As Bill Gates puts it, the age of AI has indeed begun, and it's a wild time to be a tech exec. ChatGPT has catapulted AI into the global consciousness and every large organization is racing to manage risks and capture opportunities.

AI's advancements, in March alone, are quickly changing how we live, work, and interact with the world. NVIDIA CEO, Jensen Huang, emphasized the "sense of urgency for companies to reimagine their products and business models," OpenAI's CEO touched on AI's risks in an interview with ABC News, and a report by Goldman Sachs says it could replace 300 million jobs. For the optimist, we're embarking on a new era of limitless possibilities. For the pessimist, it might be time to learn a trade.

Software Engineering and Development

March saw the release of over 1000 "AI-powered" products, so it's clear that software engineers have been pretty busy. While OpenAI's models continue to be the driving force behind many of these products and the conversation in general, there's a lot happening at the macro level. Let's look at some of the key happenings shaping AI last month.

AI Training Costs are Falling

As hardware advances, its cost decreases exponentially. Applying Wright's Law, Ark Investments forecast that the cost of producing the hardware needed for AI training will fall by 57% annually. Combining the falling costs of hardware with software optimization, they estimate a 70% annual decline in the price of AI training up to 2030. Wright's Law provides a framework for forecasting cost declines as a function of cumulative production. In this case, it refers to the cumulative output of computational power/hardware that underpins AI training. 

At the same time, competition and improvements within cloud infrastructure continue to break down significant barriers to AI development. For example, Oracle Gen 2 Cloud claims to be faster and cheaper than their competitors, Azure previewed a scalable virtual machine series interconnecting thousands of NVIDIA H100 GPUs at 400 Gb/s to accelerate generative AI, and NVIDIA outlined how they will offer everything from training to deployment for cutting-edge AI services with their partners. These trends meant that competing LLM vendors such as OpenAI, Anthropic, and Cohere could pass on reduced costs to businesses. 

Software optimization is also a major factor in falling costs. Ark Investments estimate that annual neural network software training costs would decline by 47%, and at the start of last month, OpenAI went further. They blew competitors out of the water after the announcement of a public API as well as a 90% cost reduction (before going public, the API was only available to approved users/businesses such as Snapchat) through system-wide software optimization. It’s important to note that these reductions apply to training costs, not operational costs. This dramatic price drop could also potentially be a strategic move to drive early adoption. The public API announcement gave developers direct access to ChatGPT-3.5 early last month. OpenAI then released ChatGPT 4, boasting image understanding, fewer inappropriate/biased responses, and performance advancements.

AI/ML Skills (and Potential Chip) Shortages

With the arms race for AI underway, the field has a significant skills shortage. Machine Learning engineers are in high demand, with some jobs offering over $200k+ salaries for non-management roles. 

https://twitter.com/benmschmidt/status/1635692487258800128

Stanford's AI Index Report cited an increase in AI job postings throughout all sectors, and as companies are rushing to get their AI software to production, there have been many products launched with underwhelming results (*cough* Bard *cough* Ernie).

In March, Microsoft announced they built a supercomputer to power OpenAI's ChatGPT, while it was also reported that they are rationing access to AI hardware for internal teams. With the Enterprise Artificial Intelligence (AI) Market expected to generate over $50 billion by 2026, there's trepidation around hardware demand exceeding supply, potentially disrupting the market.

The Future of Software Engineering and Development: Dev Tools 2.0

Towards the end of March, GitHub launched Copilot X utilizing GPT-4. We're now also starting to see how generative AI will influence software throughout the development lifecycle, as the already widely adopted AI coding assistant can now tag pull requests and answer questions about documentation using an in-editor, ChatGPT-style experience. 

Sequoia's article on Developer Tools 2.0 and generative AI for software engineering makes predictions for the future and highlights some key players that are revolutionizing the industry. We agree that we will see many Copilot-like tools over the next few years. The article describes potential opportunities outside of "get-help-while-you're-coding" software, such as software that makes writing code more secure. Coincidentally, on the same day Sequoia published their article, Codium AI released a beta version of their AI-powered code integrity solution called TestGPT. 

https://twitter.com/benmschmidt/status/1635692487258800128

AI will have a monumental impact on most industries we work in today and in the future. With the current skills shortage and tens of billions of investments in the field, there's never been a better time to get into AI and ML.

