Latest / Columbus' Morning News / What to know about China's new A.I. model
Transcript
- 0:00Speaker 1: All right, it's seven thirty nine in the morning. Mike Elliott,
- 0:02Speaker 1: glad you're here. Let's get to the Legacy Retirement Group
- 0:06Speaker 1: dot Com phone line and say good morning to Mike Debuski,
- 0:09Speaker 1: ABC News Technology reporter.
- 0:11Speaker 2: Michael.
- 0:11Speaker 1: See, there's a couple of new AI models out there,
- 0:15Speaker 1: and these are from China.
- 0:16Speaker 2: I got people talking. What's the story.
- 0:18Speaker 3: Yeah, So, in the last week, two Chinese companies unveiled
- 0:21Speaker 3: new AI models. The first one was called Moonshot Ai.
- 0:25Speaker 3: They released a model called Kimmy K three on Friday. Seemingly,
- 0:30Speaker 3: wherever you land, no one understands out to name an
- 0:32Speaker 3: AI model properly, but even still, they unveiled this new model,
- 0:36Speaker 3: and they claimed that in their own testing, this model
- 0:39Speaker 3: was better than nearly all American AI models out there
- 0:42Speaker 3: from the likes of OpenAI, Anthropic, Google and everyone else.
- 0:46Speaker 3: And sure, it's all well and good to like claim
- 0:48Speaker 3: that your technology is more intelligent and smarter than the
- 0:51Speaker 3: best of the best, but we actually have like an
- 0:54Speaker 3: objective way to evaluate this through benchmark testing, and according
- 0:58Speaker 3: to Moonshot Ai, going to release the scores of their
- 1:02Speaker 3: testing as soon as next week. And then things got
- 1:05Speaker 3: a little bit more interesting. Over the weekend, Ali Baba,
- 1:08Speaker 3: another Chinese company, unveiled a model called Quen three point eight,
- 1:12Speaker 3: again thought to be this super advanced AI model that
- 1:15Speaker 3: can go toe to toe with the best that America
- 1:17Speaker 3: has to offer, and they say they will soon also
- 1:20Speaker 3: release their model weights, their their benchmark score testing, you know,
- 1:25Speaker 3: in the next few weeks.
- 1:26Speaker 2: Here they are hedging a little bit.
- 1:28Speaker 3: Both companies say their respective models are better than most
- 1:31Speaker 3: of what the American companies have cooked up. They don't
- 1:33Speaker 3: say they're better than the best. For example, both Chinese
- 1:36Speaker 3: firms cite fable five from Anthropic as probably more powerful
- 1:40Speaker 3: than what they have cooked up here. But even still,
- 1:43Speaker 3: these companies would not make such a fuss over this
- 1:45Speaker 3: new technology if their models were not at least competitive,
- 1:49Speaker 3: and that has been enough to worry domestic tech companies
- 1:52Speaker 3: and their supporters on Wall Street. After Moonshot made its
- 1:54Speaker 3: announcement on Friday, the tech heavy NASDAK closed down by
- 1:58Speaker 3: about a percentage point on worries that this race with
- 2:01Speaker 3: China is maybe a lot closer than we were anticipating.
- 2:03Speaker 2: Mike, what makes an AI model better than another?
- 2:07Speaker 1: From an average, normal everyday user's perspective? What differences would
- 2:12Speaker 1: I see between an Okai agent versus a really good one.
- 2:17Speaker 2: Yeah, a great question.
- 2:18Speaker 3: So to zoom out here, what we're talking about are
- 2:21Speaker 3: not products, they're talking about engines. Really, So if we
- 2:24Speaker 3: look at like open Ai as an example, their flagship
- 2:27Speaker 3: product is chat GPT, it's a chatbot, but it's ultimately
- 2:30Speaker 3: just like a computer program, right, It's an app, it's
- 2:32Speaker 3: a website that you can go to.
- 2:34Speaker 2: The chat GPT sort of idea.
