Know Your Inference (KYI) - Chief AI Officer’s “Holy Grail”

Know Your Inference (KYI) - Chief AI Officer’s “Holy Grail”

I am attending AI Summit from Oxford University today and tomorrow where the Keynote speaker was David Knott, CTO, Cabinet Office, UK Government. He shared an interesting slide which is very much linked to the topic of KYI and how important it becomes for the Chief AI Officers to have clear understanding of the Use-case they are supporting the GenAI to be used.

Above image (referred from David’s keynote), clearly illustrates two distinct use cases that emphasise the critical role of “Know Your Inference” (KYI) in ensuring the Return on Investment (RoI) for AI projects. Let’s break it down:

1. Use Case 1: Urgent Letter Processing

Problem: The system processes 22,000 letters daily.

Solution:

• A model is deployed to identify letters from individuals requiring urgent assistance.

• This model classifies letters into two categories: those needing urgent review and action, and those suitable for normal processing.

Outcome: By accurately identifying and prioritizing urgent cases, the model ensures that critical issues are addressed promptly, thereby enhancing efficiency and service quality.

2. Use Case 2: Querying Government Websites

Problem: Millions of web pages on gov.uk need efficient navigation and querying.

Solution:

• An LLM (Large Language Model) based query tool is utilized to manage this vast data.

Challenges Identified:

• Persistent inaccuracies in responses.

• High public trust in gov.uk materials, meaning inaccuracies can significantly impact trust.

Outcome: The combination of inaccuracies and high public trust signals that more work is needed to improve the model’s precision to maintain credibility and trust.

Importance of KYI

“Know Your Inference” (KYI) is a critical concept for any Generative AI project. Here are some key reasons why KYI is essential:

1. Accuracy and Reliability: In both use cases, the accuracy of the AI model’s inference directly impacts effectiveness and reliability. Misclassification or incorrect responses can lead to significant issues, from failing to address urgent needs to eroding public trust.

2. Trust and Credibility: Particularly in scenarios involving high public trust, such as government websites, maintaining accuracy is paramount. Any errors can quickly undermine credibility, leading to a loss of user confidence.

3. Efficiency and Resource Management: Proper inference ensures that resources are allocated effectively. For instance, by correctly identifying urgent cases, resources can be focused where they are most needed, improving overall efficiency.

4. Continuous Improvement: Recognising areas where the model falls short (e.g., persistent inaccuracies) highlights the need for ongoing evaluation and improvement. This iterative process is essential for maintaining and enhancing model performance.

Open Source vs. Closed AI Models consideration for KYI

When considering AI models for implementing KYI, it’s important to weigh the pros and cons of open-source and closed-source options.

1. Open Source Models:

Advantages:

Transparency: Code and model architectures are open for review, fostering greater understanding and trust in the inference process.

Customisation: Open-source models can be tailored to specific needs, allowing for fine-tuning and optimisation.

Community Support: A broad community can contribute to the improvement and debugging of models.

Disadvantages:

Security: Open access to the model means potential security vulnerabilities.

Maintenance: Requires dedicated resources to maintain and update the model.

2. Closed Source Models:

Advantages:

• Support and Reliability: Typically come with dedicated support and reliability assurances from the providing company.

• Ease of Use: Often more user-friendly with comprehensive documentation and integration support.

Disadvantages:

• Lack of Transparency: The inference process may not be as transparent, leading to potential trust issues, especially in high-stakes environments.

• Limited Customisation: Modifying the model to fit specific needs can be more challenging and might not be possible at all.

• Cost: Often come with licensing fees which can be a significant investment, especially for smaller organisations.

Conclusion

For a Chief AI Officer, understanding “Know Your Inference” (KYI) is essential for maximising the ROI of any Generative AI project. The ability to accurately interpret and act on AI inferences ensures operational efficiency, maintains public trust, and allows for continuous improvement. When choosing between open-source and closed-source AI models, it’s crucial to consider the specific needs of the project, the level of support required, and the importance of transparency and customisation in the inference process.

By leveraging the right AI models and maintaining a strong grasp of KYI principles, organisations can effectively manage their AI initiatives, ensuring that they deliver tangible value and meet their strategic objectives.

Kind regards

Shen Pandi

Click to follow Shen on LinkedIn

TheGen.AI News

TikTok Embraces Generative AI to Revolutionize Advertising Strategy

TikTok recently announced the introduction of its "TikTok Symphony" AI suite aimed at enhancing its advertising services. This suite includes "Symphony Creative Studio," an AI-powered video generator that creates ready-to-publish videos based on simple inputs from advertisers. Additionally, the suite features "Symphony Assistant," an AI tool designed to assist in refining ad scripts and providing campaign optimization suggestions. For instance, it can generate catchy phrases for product launches or offer insights into current TikTok trends.

