AI alignment problem

Royal Caribbean is using AI to reduce food waste on board their vessels.

Reddit has struck a new deal to sell user data for AI training. The deal is for ten years and is with an undisclosed AI company.

In a move that surprised many OpenAI released Sora AI. Sora's remarkable advancements not only push the boundaries of what's possible in text-to-video technology but also hint at significant progress toward achieving more advanced AI capabilities.

OpenAI is working on a search engine to compete with Google. OpenAi is moving into several areas all at once. We will see if they can execute.

Sora reference papers for OpenAi’s new video creation model Sora.

GPTScript is a new scripting language used to automate your interaction with LLMs.

New work is being done in the area of fixing AI hallucinations.

Softr Ai can help you create web apps

Another AI video creator is HeyGen.

Personalized chatbot for your company website. Use SiteGPT for that.

Braintrust is one of the most interesting books I have ever read. It is a deep science look at how our brains develop and evolve to observe and react to our environment. It is a priceless read.

Alignment and AI

AI alignment, also known as AI safety or friendly AI, is a research field that aims to ensure that AI systems work with human values, goals, and preferences in mind. This field is concerned with developing AI systems that behave in a way that is beneficial to humanity and that avoids causing unintended consequences or harm to humans.

Knowing what to research

One question in AI alignment is that humans have desires and preferences of various kinds, and if we engage in reinforcement learning, we also have reward functions. How do we organize these different desires into an algorithm? Another area of concern is how AI algorithms ‘lock in’ different value systems, These values should be endorsed by groups and communities. There is also work being done in this field with the development of KPIs.

The above is research into why AI may not be aligned with our values but a different set. Or, when AI is not aligned you could think of it as AI does not understand our values well.

So what can we do about this?

One approach is to use machine learning techniques such as Inverse Reinforcement Learning (IRL) to train machines to infer human preferences, goals, and values by observing human behavior.

Another approach involves designing highly autonomous AI systems so that their goals and behaviors align with human values. Additionally, drawing from philosophy, researchers have proposed the concept of the "veil of ignorance" to identify fair principles for ethical AI, which encourages decision-making based on what is fair to everyone involved.

However, there is no consensus on a single set of human values to govern AI, as human values are diverse and culturally rooted. Phenomenological reflection has also led to the proposal of common principles for creating and evaluating AI alignment with human values, emphasizing the need to acknowledge and appreciate the differences in how values function in different human worldviews.

Where will the alignment issue occur?

Deep fakes are an AI alignment issue. Using deepfakes can be considered an AI alignment problem. Deepfakes, which are AI-generated synthetic media that can be used to depict individuals saying or doing things they never did, raise concerns about the potential misuse of AI and its implications for society. If you have heard of the Taylor Swift deepfake episode.

There are also implications for the coming US elections. So several, not all, AI companies have signed an international pact to fight against deep fakes in elections this year. And here is another take from Al Jazeera.

Deepfakes have the potential to undermine trust, spread disinformation, and cause harm by manipulating content in ways that are difficult to detect. Addressing the challenges posed by deep fakes requires efforts to align AI with human values, such as promoting the development of robust detection technologies and establishing ethical guidelines for the creation and use of synthetic media.

Deepfakes pose significant challenges to AI alignment with human values, including trust, disinformation, and harm. To address these challenges, efforts should focus on aligning AI with human values by developing robust detection technologies, establishing ethical guidelines for synthetic media, and promoting transparency and accountability in the use of AI.