- 2:37Speaker 3: Runs on an engine known as GPT five, And that's
- 2:40Speaker 3: really open AI's core technology. Kimmi K three from Moonshot,
- 2:44Speaker 3: Quen three point eight from Ali Baba. Those are engines,
- 2:47Speaker 3: so they can be built into a whole variety of
- 2:50Speaker 3: different things. And really, how we evaluate the difference in
- 2:53Speaker 3: how quote unquote good these things are is their range
- 2:56Speaker 3: of abilities.
- 2:57Speaker 2: Right.
- 2:57Speaker 3: There are models that are really good at, you know,
- 3:00Speaker 3: math problems, some are really good at like logic problems.
- 3:03Speaker 3: Some are really good at, you know, writing in a
- 3:05Speaker 3: human sort of way. Some are good at sort of
- 3:07Speaker 3: understanding large scientific concepts and boiling them down into something
- 3:11Speaker 3: that we can understand, or coding or you know this,
- 3:14Speaker 3: that and the other thing. There's a lot of different
- 3:16Speaker 3: like sectors that you can you know, evaluate a model
- 3:19Speaker 3: on and the way we talk about like powerful models
- 3:22Speaker 3: are ones that can kind of do it all right.
- 3:24Speaker 3: You think about like an SAT test, for example, it's
- 3:27Speaker 3: all well and good to be good at the literature section,
- 3:29Speaker 3: you know, but you don't want it to come at
- 3:31Speaker 3: the expense of like the math section. That's kind of
- 3:33Speaker 3: how we test these and that's basically how benchmark tests
- 3:36Speaker 3: go about, you know, kind of evaluating these problems. It's
- 3:39Speaker 3: like a SAT test that's been turned up over and
- 3:42Speaker 3: when you look at scores of that, you know, that's
- 3:44Speaker 3: a good way to kind of put everybody on the
- 3:46Speaker 3: same page about their capabilities. It's worth mentioning there are
- 3:49Speaker 3: some complaints about how benchmark testing is done. For example,
- 3:52Speaker 3: a lot of AI companies do have a hand in
- 3:54Speaker 3: how benchmark tests are written, but that's kind of a
- 3:57Speaker 3: you know, a different story and something that we're kind
- 3:59Speaker 3: of working on figuring out understanding. Even still, it's kind
- 4:03Speaker 3: of the best thing we have right now to evaluate
- 4:05Speaker 3: the relative power of one model versus another.
- 4:07Speaker 1: Are there any security or safety concerns from an Chinese
- 4:12Speaker 1: own AI platform?
- 4:14Speaker 3: Yeah, certainly, And I mean this is something that, you know,
- 4:16Speaker 3: the American tech companies have raised as the potential here
- 4:20Speaker 3: for this to go wrong, right, like, not necessarily that
- 4:23Speaker 3: this model would like steal people's information or anything like that,
- 4:26Speaker 3: but that it could you know, enter into a less
- 4:29Speaker 3: regulated environment and be used to exact great harm globally.
- 4:33Speaker 3: Interestingly enough, how both Ali Baba and Moonshot have decided
- 4:36Speaker 3: to go about this, in addition to a number of
- 4:37Speaker 3: other Chinese AI firms, is to make these models open source,
- 4:41Speaker 3: meaning that you know, developers, you or I basically anyone
- 4:44Speaker 3: can go in and like tinker around with the source
- 4:47Speaker 3: code of this and see how it's built, see how
- 4:50Speaker 3: it kind of thinks about different problems and what have you.
- 4:54Speaker 3: That is a competitive difference from how the Americans have
- 4:57Speaker 3: thought about doing this open AI anthropic. Their models are
- 5:00Speaker 3: proprietary to themselves, right, they want to keep that trade
- 5:03Speaker 3: secret to themselves. You know, maybe this sort of open
- 5:07Speaker 3: source idea that is popular among Chinese tech firms is
- 5:10Speaker 3: a way to sort of assuade those concerns and get
- 5:12Speaker 3: these models out in front of more people, you know,
- 5:15Speaker 3: by just kind of giving over a little bit more
- 5:17Speaker 3: of you know, what they're built on top of.