Another tool, "Symphony Ads Manager Integration," is designed to improve and optimize pre-existing videos for more effective presentations. TikTok is also launching "TikTok One," a hub where marketers can connect with nearly two million creators, find agency partners, and utilize TikTok's creative tools.

Moreover, TikTok is enhancing its advertising solutions with predictive AI that helps advertisers tailor their budgets and goals to select the most suitable creative assets and target audiences. The company also shared insights indicating a significant influence on consumer behavior, with a large percentage of users making purchases or decisions based on TikTok ads.

However, TikTok faces challenges in the U.S. market, following President Joe Biden's recent legislation that could lead to a ban on the app unless its parent company, ByteDance, divests it. This presents an opportunity for other tech firms and startups to fill the void if TikTok were to exit the U.S. market.

Why US Intelligence Remains Skeptical About Generative AI?

The AI revolution is gaining momentum and no one wants to be left behind, including U.S. intelligence agencies. These agencies are increasingly adopting AI as data becomes more prevalent and as they seek any competitive edge.

However, the technology still faces significant hurdles. It's in the early stages of development and is often found to be unreliable. Long before the surge in popularity of generative AI tools like OpenAI's ChatGPT, U.S. intelligence and defense sectors were already exploring AI capabilities.

For example, Rhombus Power utilized AI in 2019 to significantly improve the detection of fentanyl trafficking from China, surpassing the efficiency of traditional human analysis. Additionally, Rhombus successfully predicted Russia’s major military actions in Ukraine with 80% accuracy four months before they occurred.

CIA Director William Burns emphasized in a Foreign Affairs article the need for advanced AI models capable of processing vast amounts of data, both open-source and covertly collected. However, the CIA’s first Chief Technology Officer, Nand Mulchandani, warns of the limitations of generative AI, likening it to a "crazy, drunk friend" — insightful but also prone to errors and biases.

Security and privacy concerns are significant. There's the risk of adversaries tampering with or stealing the AI systems, which might also inadvertently process sensitive data not cleared for access by agents. According to Mulchandani, the primary use of generative AI is akin to that of a virtual assistant, aiding in sifting through massive amounts of data.

Officials are adamant that AI will not replace human analysts. Despite the secrecy around its use on classified networks, it is known that thousands of analysts from the 18 U.S. intelligence agencies are now using a CIA-developed AI tool named Osiris. This tool, which incorporates several commercial AI models, summarizes unclassified data and includes a chatbot for interactive queries.

The potential applications of generative AI in intelligence are vast, ranging from predictive analytics and war-gaming to brainstorming scenarios.

Before the advent of generative AI, intelligence agencies were already employing machine learning for tasks such as alerting analysts to critical developments around-the-clock.

Big players in AI technology, like Microsoft, are actively seeking to integrate their advanced AI models, like GPT-4, into top-secret U.S. intelligence networks. Another significant player, Primer AI, is already serving two intelligence agencies, providing tools that monitor a wide range of news and social media sources in numerous languages to detect early signs of significant events. These technologies are crucial in distinguishing accurate information from misinformation during critical times, as demonstrated in a recent demonstration following the October 7 Hamas attack on Israel.

Introducing WhatsApp's AI: Instant Stunning Profile Photos at Your Fingertips

WhatsApp has introduced a new AI-driven feature that allows users to easily create distinctive and visually appealing profile pictures. This feature, integrated into WhatsApp's user interface, streamlines the process of enhancing digital identities with a touch of personalization.

How It Operates: A User-Friendly Interface

The AI profile photo generator is easily accessible under profile settings. Users begin by uploading an image from their gallery or taking a new photo. The AI then provides various customization options, such as artistic filters and background changes, to tailor the photo. Once adjustments are made, the image can be instantly saved as the user's profile picture, simplifying the creation of professional-looking images for all users, regardless of technical expertise.

Behind the Scenes: Advanced AI Technology

Powered by sophisticated machine learning, WhatsApp's feature focuses on facial recognition to apply precise edits and uses image enhancement techniques to adjust elements like lighting and sharpness. The technology supports a wide array of artistic styles, ensuring versatility in photo customization. This AI system has been trained on a diverse set of images to handle numerous photo types and user requests efficiently.

User Advantages

WhatsApp's new tool offers several benefits:

• Personalization: Users can forge a unique digital presence that mirrors their personality.

• Professionalism: Enhanced profile pictures can boost a user's professional image on the platform.

• Ease of Use: The feature removes the complexity of manual photo editing, delivering professional results effortlessly.

• Privacy and Security: The AI processes images directly on the device, maintaining WhatsApp’s commitment to privacy by ensuring that user data does not leave the device.

Addressing Privacy Concerns

In response to privacy concerns, WhatsApp confirms that all image processing is conducted on the user’s device, avoiding external data transmission. This secure approach is consistent with WhatsApp's priority on protecting user privacy and data.

