- Secure AI - Thinks and Links
- Posts
- October 21, 2023
October 21, 2023
Randy Lariar
April 20, 2024
October 21, 2023
Happy Weekend!
GPT-4: Beyond Text and Into the Future
The future has landed, and it's multi-modal. For those who've played with OpenAI's chat GPT, there's a new realm to explore, and it's nothing short of exhilarating. Gone are the days when we were limited to just words. Now, we can feed GPT-4 photos or even voice notes and watch its creative prowess soar.
Imagine the possibilities! Beyond just consumer fun, we're talking game-changing business solutions that we've never seen before. Take a simple snapshot, and chat with GPT-4 about it. To give you a taste, I'm penning this post conversationally, chatting with my GPT-4 assistant. No filters, no overthinking—just raw thoughts spilling into the digital cosmos. And GPT-4? It's my editor-in-chief, tidying things up.
Want to see magic? Here's a snap of my fridge—yes, sorry about the clutter! From just this pic, chat GPT gave me a smorgasbord of recipe suggestions. Dive deeper, and voilà, detailed instructions at your fingertips.
Multimodal AI isn't just opportunity; it's a responsibility.
With the power of GPT-4, we're diving into uncharted territories of capability. But it's a double-edged sword. Say an employee snaps a screenshot of confidential data and queries ChatGPT for insights. Boom! That data is in the cloud. Suddenly, we're grappling with potential data loss.
The challenge? Crafting smart policies, ramping up employee training, and relying a tad on the honor system.
But there's another wrinkle: prompt injection. I stumbled upon a clever hack a researcher shared. It appeared as a blank, white image. But feed it to ChatGPT and hidden messages emerge—white text unseen by the human eye. Imagine this lurking on a website. An innocent snapshot by an employee might carry a hidden directive, a ticking time bomb for later action.
This tech frontier is thrilling, but it widens the risk surface. Security and risk pros, it's game time. We have to navigate this new landscape with vigilance.
A few Additional (Human-Written) Notes
ChatGPT wrote the entry above via a conversation, and I made only small edits where it had mis-heard some of my words. The whole thing can be done conversationally, with my phone in my pocket and headphones on. While I’m not too concerned about early adopters rushing to share secretive refrigerator photos or prompt inject white text, it’s important to remember OpenAI isn’t the only player in the multi-modal AI game. There are closed and open source alternatives available now, with more coming. As with earlier technology waves, there’s very little we can do in cybersecurity to prevent someone determined to take photos of sensitive documents while working from home. We can, however set clear policies, communicate them frequently, and hold individuals responsible if they’re discovered to be in violation of those policies.
Perhaps we are entering the “Judgement Economy” – where success is not necessarily about the tools and information you have access to, but rather how wisely you apply them.
State of AI Report 2023
This is a great report put out annually by Nathan Benaich and his Air Street Capital investment team. I’ve been looking forward to the 2023 edition which certainly has a ton more mainstream impact for AI versus previous years. Some of the key take aways about the state of AI:
GPT-4 is still the best LLM as measured against classic Natural Language Processing benchmarks as well as at human tasks like taking the LSATs.
LLMs and other AI models are driving research breakthroughs in other fields such as biology and drug discovery.
Safety discussions are mainstream, however there’s no emerging standard or consensus towards global governance leaving different regions to adopt different approaches.
There are still major gaps in evaluation methodologies to monitoring that an AI is behaving.
Third-party AI Tools Pose Increasing Risks for Organizations
You don’t need to be developing sophisticated new AI models to be exposed to the risk from AI. This point is driven home in this article from MIT Sloan discusses the growing risks associated with the use of third-party AI tools by organizations. As AI becomes more potent and prevalent, errors or misuse could lead to reputational, financial, and legal challenges. A significant concern is third-party AI tools, which 78% of organizations use, and over half use exclusively. These third-party tools account for 55% of all AI failures. The article suggests expanding responsible AI programs, properly evaluating third-party tools, and preparing for regulation to mitigate these risks.
