At one stage, the hype around the potential of generative AI and large-scale language models (LLM), led by OpenAI's ChatGPT, seemed virtually insurmountable. It was certainly inevitable. More than $1 in every $4 invested in U.S. startups this year has gone to AI-related companies, while OpenAI announced at a recent developer conference that ChatGPT is one of the fastest-growing services in history. It was revealed that.
However, something continues to go wrong. In fact, something weird keeps being added.
One of the biggest problems with LLMs is the ability to hallucinate. In other words, it makes things up. Numbers vary, but commonly cited percentages are between 15% and 20%. One Google system saw an increase of 27%. This wouldn't be so bad if it didn't come across so aggressively when doing so. Jon McLoone, director of technical communications and strategy at Wolfram Research, likens this to “the loud know-it-all you meet at the pub.” “He'll say anything that makes him seem smart,” McCrone says. AI news. “It doesn't have to be right.”
But the truth is that such illusions are inevitable when dealing with LLM. As McCrone explains, it's all a matter of purpose. “One of the things that people forget with this 'thinking machine' idea,” he says, “is that all of these tools are designed with a purpose in mind, and the machine performs based on that purpose.” I think so,” McCrone said. “And the purpose was not to know the facts.
“The driving purpose behind its creation was to be fluid. Say what a human would say. That's plausible,” McCrone added. “It's very plausible to say the right answer, to tell the truth, but that's not a requirement of plausibility.
“So you can do something fun like, 'Explain to me why zebras like to eat cacti,' and it does a plausible job,” McCrone says. “It's saying things that might sound right, but of course it's all nonsense, because it's just being asked to sound plausible.”
What is needed, therefore, is a kind of mediator who can inject a little objectivity into the proceedings. And this is where Wolfram comes into play. In March, the company released the ChatGPT plugin. The goal is to “make ChatGPT smarter.”Powerful calculations, access to accurate calculations[s], Curated Knowledge, Real-Time Data, Visualization”. In addition to being a general extension to ChatGPT, the Wolfram plugin can also synthesize code.
“This teaches LLM to recognize the kinds of things that Wolfram|Alpha might know, our knowledge engine,” McLoone explains. “Our approach to it is completely different. We don't scrape the web. We have human curators who give meaning and structure to the data, and we do calculations based on that to synthesize new knowledge. , users can ask questions about the data. There are thousands of datasets built into it.”
Wolfram has always been on the side of computing technology, and McLoone describes himself as a “lifelong computing guy,” having been with the company for nearly 32 of its 36-year history. Therefore, when it comes to AI, Wolfram is on the symbolic side, better suited for logical reasoning use cases, rather than statistical AI, better suited for pattern recognition and object classification.
Although the two systems seem diametrically opposed, they have more in common than you might think. “From what I see, [approaches to AI] All have something in common. It’s about using computational machinery to automate knowledge,” McCrone says. “What has changed during this time is the concept of at what level you automate knowledge.
“In the good old world of AI computing, humans come up with rules for behavior and machines automate the execution of those rules,” McCrone adds. “So computers will expand the brain's ability to do these things in the same way that a stick expands a caveman's reach, but we still haven't solved the problem in advance.
“With generative AI, we're no longer saying, 'Let's focus on the problem and discover the rules of the problem.' We're now saying, 'Let's find the rules of the world.' That way you have a model that you can try and apply to a variety of problems rather than a specific problem.
“So as automation moves higher up the intellectual spectrum, things are becoming more commonplace, but at the end of the day, it's all just executing rules,” McCrone says.
Additionally, companies on both sides share the same goals, as different approaches to AI share common goals. When OpenAI was building his plugin architecture, Wolfram was asked to be one of his first providers. “Once the LLM revolution started, we started doing a ton of analysis of what LLMs were actually capable of,” McLoone explains. “Then we realized what the pros and cons were, and it was at that point that OpenAI started working on the plugin architecture.
“They approached us early on because they were able to think about this a little bit longer than we did. They've been anticipating it for two years,” McCrone said. added. “They already knew exactly what the problem was.”
McLoone will demonstrate the plugin with an example and speak at the AI & Big Data Expo Global event in London from November 30th to December 1st. Still, he is keen to emphasize that there are more diverse use cases that can benefit from the combination of ChatGPT's unstructured language mastery and Wolfram's computational mathematics mastery.
One such example is performing data science on unstructured GP medical records. This ranges from special transcription corrections on the LLM side (replacing “peacemaker” with “pacemaker” as one example) to using old-fashioned calculations to look for correlations in the data. “We focus on chat because being able to talk to a computer is the most amazing thing at the moment. But the LLM is much more than just chat,” he says. “It's very good at handling unstructured data.”
How does McLoone think LLM will evolve over the next few years? With various incremental improvements and training best practices, not to mention potential speed increases through hardware acceleration, You will get good results. “Where big money moves, architecture follows,” McCrone says. But changes of the magnitude of the past 12 months or so could likely be ruled out. Computing costs are disastrous in part, but it may also have peaked in terms of training sets. If the copyright ruling goes against his LLM provider, the training set will be reduced going forward.
However, the issue of LLM reliability will come to the forefront in McLoone's presentation. “Anything that's computational, that's the weakest part, it can't follow rules beyond the really basic ones,” he explains. “If you want to integrate new knowledge or do computing in a data-oriented rather than story-oriented way, computing is still a way to do that.”
Still, the response may vary, and at the end of the day, you have to consider the degree of randomness in ChatGPT, but as long as you give strong instructions to the LLM, this combination seems to work. “I don't know if you've ever seen it, but [an LLM] It actually overrides the facts I gave you,” McCrone said. “When we let plugins go, they often think, 'I don't need to bother calling Wolfram, I know the answer,' and they make something up.
“So if you're in charge, you have to provide very strong and agile engineering,” he added. “Say, 'If you have anything to do with this, make sure you use the tools. Don't try to do it alone.'” But the opposite is true: computation generates knowledge, and LLM I've never seen it ignore the fact.
“It's like the nagging guy in the pub. You whisper facts into his ear and he's happy to take the credit.”
Wolfram will be attending AI & Big Data Expo Global. Want to learn more about AI and big data from industry leaders? Check out the AI & Big Data Expos in Amsterdam, California, and London. This comprehensive event will be held concurrently with Cyber Security & Cloud Expo and Digital Transformation Week.
Learn about other upcoming enterprise technology events and webinars from TechForge here.
