Generative AI is advancing fast. Some would say too fast, others — not fast enough. The scope of AI safety and security encompasses a wide range of topics. We definitely won’t cover everything, but I still want to explore several perspectives at a high level.
Like the internet revolutionized access to knowledge, Large Language Models (LLMs) are transforming how we interact with information in a much more personalized manner. The need to process several sources and interpret a solution for a given task or question have somewhat faded, instead you can now obtain the exact answer within the context of your problem in an instant (or as fast as the text can be streamed back to you).
The Risks
While the majority of individuals will leverage this enhanced access to information for boosting productivity, it is inevitable that a portion of the global population will misuse it for detrimental purposes. To counter this, companies developing foundational models are diligently working to implement safeguards against sensitive content, though these efforts have not always yielded complete success. This brings us to the concept of prompt injection:
The landscape of jailbreak prompts is continuing to evolve, potentially creating a fresh layer of complexity to the familiar narrative of white hat versus black hat exploit discoveries. Additionally, open source models are edging closer to the benchmarks set by GPT-3.5 and Claude 2.1, such as Mixtral's mixture of experts approach (with promises of a new MoE in 2024 at par with GPT-4) that will be even more difficult to gate.
The ease of access to information is now matched by the ease of its creation, leading to a significant uptick in AI-generated content. This proliferation is increasingly flooding the internet, amplifying a host of issues such as automated bots, deepfakes, misinformation and harassment. This problem might be the most immediate challenge that lacks a definitive solution. Although several companies are striving to develop technologies capable of detecting deepfakes and AI-generated content, their success has not yet met the expectations required to keep pace with the rapid advancements. This could have major consequences in 2024 with elections approaching if we don’t find ways to effectively counter these issues.
Finally, there is AI automation that falls under two primary areas: replacing humans with job automation and AI agency, where it acts with a degree of independence. A key question is whether these advancements in AI automation will lead to widespread job displacement. What are the potential socioeconomic repercussions if this transition occurs too rapidly? Software engineers have already observed that an AI co-pilot can boost their productivity by 40-75% (more so than other job functions so far). This raises an important point: companies will be able to maintain or even increase output and quality with fewer engineers in the near future. It seems we're still only scratching the surface in understanding how this surge in productivity might affect overall employment numbers.
The idea of self-aware AI, particularly in the form of Artificial General Intelligence (AGI), remains a more speculative concern. We haven’t come close to AGI let alone perfecting logical reasoning and complex math problems. Additionally, the Turing Test, once considered a benchmark for identifying AGI, has fallen out as a reliable measure. Of course we can’t overlook the long term impact on humanity once AGI is discovered, but how do we find the right balance between embracing rapid technological progress and ensuring control and safety.
Similar Concerns, Different Perspectives
There are leaders all over racing to develop the most advanced AI: OpenAI, Anthropic, Meta, Google, Mistral etc.
“I think we are seeing the most disruptive force in history here… It's hard to say exactly what that moment is, but there will come a point where no job is needed.” - Elon Musk
“I think AGI will be the most powerful technology humanity has yet invented… A thing that I’m more concerned about is what happens if an AI reads everything you’ve ever written online … and then right at the exact moment, sends you one message customized for you that really changes the way you think about the world.”Sam Altman
“[more must be done] to try and control, measure, steer these models.” - Dario Amodei
“Unlike other things, climate change, pandemic etc. — they don’t have anything in the positive column. When you get to pandemic maybe the only way to solve pandemic is AI. A certain help for climate is AI. [AI] is net strongly in the positive column.” - Reid Hoffman
There's a general consensus on the risks associated with advancing AI, but how to manage the advancement is where everyone differs. Elon Musk has frequently voiced concerns, suggesting that AI development is moving too quickly and poses the greatest threat we face. On the other hand, Sam Altman acknowledges these risks and advocates for collaboration with legislative bodies to establish safety regulations. Despite this, OpenAI shows no indication of decelerating, keeping them at the forefront of the field. Dario Amodei, having been a key figure early at OpenAI, shifted his focus to prioritize safety and security right from the start. This led him to found Anthropic, which has an emphasis on a more foundational approach to these critical aspects in AI development.
Executive Order
Joe Biden’s executive order is an initial attempt to create a framework for safety as companies continue to improve and release foundation models. It is fairly high level covering several key areas:
AI Safety and Security Standards: New safety standards and agencies to monitor developers of high risk AI.
Privacy Protection: Funding for research and enforcing privacy focused techniques.
Equity and Civil Rights: Guidance and prevention of discrimination in various forms.
Consumer, Patient, and Student Protections: Reporting and remediation of AI related healthcare harm and resources for safe AI enabled education.
Worker Support: Mitigating AI’s impact on job displacement with worker support and federal support for workers affected by AI disruption.
Innovation and Competition: Resources for AI research including grants, support for small developers, immigration for skilled professionals.
Global Leadership: An international framework for AI safety and benefits.
Government Use of AI: Guidance for responsible AI deployment, contracting and hiring within government agencies.
These approaches seem like a reasonable starting point, setting a baseline that can be built upon in the future without impacting the current pace of innovation. It's clear that the industry isn't showing any signs of slowing down in the near term.
Embrace the Future of AI
In my view, the advantages of AI currently far surpass the risks it presents. It's crucial for individuals and organizations alike to integrate AI into their workflows, enhancing productivity and adapting to developments in their respective areas. AI is not just a technological advancement; it's leveling the playing field in terms of global knowledge access, with its impact already being felt in sectors like software development, education and healthcare. As we all navigate these rapid technological changes, it's important not to get left behind. Embracing AI means ensuring that we are part of the ongoing evolution that is reshaping our world.