“A responsible AI framework helps companies guard against financial, reputational, and legal risks.”
4 Reasons Why Gen AI Projects Fail
A New Zealand supermarket chain released the “Savey Meal-Bot” to help suggest recipes for customers. Examples were shared online of it suggesting “bleach-infused rice surprise” and “aromatic water mix” that was really a recipe for chlorine gas. Not a great experience. This article explains why many companies have had failures with Generative AI projects this year in the following four categories:
Lack of Governance – a thoughtful, disciplined, and coordinated approach is needed to built the right AI with the right guardrails
Balooning Costs – every API call comes with a fee and as projects move from proof of concept to production those costs can climb more quickly than businesses realized
Unrealistic Expectations – while the latest generation of AI models are extremely capable, they’re still a long way off from science fiction and those limits are sometimes painful to learn about in production
Data issues – garbage in, garbage out. AI models need data for fine tuning, prompting, and to work in connected corporate systems. Lack of clean, usable data will produce problems
Decoding Trust: Trustworthiness Evaluation Framework
The "DecodingTrust" project by Microsoft Research and other institutions assessed the trustworthiness of GPT-3.5 and GPT-4 models across various dimensions like toxicity, bias, and privacy. They discovered GPT models could be misled into generating biased responses or leaking private information. Despite GPT-4 performing better on standard benchmarks, it was found more susceptible to specific adversarial prompts because it was better at following directions. In addition to the research report, Microsoft has made a Github repository available for use in benchmarking other language models.
Gartner Predictions for IT Organizations and Users in 2024 and Beyond
AI as Economic Indicator - Nations are emphasizing AI strategies to enhance productivity and digital economy, bolstered by regulatory frameworks.
GenAI for Legacy Modernization - By 2027, GenAI tools will assist in modernizing legacy business applications, reducing modernization costs by 70%.
Malinformation Combat Costs - By 2028, enterprises will spend over $30 billion to tackle malinformation, reallocating funds from marketing and cybersecurity.
CISOs' Expanded Role - Due to regulatory pressure, by 2027, 45% of CISOs will have broader responsibilities beyond cybersecurity to unify security management.
Unionization Surge - By 2028, unionization among knowledge workers will rise by 1,000% due to GenAI adoption, impacting turnover rates for uncommunicative organizations.
Digital Charisma Filters - By 2026, 30% of workers will use digital charisma filters to enhance social effectiveness, aiding career advancement and inclusive hiring.
Neurodivergent Talent Recruitment - By 2027, 25% of Fortune 500 companies will actively recruit neurodivergent talent for better engagement, productivity, and innovation.
Machine Customer Channels - Through 2026, 30% of large companies will create dedicated units or channels to cater to rapidly growing machine customer markets.
Robots Surpassing Frontline Workers - By 2028, smart robots will outnumber frontline workers in manufacturing, retail, and logistics sectors due to labor shortages.
Electricity Rationing - By 2026, monthly electricity rationing will affect 50% of G20 members, making energy-aware operations a competitive necessity.
When You Play the Game of Thrones, you Win or You Sue OpenAI
The Authors Guild filed a lawsuit against OpenAI, accusing the company of copying entire books without permission which allegedly harmed the authors' book sales and incomes. Notable authors such as George R. R. Martin and John Grisham, among others, are part of this class action lawsuit against OpenAI. Use of Large Language Models that have unclear IP rights for the underlying data is a risk. Microsoft acknowledges this risk and compensates by assuming responsibility for any legal risks involved. If you use other large language models or Open Source AI from other places, you may find yourself saying “Winter is coming.”
Awesome-machine-learning-interpretability
If you want a broad understanding of AI risk and safety, you can look through previous issues of this newsletter. If you’re ready to go deep, bookmark this link. This is a fantastic, curated set of resources focused on AI interpretability, risk management, and security. Reading through these references will help clarify what state of the art is in understanding how machine learning models make decisions and what to do about the situations where such clarity isn’t available (e.g. opaque deep learning models). Risk and security rely on the application of these kinds of tools aligned with policy.