OpenAI, Google ink deals to augment AI efforts with news – it was Time for better sources
- Reference: 1719570254
- News link: https://www.theregister.co.uk/2024/06/28/openai_google_ai/
- Source link:
OpenAI is now working with Time Magazine under a multi-year [1]agreement , the terms of which weren't disclosed, that grants the machine-learning mega-lab access to the publisher’s assets – from the present way back to 1923.
"We're partnering with Time to make it easier for people to access news content through our AI tools, and to support reputable journalism by providing proper attribution to original sources," OpenAI COO Brad Lightcap declared in a canned statement.
[2]
It isn't entirely clear how Time's content will be integrated into OpenAI’s products. The AI upstart simply noted the catalog will "enhance its products and display in response to user inquiries – featuring a citation and link back to the original source on Time.com."
[3]
[4]
This suggests the use of retrieval augmented generation (RAG) to parse Time articles, in order to generate more accurate answers to ChatGPT users' queries.
We covered RAG in depth in a recent [5]hands-on tutorial . In a nutshell, the technique ingests content, converts it into an embedding model and feeds it into a vector database that a large language model, such as GPT-4, can use to generate answers. In a sense RAG is a bit like sending your LLM to the library to research a topic – or in the case of this deal, a newsstand.
[6]
There's nothing in OpenAI's announcement to suggest it won't integrate Time's archives into its training datasets. We've asked OpenAI and Time for comment.
There's also no word on how much cash is changing hands, but the publisher will receive access to OpenAI's technology to develop "new products for its audiences." Additionally, the publication will get the opportunity to share its input on how to "refine and enhance the delivery of journalism in ChatGPT and other OpenAI products."
Exactly how much sway Time journalists will actually have over OpenAI – or the use of Time content – is unclear.
[7]
The announcement comes weeks after OpenAI announced a partnership with Reddit and News Corp to feed its models with content.
Google does deals, too
Google, which [8]announced a similar agreement with Reddit back in February, revealed on Thursday that it's [9]expanding its content partnerships to include several more content and data providers.
The search and ad giant's collaborators include credit rating outfit Moody's, investment research firm MSCI, Thomson Reuters, and Zoominfo. However, unlike OpenAI's partnership with Time, Google specifically mentions the use of these sites as fodder for its RAG databases.
Google will offer its customers an option to "ground" their AI apps in facts sourced from the sites it’s signed. The hope is that more outside sources will reduce the likelihood of hallucinations – aka errors.
[10]Reddit hopes robots.txt tweak will do the trick in scaring off AI training data scrapers
[11]Record labels gang up to sue AI music generator duo into utter oblivion
[12]Dominant AI players Nvidia, Microsoft, OpenAI face US antitrust inquiries
[13]Read AI about it... OpenAI does deal with News Corp
RAG should improve the quality of AI-generated results, but it isn't a silver bullet. Even if a model doesn't make something up, inaccurate source data remains a problem. As we saw last month with Google's own AI search function, the quality of the answers relies heavily on the quality of the source. This led to the AI search [14]presenting troll posts – including one suggesting adding "non-toxic" glue to pizza sauce to help the cheese stick – as fact.
However, as customer expectations of machine learning tools grow, so too has the demand for high quality datasets and – by extension – the need to protect curated sources of content from being hoovered up into the AI vacuum cleaner.
Earlier this week, Reddit said it would start [15]making changes to its Robots.txt file to dissuade AI startups from scraping its site without paying for it – something we'll note OpenAI and Google have already agreed to do.
These deals are also rather lucrative, with Google reportedly [16]paying Reddit $60 million a year for the right to peruse its posts. ®
PS: That's nice for Time. But the Center for Investigative Reporting, a nonprofit that operates Mother Jones, [17]is now suing OpenAI and its backer Microsoft for ingesting its copyrighted content allegedly without permission nor compensation.
Get our [18]Tech Resources
[1] https://openai.com/index/strategic-content-partnership-with-time/
[2] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2Zn7eIoglpxlobQGzgPRdfAAAANY&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[3] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44Zn7eIoglpxlobQGzgPRdfAAAANY&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[4] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33Zn7eIoglpxlobQGzgPRdfAAAANY&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[5] https://www.theregister.com/2024/06/15/ai_rag_guide/?td=rt-9cp
[6] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44Zn7eIoglpxlobQGzgPRdfAAAANY&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[7] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33Zn7eIoglpxlobQGzgPRdfAAAANY&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[8] https://www.theregister.com/2024/02/22/reddit_google_license_ipo_altman/
[9] https://cloud.google.com/blog/products/ai-machine-learning/rag-and-grounding-on-vertex-ai
[10] https://www.theregister.com/2024/06/27/reddit_ai_info/
[11] https://www.theregister.com/2024/06/24/udio_suno_riaa/
[12] https://www.theregister.com/2024/06/06/nvidia_microsoft_and_openai_face/
[13] https://www.theregister.com/2024/05/23/openai_news_corp/
[14] https://www.theregister.com/2024/05/31/google_ai_search_update/
[15] https://www.theregister.com/2024/06/27/reddit_ai_info/
[16] https://www.theregister.com/2024/02/22/reddit_google_license_ipo_altman/
[17] https://revealnews.org/press/cir-sues-openai/
[18] https://whitepapers.theregister.com/
good stuff
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.
Re: good stuff
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.
Re: good stuff
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.
hallucinations
Do they really expect factual information from News Corp?
A Logical Next Step of Raw Rare Opportunity
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.