchwork fashion with whatever BS Elon comes up with.
Given that every country seems to be voting for further and further right-wing parties, maybe the AI outputs should be matching this trend to suit their audience rather than giving them fact-based answers? /sarcasm
I wonder if such skewing is not more common with the other LLMs. If GROKs competitors don't pounce, you'll know they are doing it too.
Plato’s cave needs another chamber….
Digging
I would never trust an LLM unless I trusted the creator. Trust Elon Musk? Lol no.
No serious person or institution will use Grok for anything meaningful. Anyone that does use Grok will be greeted with derision amd suspicion. Grok is now the official AI of racist trash.
It is almost sure none of the big LLMs are currently commercially viable. Even the most successful one (OpenAI) is losing significant amount of money, have minimal moat to protect themselves from new entrants and open source / open models are only something like six months behind.
This is a bit overly simplistic. OpenAI and other labs make a big profit margin on actual inference provision - their losses are due to their ginormous R&D spend to improve the models, which may not be necessary forever
Fair point about R&D costs being the biggest part of OpenAI bills. But if they decided to ditch their R&D, it would take about 6 months for open models to catch up, at which point the cost of inference would be equal to the cost of hardware.
Agree R&D costs will remain high to keep them ahead, but since inference itself seems to be profitable (and becomes more so as models become more useful + inference stack is optimised) it seems plausible that inference profits can ultimately exceed R&D costs
My guess is that all LLMs are currently trying to capture a market share with a view that they will commercialise it later on. From that perspective EM may be trying to capture a market share of people who do appreciate 'fancy Daily Mail'.
Much like Google did with search - capture the market, then find and settle on a business model.
This is what happens when you kill the college research/military research kickdown pipeline. You get a lot of money spent that is always inadequete because true innovation requires the failures experimentation brings, and that doesnt work for a bottom line.
More it requires continued efforts past where a commercial enterprise will demand "now where is the product we can sell?" And companies fail at making standards because what it takes is working to make the pie bigger even if you can't own it all. They'd rather dominate than succeed.
It’s a feature not a bug. They hallucinate intentionally- as they work to support whatever conjecture is made and then make stuff up to “support “ the position It’s messed up for sure
There's a dedicated customer base for that, though.
Well but what if nazibot3000 IS the commercial preposition?
AI's are BS (bias substantiation) machines and Grok is programmed to spread Musk's biases. Anyone using it in particular is fine with being indoctrinated into the world of white supremacy and Nazi beliefs. Ask people which AI they use, and shun anyone that says Grok.
Or any answer really
The Daily Mail has a thicker veneer of respectability than the stuff Grok is churning out at this point.
Are you using LLMs for fiction or nonfiction? 🤖🤔
Polymarket has odds of XAi having the best model at 14% still
FWIW the core model seems unchanged, but they augmented systems around it for twitter
Ah very interesting
Although that in of itself raises all sorts of questions. I guess though it’s not surprising that any corporate version of this sort of thing is going to be deliberately skewed.
i think its worth saying even for OpenAi i dont think serving these models is a big profit centre for them. Commercially the per token prices have just been crushed
Their per token prices are still multiple times their inference costs, though
i just mean in general these ventures will make their money on other products that arent the API
Not convinced tbh. Smaller, more targeted companies (Cursor being the most obvious) are beating frontier labs at building products targeting specific use cases, so it seems that selling general-purpose API inference will be critical for the frontier labs
Think the primary lesson from grok and deepseek is that there is no moat and there won’t actually be any big money in general purpose frontier models. Just as there was no big money in being an ISP.
For me Grok and DeepSeek show that though many labs can produce a model at 80% of frontier quality, we have still never seen anyone (arguably Anthropic briefly) catch OpenAI / DeepMind, and power/commercial users are willing to pay a premium to have the best model over a model just nearly as good
Maybe Grok isn't intended to be commercially viable on its own; like GB News, it's a loss leader, whose value is as a propaganda outlet to make the political and social landscape favourable to oligarchs and their schemes which _are_ designed to make profits.
Elon Musk the Commercial Genius strikes again!
Your dealing with FASCISM
I am convinced Grok's primary source is Twitter itself, and the model is getting updated close to real time on Twitter feed, which is why it often comes off so awful. It's just regurgitating the Twitter mess.
It's not possible for it to be trained in real time. It's quite possible they're stuffing masses of information into the context before you type a prompt.
Yeah I know I am super over simplifying. A topical RAG curated in realtime for examine could be a thing, or LORAs made daily.
It is being trained on Twitter data though...
Of course, but surely that's what all these LLMs are for? A way to present their alternative facts in a way that people believe?
Your mistake here is assuming that "Fancy Daily Mail" is not commercially viable.
In fact, I bet people would pay extra to keep an LLM from ever spitting out a fact that makes them uncomfortable.