OpenAI, Google ink deals to augment AI efforts with news – it was Time for better sources OpenAI and Google on Thursday independently announced fresh collaborations with major publishers as they work to expand paid access to information used by their AI products and services. OpenAI is now working with Time Magazine under a multi-year agreement, the terms of which weren't disclosed, that grants the machine-learning …
Hi, Tobias,
Can we expect and look forward to lots of enlightening future news should OpenAI and Google come a'knocking on El Reg/Situation Publishing doors/stores for mature content?
El Reg archives into training datasets would certainly be at least a novel venture and also, whenever quite possible, probably extremely problematical too for an unchosen few in some very specific areas entertaining and exploring and exploiting remotely leading virtual interests.
Although El Reg could surely always boot up its very own AI and keep its growing content for exclusive future leading use, .......and in so doing make any number of killings on markets susceptible and vulnerable to savvy high tech disciples and renegade rogue 0-day traders alike.
I enjoy your coverage of AI and it isnt too far out of date and much improved.
And I was so enjoying your article until... you shoehorned RAG in! Are you an old database man that is trying desperately to keep traditional databases in the previously elevated position? Kind of reminds me of the news hacks when the internet came along and they tried to get Quark to do HTML! Oh how I laughed.
RAG is alright but what is the point of having another massive database of structured data again! So, with your beloved RAG, we will get to the situation where more and more content providers are added and then we end up back where we started except this time the answers are in NLP. RAG is much more processor intesive too.
OpenAI have a much more complicated system then just RAG. RAG isnt used there, they have their own 4 part modular data system of which a highly advanced version of RAG (CRS) WITHOUT the structured databases is part of.
You mention that the downside of RAG is the purity of the data, but that's an old database man argument that is applicable to any data. That isn't noting the negative impacts of running what will be DB #3. #1 the internet as a structured (of sorts) database that goes into #2 AI model and becomes abstract data that is then superseded by another #3 structured database containing the same data - albeit limited to only the major 'trusted' outlets of propaganda. And not only that but that #3 step means lag, big lag. Im sure this deal isn't API calls. It's their archive. It will be hosted by OpenAI. It's too slow otherwise and energy costs will rocket as RAG is a very CPU intensive task.
And that's just the archive. What about the stream of content for the next 5 years! How can you integrate real-time into a snapshot model LLM. You cant. You can fudge it like RAG but it will bite u in the arse for the reasons mentioned above.
RAG is useful but it is a best a clumsy stop gap that is a fake real-time model. The only way we can get LLMs or similar to work properly is for them not to be snapshot databases that we have now, but real-time models.
We have just finished a model specially designed to estimate how long it will take for desktop/mobile devices to be able to run real-time neural networks based on the current trajectory of tech and including random speed bursts to account for little discoveries on the way, like RAG. We were convinced it was wrong for about a month but could not disprove or disagree with its projections no matter how much we tried.
Server-based real-time 10-15 years. Commercial rollout 15-20. Local versions using same model: 2 decades. 2 decades till we get the thing to even the level of the most stupid of humans. Even then, we will probably discover that there is a whole load of new problems that will stop as ever getting there.
RAG is a waste of time. Spend the time and research on real-time fully blown AI models - not crappy little Apple basic OS actions AI. RAG is XLST - a hopeless deadend that offers initially interesting and powerful results until you use it in production and everybody hates it.
I hop you'll forgive me for saying so, but it sounds like you are in the business and have your own agenda here. Also, you claim that there is currently no proper real-time solution and people should not try, but the nature of the business is that those who wait until they have a proper solution will get eaten by those who'll offer a half-baked solution right now.
No that's fine. Very polite counter and your criticisms are valid.
I dont have an agenda anymore. I did 2 months back but not now. My only agenda is too extract as much money as possible from greedy people before somebody else works out what we have.
We might be wrong, but as far as I can tell, it is the most comprehensive model on a subject we have ever done. We tried for a whole month to keep expanding the related datasets to improve the result. I was heartbreaking TBH.
How can we be sure of the accuracy? We can't. Why? Cause real-time models are the merest hint of in the theorists mind at the moment. There are 3 options but the best is a second level of abstraction on the abstract data from the model. And this is where the darkness creeps in. The abstraction of the abstract theories is to use of the spaces around the abstract data to enable even greater compression and less runtime demands.
"no proper real-time solution and people should not try"
I was mistaken if i said that, but I dont think I said people should not try. The opposite in fact. Stopping dicking around with RAG and focus on real-time neural networks.Like I said above, there is only high-level theory at the moment.
"but the nature of the business is that those who wait until they have a proper solution will get eaten by those who'll offer a half-baked solution right now.£"
very true and Ive done some half-baked solutions on massive projects before which turned out good enough. The author is banging on about the wonders of RAG and I was pointing out that OpenAI don't use structured RAG databases or even RAG. Their system is much more developed and that RAG is a step back.
And that we cant be the only agency to have built a model like this. Others must know this too, but not very many. I just dont think anybody has told the investors the truth about AI's future development curve.
It's the disappointment. The sadness of the reality of the next 10-20 years. We are running out of the steam Denis gave us in 2016. The bubble burst is coming soonish. Once people realise that Siri 2.0 and ChatGPT are pretty much as good as it is going to get for a decade, and we are stuck with the IE4 of AI.
The only hope I can see is maybe quantum computing but that's miles off.
I dont have an agenda anymore. I did 2 months back but not now. My only agenda is too extract as much money as possible from greedy people before somebody else works out what we have. .... Anonymous Coward
Sounds like a pretty good, real agenda to me, AC, with more than just you realising it is one very attractive and rewarding way to go to rid worlds of their wannabe fake JanI [Janus’s] and mad attack dogs and warware mongers.
Here's some live news of A.N. Others similarly enamoured of NeuroLinguistic Programming Projects to extract as much money as possible from greedy people before somebody else works out what we have, and they can do too with AI help assisting them and future developments ...... https://forums.theregister.com/forum/all/2024/06/28/microsoft_ceo_ai/#c_4887160
> Providing proper attribution to original sources
Sources? Plural?
> featuring a citation and link back to the original source on Time.com.
So Time is (presumably) being paid for this access (so that's alright then) *and* they effectively get advertising, which basically credits it as the only source of information In The World. Meanwhile, all the sources that were scraped for years beforehand are still ignored.
A cunning ploy to get other sources to give in[1] and allow their materials to be scraped again, just to even out the citations playing field[3]? Maybe even get a few to pay for the "privilege".
In the meantime: hmm, wonder if an LLM will spot that there is a pattern to the way that approved citations are listed and then starts applying that whenever it is prompted to give citations: "Put glue on your Pizza (Time, April 37th 1876)".
[1] "you know your material was scraped[2] so it'll be spat back out in one form or another; wouldn't you prefer it be shown as an accurate quote from our database, with citations, or just munged in with everything else?"
[2] even though we won't admit it publicly and you can't prove it strongly enough to win
[3] and to get a piece of paper saying they let us in, so nyaah to all your sueballs ("No, we never scraped a copy of Harry Potter, we just assembled a complete copy of the plot and all the characters by piecing together the details from the World-Class reporting of the books and movies that Time is so well known for")
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
