AI, particularly generative AI and large-scale language models (LLM), have made significant technological advances and are reaching a tipping point for widespread industry adoption.and McKinsey report AI talent is already “fully committed to artificial intelligence,” so businesses know they will be left behind if they don’t embrace the latest AI technologies.
However, the field of AI safety is still in its infancy, which poses significant risks for companies using this technology. It's not hard to find examples of AI and machine learning (ML) going rogue.in the field from medicine to law enforcementalgorithms that are supposed to be fair and unbiased are exposed as having hidden biases that exacerbate existing social inequalities, posing significant reputational risk to their creators.
Microsoft's Tay Chatbot is perhaps the most well-known lesson for businesses. The disgraced tech giant who was trained to speak in her conversational teenage patois before being retrained by internet trolls to spew unfiltered racist misogynistic bile It was quickly removed by the publisher, but not before its reputation deteriorated. Damage has occurred. Even her very proud ChatGPT is called „.stupider than you think”
Corporate leaders and boards of directors understand that their companies must begin to tap into their innovative potential. AI generation. But how do you even begin to think about identifying and prototyping early use cases when operating in a minefield of AI safety concerns?
The answer lies in focusing on a class of use cases that I call the „needle in a haystack“ problem. Haystack problems are relatively difficult problems for humans to search for or generate potential solutions to, but relatively easy to test possible solutions. The unique nature of these problems makes them ideally suited for early industry use cases and adoption. And if you recognize the pattern, you'll see that there are a lot of haystack problems.
Here are some examples.
1: Copy editing
It's difficult to check long documents for spelling and grammatical errors. Computers have been able to spot spelling mistakes since the early days of Word, but finding grammatical mistakes accurately has been difficult. Emergence of AI generationand even these often result in perfectly valid phrases being incorrectly flagged as ungrammatical.
You can see how copy editing fits into the Haystack paradigm. It can be difficult for humans to spot grammatical errors in long documents. Once AI identifies a potential error, humans can easily verify whether it is actually ungrammatical. This last step is critical, as even modern AI-powered tools are imperfect. Services like Grammarly already exploited This is what the LLM does.
2: Creating boilerplate code
The most time-consuming part of writing code is learning the syntax and conventions of a new API or library. This process takes a lot of research through documentation and tutorials, and is repeated every day by millions of software engineers. Generative AI trained on the collective code created by these engineers enables services such as: Github copilot and tab nine We've automated the tedious step of generating boilerplate code on demand.
This problem fits well into the Haystack paradigm. Although it is time-consuming for a human to do the research required to generate code that works with an unfamiliar library, it is relatively easy to verify that the code works correctly (i.e., run it). Finally, like everything else, AI-generated contentengineers must further validate that the code works as intended before shipping it to production.
3: Search scientific literature
Keeping up with the scientific literature is difficult, even for trained scientists. millions of papers Published annually. But these papers are a treasure trove of scientific knowledge, and if that knowledge can be processed, absorbed, and combined, patents, medicines, and inventions can soon be discovered.
Particularly challenging are interdisciplinary insights that require expertise in two often very unrelated fields, with few experts having mastery of both fields. Fortunately, this problem also applies to his Haystack class. It is much easier to check the sanity of new ideas generated by AI by reading the original papers than it is to generate new ideas across millions of scientific studies.
And if AI can learn molecular biology in broad strokes the same way it learns mathematics, it will no longer be limited by the disciplinary constraints faced by human scientists.products like typeset It is already a promising step in this direction.
Human verification is important
The key insight in all of the above use cases is that while solutions may be generated by AI, they are always validated by humans. Having AI interact directly with the world (or act) on behalf of large corporations is horribly risky, and history is full of past failures.
Human verification of the output of AI-generated content is critical to AI safety. Focusing on the haystack problem improves the cost-benefit analysis of human validation. This allows AI to focus on solving problems that are difficult for humans, while maintaining easy but important decision-making and double-checking for human operators.
In the early days of LLM, Haystack's focus on use cases can help companies build AI experiences while mitigating potentially serious AI safety concerns.
Mr. Tianhui Michael Lee Pragmatic Institute and founder and president data incubatora data science training and placement company.
data decision maker
Welcome to the VentureBeat community!
DataDecisionMakers are experts, including data technologists, who can share data-related insights and innovations.
If you want to read about cutting-edge ideas, updates, best practices, and the future of data and data technology, join DataDecisionMakers.
You may also consider Submit an article It's your own!