Britannica is now an AI company

[thewayne]

This is interesting.

The company who, up until 2012, published the book-form of the Encyclopedia Britannica, is now turning that huge trove of facts into an LLM engine with the goal of selling it as a service to the education market.

While this might seem as a bit of a snoozer, there's one very interesting aspect to this: AI hallucinations.

Most LLM models have hallucination problems, seemingly stemming from their snarfing up their training data from hoovering up the internet with all of its crappy and contradictory information. This is where Britannica shines: they paid a literal fortune over two centuries collecting vetted materials from recognized scholars using quality editors to compile it into a trusted source. Thus, the quality of their training model will be very, very high.

The question will be if their code that ingests this training model will still hallucinate. And we'll only see that with testing when it goes public and really gets pummeled. But I do like the idea: starting with a very high quality training set, I think it shows promise.

Though we still have the problem of AI systems consuming stupid godawful amounts of energy.

Britannica's encyclopedia is still available online, just not in a print edition.

https://gizmodo.com/encyclopedia-britannica-is-now-an-ai-company-2000542600

Comments

[kathmandu]

I predict it will hallucinate. Generative LLMs aren't search engines, they aren't databases even if they "train" on databases, and they have no reality-checking.

[garote]

Yes. It will absolutely hallucinate. As the linguist professor in my life puts it: “LLMS are doing language with math.”

The whole notion of hallucination is foreign to math. Ergo, the notion of sanity.

[thewayne]

My programming background is most strong in relational database which is working in sets.  I know effectively zero about LLMs and how they train and make correlations, though it's something that I'd like to learn about.  I expect it will, but I think the rate of hallucinations will be much lower and may be more preventable.  But I don't know.

[disneydream06]

I have a copy of the print edition in the basement. Not sure what year it's from, they were dad's.
Also have a copy of The Year Book encyclopedia too. I think from the 70s.
Hugs, Jon

[thewayne]

We gave away our encyclopedia set to a retirement home, it was a World Book from the early '70s.

[disneydream06]

That was a good thought.

[moonhare]

Our library refused to take encyclopedias for book sale donations; naturally I welcomed these along with other ‘unwanted’ materials :o) We may or may not still have a set here that I brought home for my kids.

One of the first software programs I bought for the 386 was a Compton’s encyclopedia. This was a nod to the 1955 set I grew up with at home.

[silveradept]

As soon as the chatbot is given any freedom other than regurgitation of the exact material in the data set, the whole thing springs apart at the seams and seals. Any other entity training on that data set has a higher probability of returning correct information, but that only weights the probability, not makes things a certainty. LLMs will never be sufficiently authoritative and accurate to be used for anything that requires a modicum of either.

[thewayne]

The sad thing is that unis are pretty much requiring LLMs to be incorporated into curriculum.  Going to be interesting to see what the result is if that industry turns out to be a bubble that bursts.  I don't think enough critical thinking is being taught as a pre-req to recognize whether or not the AI is returning good information, and that critical thinking is a must before AI is relied upon broadly.  In certain fields it's fine, such as the AI that is trained upon antibiotics and bacterial pathogens and can create new antibiotics.  But general application AI is a whole other matter.

[garote]

Frankly I feel like the whole usage of the term “AI” for this business is a marketing coup. It’s more like “fuzzy logic” from the 1960’s with ten orders of magnitude more processing power and storage behind it. But “fuzzy logic” doesn’t give people starry-eyed visions of grinning robots holding paintbrushes and serving trays.

What even is “artificial” intelligence when it takes “natural” (human) intelligence to design, feed, and leverage it? It’s a deceitful phrase, custom made to generate hype.

[thewayne]

Gotta have a great buzz phrase to pry money out of investor wallets!

[benicek]

Might produce something half-decent. Let's see.