This innovation marks a significant development in integrating AI with daily digital communication tools, enhancing user experience and paving the way for future technological advancements in messaging applications. As AI technology advances, users can anticipate more dynamic and sophisticated features that simplify digital communication.

In summary, WhatsApp's new AI-powered profile photo generator not only elevates the visual appeal of user profiles but also enhances the overall user experience, reflecting WhatsApp's ongoing commitment to innovation and user empowerment. This tool exemplifies the potential future directions of AI in enhancing digital communication and personal expression.

TheOpen.AI News

Epic Unveils Open-Source AI Validation Tool for Healthcare Systems

Epic, a prominent player in healthcare technology, has recently launched its first open source tool on GitHub, according to the company’s vice president of research and development, Corey Miller, in an interview with Fierce Healthcare. This new AI validation software suite, which is freely available to the public, enables healthcare organizations to test and monitor AI models within their electronic health record (EHR) systems. The tool supports both Epic-developed AI models and those from other creators, facilitating adherence to emerging best practices in AI.

Miller highlighted the accessibility of the tool, emphasizing its public availability on GitHub without any restrictions imposed by Epic. This move reflects Epic's commitment to promoting equitable AI use in healthcare by allowing global contributions to the software.

The suite, termed an "AI trust and assurance software suite" by Epic, was announced earlier in April. It features automated data collection and mapping to provide near-real-time analytics on AI models, explained Seth Hain, Epic's senior vice president of R&D. This automation aims to standardize validation processes and reduce the labor-intensive task of data mapping. Hain underscored the importance of enabling local-level AI testing and ongoing monitoring on a large scale.

Currently, the tool does not support generative AI models, but there are plans for future expansions to include more AI model types. The Health AI Partnership (HAIP), which includes organizations like Duke Health, Mayo Clinic, and Kaiser Permanente, plans to use Epic’s suite for validating AI models for its member health systems.

The suite’s ability to integrate with local data and workflows allows health organizations to analyze AI performance across various patient demographics, ensuring that AI applications are equitable and effective across all patient groups. The intuitive dashboards and common data schema are designed to facilitate future extensions of the tool to new AI models.

Concerns have been raised regarding potential conflicts of interest given Epic’s significant influence in the health IT market. However, Miller assures that the tool is designed to be neutral, capable of working with any predictive models, whether developed by Epic or others. He believes this initiative aligns with Epic’s ethos of contributing positively to the community and improving global capacity to safely utilize AI in healthcare.

IBM Expands Open-Source AI, Secures Partnership with Saudi Arabia

On Tuesday, International Business Machines (IBM) announced its plan to release a series of artificial intelligence models as open-source software, contrasting with competitors like Microsoft who charge for access to their AI models. Located in Armonk, New York, IBM is adopting a strategy similar to Meta Platforms by providing free access to its "Granite" family of AI models, allowing companies to tailor these tools to their specific needs.

The Granite models are crafted to assist software developers in accelerating the coding process. While the models themselves are free, IBM plans to generate revenue through a subscription-based service called Watsonx, which facilitates the smooth operation of these models in data centers post-customization.

IBM's approach is to profit from the practical application of AI, irrespective of whether the models originate from IBM or other providers, and whether they are deployed in IBM’s data centers or elsewhere. IBM CEO Arvind Krishna emphasized the company’s commitment to both competition and safety in the burgeoning field of generative AI.

Additionally, IBM disclosed that the Saudi Data and Artificial Intelligence Authority will utilize Watsonx to train its "ALLaM" Arabic language model. This training will enhance the model's capabilities, including its understanding of various Arabic dialects, thus expanding IBM’s linguistic offerings in AI.

TheClosed.AI News

OpenAI's Recent Setback Highlights Obstacles in Chinese AI Development

The recent launch of GPT-4o by OpenAI, an advanced AI "omnimodel" capable of understanding voice, text, or video inputs, was anticipated to be a significant milestone. However, the company quickly encountered several issues that have overshadowed the release. Firstly, a majority of the safety team resigned, and actress Scarlett Johansson accused OpenAI of using her voice in the model without her permission, prompting the company to enter a phase of urgent damage control.

Further complicating matters, the tokenizer used to train GPT-4o, which helps the model process and understand text, was found to be contaminated with data from Chinese spam websites. This has led to the model's Chinese token library being heavily populated with inappropriate content related to pornography and gambling, which are known to exacerbate issues like inaccurate AI outputs (hallucinations), performance lapses, and potential misuse.

This issue was highlighted in a report I published last Friday after it was brought to my attention by various researchers and insiders in the AI field. They discovered that the majority of the longest Chinese tokens—phrases commonly associated with spam content—made up more than 90% of the non-English language support updates in GPT-4o.

