There is still a lack of transparency around how the platform models are trained, and this gap could lead to growing tensions with users as more organizations look to adopt artificial intelligence (WHO).
In Asia-Pacific, excluding China, spending on AI according to IDC, it is expected to grow 28.9% from $25.5 billion in 2022 to $90.7 billion in 2027. The research firm estimates that the majority of this spending, at 81 %, will target predictive and interpretive AI applications.
Also: Five ways to use AI responsibly
So while there are many hype surrounding creative AIChris Marshall, vice president of data, analytics, AI, sustainability and industry research at IDC Asia-Pacific, said the AI segment will account for only 19% of the region’s AI spending.
Marshall, who is speaking at the Intel AI Summit held in Singapore this week, said the research highlights a market need for a broader approach to AI, beyond generative AI.
However, 84% of Asia-Pacific organizations believe that harnessing generative AI models will bring a significant competitive advantage to their business, IDC noted. By doing so, these businesses hope to gain operational efficiency and employee productivity, improve customer satisfaction and develop new business models, the research firm added.
Also: Best AI chatbots: ChatGPT and other notable alternatives
IDC also expects the majority of organizations in the region to increase their spending on edge IT this year, with 75% of enterprise data expected to be created and processed at the edge by 2025, outside of central hubs. Traditional and cloud data centers.
“To truly bring AI everywhere, the technologies used must provide accessibility, flexibility, and transparency,” said Alexis Crowell, CTO Japan, Asia-Pacific, Intel. for individuals, industries and society in general”. “As we see increasing growth in AI investment, the next few years will be critical for markets to build their AI maturity foundations in a responsible and thoughtful way. ”
Industry and government businesses often emphasize the importance of building trust and transparency in AI and letting consumers know AI systems are “fair, explainable and safeHowever, this transparency seems to be lacking in some important respects.
When ZDNET asked whether there is currently enough transparency about how large language models (LLMs) and platform models are trained, Crowell said: “No, not enough.”
Also: Today’s AI explosion will amplify social problems if we do not act now
She points to a study by researchers from Stanford University, MIT and Princeton who evaluated the transparency of 10 major platform models, with the highest scoring platform only achieving a score of 54%. “It was a point of failure,” she said during a media briefing at the summit.
According to the study, the average score was only 37%, which evaluated models based on 100 indicators, including processes related to model building, such as information about training data, structure and risks of the model as well as the policies that govern its use. . The top scorer with 54% was Meta’s Llama 2, followed by BigScience’s Bloomz with 53% and OpenAI’s GPT-4 with 48%.
“No major platform model developer comes close to achieving full transparency, which demonstrates a fundamental lack of transparency in the AI industry,” the researchers noted.
Transparency is necessary
Crowell expressed hope that this situation could change with the availability of benchmark and the organizations that monitor these developments. She added that lawsuits, such as those filed by New York Times against OpenAI and Microsoftcould help bring greater legal clarity.
In particular, it is necessary governance framework similar to data governance laws, including Europe’s GDPR (General Data Protection Regulation), so users know how their data is being used, she noted.
Businesses also need to make purchasing decisions based on What is their data like? captured and where it went, she said, adding that tensions were growing from Users demand more transparency could spur industry action.
Thus, 54% of AI users do not trust it The data is used to train the AI systemrevealed a recent piece of information Sales force surveypolled nearly 6,000 knowledge workers across nine markets, including Singapore, India, Australia, UK, US and Germany.
Also: AI and advanced applications are straining today’s technology infrastructure
Contrary to popular belief, accuracy does not necessarily come at the expense of transparency, Crowell said, citing a research report. Led by Boston Consulting Group.
The report looked at how black- and white-box AI models performed on nearly 100 benchmark taxonomic data sets, including prices, medical diagnoses, bankruptcy predictions, and purchasing behavior. For nearly 70% of the datasets, the black-box and white-box models produced exactly the same results.
“In other words, there is frequently no balance between accuracy and explainability,” the report said. “A more interpretable model can be used without sacrificing accuracy.”
However, achieving full transparency remains challenging, Marshall said. He notes that discussions around AI’s explainability were once heated but have since subsided because it is a difficult problem to solve.
Also: 5 ways to prepare for the impact of innovative AI on the IT industry
Laurence Liew, director of AI innovation at government agency, AI Singapore (AISG), said the organizations behind large platform models may not be willing to provide their training data for fear of being exposed. to sue.
He added that the selection of training data will also affect the AI’s accuracy rate.
Liew explained that AISG has chosen not to use certain datasets due to potential problems with using all publicly available datasets with its own datasets. LLM Initiative, SEA-LION (Southeast Asian languages in a network).
As a result, open source architecture is not as precise as some of the major LLMs on the market today, he said. “It’s a fine balance,” he notes, adding that achieving high accuracy rates means taking an open approach to using whatever data is available. Choosing the “ethical” route and not touching certain data sets will mean lower accuracy rates than what commercial players achieve, he said.
Liew said that although Singapore has chosen a high ethical standard with SEA-LION, this is often challenged by users demanding to exploit more datasets to improve the accuracy of LLM.
A group Authors and publishers in Singapore last month expressed concerns about the possibility that their work could be used to train SEA-LION. Among their grievances is the lack of a clear commitment to “fair compensation” for the use of all articles. They also note the need for clarity and explicitly acknowledge that the country’s intellectual property and copyright laws as well as existing contractual arrangements will be upheld in the creation and training of the LLM.
Open source transparency
According to Red Hat CEO Matt Hicks, such recognition will also extend to open source frameworks on which AI applications can be developed.
Models are trained on large volumes of data provided by copyright holders, Hicks said, and using these AI systems responsibly means complying with the licenses under which they are built. construction, Hicks said, during a virtual media briefing this week following Red Hat Summit 2024.
Also: Do you want to work in the AI field? How to turn your career around in 5 steps
This is consistent with open source models that can have different licensing variations, including copyleft licenses like GPL and permissive licenses like Apache.
He emphasizes the importance of transparency and is responsible for understanding data models and processing the outputs the models produce. To ensure the safety and security of AI architecture, it is necessary to ensure models are protected from malicious exploits.
Red Hat is looking to help its customers with such efforts through a series of tools, including Red Hat Enterprise Linux AI (RHEL AI), announced at the summit. This product includes four components including the Open Granite language and code models from the InstructLab community, supported and compensated by Red Hat.
The open source vendor says this approach addresses challenges organizations often face in implementing AI, including model and application lifecycle management.
“[RHEL AI] it creates a foundational modeling platform to bring open source licensed GenAI models into the enterprise. With InstructLab alignment tools, Granite models, and RHEL AI, Red Hat aims to apply the benefits of truly open source projects — freely accessible and reusable, transparent and is willing to contribute — to GenAI in its efforts to remove these obstacles.”