- 5:21Speaker 2: Will it work?
- 5:21Speaker 3: I think that kind of remains an open question. But yeah,
- 5:24Speaker 3: I think, you know, with with all sort of Chinese
- 5:27Speaker 3: you know, tech products like that, you might want to
- 5:28Speaker 3: take it with a little bit of an extra grain
- 5:30Speaker 3: of salt, as we have done with you know, companies
- 5:32Speaker 3: like TikTok and you know Chinese vehicles and what have you.
- 5:36Speaker 1: Mike, you know, as a technology reporter for ABC, at
- 5:39Speaker 1: your job literally to stay on top of all of this,
- 5:41Speaker 1: and I sure I assume you use as much of
- 5:43Speaker 1: this as possible for either for personal use or you know,
- 5:47Speaker 1: for business if you're working on a story. When it
- 5:49Speaker 1: comes to you know, privacy, data, privacy, or you know,
- 5:52Speaker 1: any breach type of issues, is there anything that you
- 5:55Speaker 1: personally won't put into an AI chat bot, a chat
- 5:59Speaker 1: g PT or a quad or one of these new
- 6:02Speaker 1: ones that you're like, hey, you know, I got to
- 6:03Speaker 1: draw a line there. I won't put a picture of
- 6:05Speaker 1: myself or any kind of personal effort.
- 6:07Speaker 2: Where do you draw a line when it comes to
- 6:09Speaker 2: safety with AI? I drive the line pretty early, I say.
- 6:13Speaker 2: For one, you know, I am an employee of ABC News,
- 6:16Speaker 2: which is owned by Disney.
- 6:17Speaker 3: You know, we as a journalistic outlet are not allowed
- 6:20Speaker 3: to have our sort of work interact with artificial intelligence anyway.
- 6:24Speaker 3: And you know, obviously that you know, from the company's perspective,
- 6:28Speaker 3: makes a fair amount of sense, right, Like there's some
- 6:29Speaker 3: sense of information that flies around our emails, like we
- 6:31Speaker 3: just don't want that, you know, potentially to show up
- 6:34Speaker 3: in an AI model or you know, as a piece
- 6:36Speaker 3: of training data or what have you.
- 6:38Speaker 2: In my personal life.
- 6:39Speaker 3: I mean, like, I really do struggle to find a
- 6:42Speaker 3: consistent application of this technology outside of just playing around
- 6:45Speaker 3: with it, to explain it to you guys, and to
- 6:47Speaker 3: explain it to coworkers and to put it into stories.
- 6:50Speaker 3: I think that, again, is something that we should always
- 6:53Speaker 3: think about in the back of our minds, even as
- 6:55Speaker 3: we talk about the AI race with China and the
- 6:57Speaker 3: regulatory environment and how powerful these things are getting. No
- 7:00Speaker 3: one's really paying for this, And the reason that no
- 7:02Speaker 3: one's really paying for this is because they haven't demonstrated
- 7:05Speaker 3: a consistent use case for this technology in a way
- 7:09Speaker 3: that people will shell out, you know, twenty bucks a
- 7:11Speaker 3: month for And as costs continue to rise, as China
- 7:15Speaker 3: continues to put pressure on American tech companies, I think
- 7:17Speaker 3: that's going to become an increasingly acute problem.
- 7:20Speaker 2: A lot of people use.
- 7:21Speaker 3: AI, right According to OpenAI, some nine hundred million people,
- 7:24Speaker 3: nearly a billion people a month use Chat GBT, but
- 7:27Speaker 3: a tiny fraction of them pay for it, and again
- 7:30Speaker 3: that kind of hangs over this whole discussion in a
- 7:32Speaker 3: huge way. You know, it's all well and good to
- 7:35Speaker 3: talk about the power of this stuff, but if no
- 7:37Speaker 3: one's paying for it, then there's no business here, and
- 7:40Speaker 3: that's again going to become an increasingly acute concern.