Zhengyang Geng, a PhD student at Carnegie Mellon University, expressed disappointment over the quality of Chinese data used, questioning whether it was a result of poor data cleaning or just reflective of the available online content. This situation raises concerns about potential cultural or linguistic misrepresentations, suggesting that the tokens used might not accurately reflect genuine language usage but rather the type of content that is most accessible on the internet.

Henry Luo, a researcher based in Hong Kong, further analyzed the longest tokens from various languages and found distinct themes, indicating that the issue might be more about the sources OpenAI uses for its data.

Victor Shih, a professor at the University of California, San Diego, humorously noted that avoiding Chinese state media content in training might lead to reliance on less desirable sources, illustrating the challenge of finding balanced training material. Despite some formal tokens from Chinese state media appearing in the model, the overwhelming presence of spam tokens highlights the difficulty in sourcing quality Chinese language data for training purposes.

This problem is not unique to OpenAI; the broader AI industry in China faces similar challenges due to major tech firms like Tencent and ByteDance, which dominate social platforms and restrict access to their data, limiting its availability for training robust language models. Consequently, even powerful search engines struggle with Chinese language searches, as genuine conversational content remains locked within proprietary platforms rather than being accessible on the open internet.

OpenAI and News Corp Forge Alliance, Indicating Shift Towards Super-App Development

OpenAI, a leader in artificial intelligence, and the media conglomerate News Corp have entered into a significant global partnership, announced on Wednesday (May 22). This collaboration will integrate OpenAI's renowned ChatGPT AI chatbot with content from more than a dozen of News Corp's publications, such as The Wall Street Journal, Barron’s, MarketWatch, and the New York Post.

This agreement grants OpenAI access to both current and archived material across a diverse range of subjects including business, finance, politics, and entertainment from News Corp's extensive publication library. Additionally, News Corp will contribute the expertise of its journalists to aid in enhancing ChatGPT's understanding of the news landscape.

Robert Thomson, CEO of News Corp, expressed his enthusiasm about the partnership, highlighting its potential to redefine standards of accuracy and integrity in the digital era. "This landmark accord is not an end, but the beginning of a beautiful friendship in which we are jointly committed to creating and delivering insight and integrity instantaneously,” Thomson stated.

Experts suggest that the implications of this partnership extend beyond merely improving ChatGPT’s data training. According to Nathaniel Whittemore, CEO of the AI education company Superintelligent, the collaboration is strategically aimed at positioning ChatGPT as a primary information conduit, competing directly with Google Search. “It’s about turning ChatGPT into your primary gateway to the rest of the world," Whittemore explained, noting that the partnership is aimed at "organizing the world’s information."

The deal comes at a time when there is a growing consumer interest in a single platform capable of managing various aspects of daily life. Research from PYMNTS Intelligence indicates that nearly 100 million consumers in the U.S. and Australia are interested in a super app that can handle multiple tasks.

This partnership also emerges amidst growing concerns about AI's role in the media sector, with debates about its potential to disseminate misinformation or even replace human journalists. Nonetheless, many view AI as a valuable tool that can enhance traditional media practices.

While the financial details of the agreement were not disclosed, both companies underscored their long-term commitment to collaboratively shape the future of news and information in the AI era.

OpenAI Fails to Deliver on Computing Power Promise for Critical AI Safety Team

In July 2023, OpenAI launched a specialized team named Superalignment, tasked with developing methods to safely manage AI systems potentially smarter than humanity. The company underscored its commitment by pledging 20% of its computational resources to the team's efforts. However, within a year, the team was disbanded amid a series of resignations and accusations that OpenAI prioritized product development over AI safety.

Sources familiar with the team's operations revealed that the promised resources were never fully allocated. Requests for essential graphics processing units (GPUs) were often denied, with the team's actual compute budget falling significantly short of the 20% threshold.

This situation casts doubt on OpenAI's sincerity in fulfilling its public commitments. The company is currently under scrutiny for other issues, including the use of an AI-generated voice similar to that of actress Scarlett Johansson in its AI speech generation features. Johansson alleges that OpenAI's CEO, Sam Altman, approached her for permission to use her voice, which she declined. Despite OpenAI’s claims of coincidental similarity and their statement about hiring another actress for the voice, skepticism remains, fueled by Johansson's assertions and social media speculation that her voice may have been cloned or mixed.

The Superalignment team, initially led by OpenAI cofounder Ilya Sutskever and researcher Jan Leike, was abruptly disbanded following their resignations. The remaining team members, approximately 25, were reassigned within OpenAI. This quick dissolution of a team tasked with addressing potential existential risks posed by superintelligence—AI systems that could surpass human intelligence—raises further questions about OpenAI’s dedication to its visionary goals.

Join the movement - If you wish to contribute thought leadership, then please fill the form below. One of our team members will be back in touch with you shortly.

Form link

Keep reading