Transparency and Ethics

Ethical Dilemmas 

Popular communication platform, Discord, launched an AI-powered chatbot last month and simultaneously deleted a clause from its privacy policy that said it wouldn't store the contents of calls, streams, or channels. Fortunately, consumers noticed, and Discord made a swift U-turn a day later. Following backlash and the reinstatement of their policy, Discord said, "We recognize that when we recently issued adjusted language in our Privacy Policy, we inadvertently confused our users. To be clear, nothing has changed and we have reinserted the language back into our Privacy Policy, along with some additional clarifying information."

Microsoft has also been in the news after they laid off the ethics and society team that taught employees how to make AI tools responsibly. In a statement, the company said, “Microsoft is committed to developing AI products and experiences safely and responsibly, and does so by investing in people, processes, and partnerships that prioritize this". They pointed out that they have increased the number of employees working for their Office of Responsible AI and went on to give kudos to the former ethics and society team "We appreciate the trailblazing work the ethics and society team did to help us on our ongoing responsible AI journey." However, employees highlighted that the ethics and society team ensured that the principles coming from the Office of Responsible AI were actually reflected in the company's products.

Pausing AI Experiments

At the time of writing, the open letter to pause AI experiments has over 50,000 signatures, including high-profile names such as Yoshua Bengio and Stuart Russell. With AI often framed as an arms race, the open letter proposes a six-month ceasefire to "jointly develop and implement a set of shared safety protocols for advanced AI design and development that are rigorously audited and overseen by independent outside experts."

Adhering to this open letter would help "humanity  . . .  enjoy a flourishing future with AI." However, putting this into practice raises concerns over the potential for others countries and institutions to make gains in AI development. China is of particular concern as the country houses 9/10 top institutions publishing AI research, and their government has been accused of cyber espionage and IP theft. 

Realistically, there's no way of slowing down AI's advancements. Still, practitioners should think carefully about the ethical implications of developing AI from ideation to data collection through to production.

Black Box AI

Elon Musk, a co-founder of Open AI, criticized the company on Twitter in mid-February:

https://twitter.com/benmschmidt/status/1635692487258800128

In March 2019, OpenAI LP was created as a "capped-profit" company, with the non-profit OpenAI Inc owning a controlling interest. Four months later, in July, Microsoft invested $1 billion in the LP company and became OpenAI's "exclusive" cloud computing provider. Despite being open source up to GPT-2, the following installment of GPT-3 became a closed source application. As the majority of large LLMs are closed source, we have to make educated guesses about the inner workings of these tools—researchers and practitioners such as Damien Benveniste, Ph.D. Author of The AiEdge Newsletter and Ex-ML Tech Lead at Meta, outline reasoned hypotheses on GPT-4's architecture and training.

Microsoft increased its investment to $10bn at the start of this year, and two months later GPT-4 was released on March 14th. It was accompanied by a 98-page introduction paper that boasts its prowess when tested in various professional and academic settings, even passing a simulated Bar Exam with flying colors. Within the "Scope and Limitations" section of the paper, it states:

"Given both the competitive landscape and the safety implications of large-scale models like GPT-4, this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar."

In not sharing these details, OpenAI also garnered criticism:

https://twitter.com/benmschmidt/status/1635692487258800128

Verifying OpenAI's claim that GPT-4 has fewer inappropriate/biased results without information on its training data is difficult. In the absence of such information, it is not possible to analyze the cause. Only the effect can be observed.

The paper did cherry-pick and highlight some efforts to reduce the harms and biases that GPT-4 can produce.

https://twitter.com/benmschmidt/status/1635692487258800128

GPT-4 Technical Report, Page 50

However, Schmidt goes further in his blog post saying, “Their argument is basically a combination of ‘trust us’ and ‘fine-tuning will fix it all.’ But the way they’ve built corpora in the past shouldn’t inspire trust. When OpenAI launched GPT-2, their brilliant idea was to find ‘high quality’ pages by using Reddit upvotes.”

The paper also references OpenAI research on how they went about red teaming, as well as "A Hazard Analysis Framework for Code Synthesis Large Language Models" which documents hazard severity categories, loss definitions, a risk assessment, and hazard analysis framework for OpenAI’s Codex (the model that powers GitHub’s Copilot).

Balancing individual profit and collective welfare in AI is tough, especially with so much money involved. Even putting aside the money, it's still extremely challenging to understand the many interconnecting ethical, social, economic, and environmental dimensions of AI while new advancements are being pushed to production as fast as they can be made.

AI for Good 

Despite the doom and gloom, we've seen some promising developments in responsible AI in March. Mozilla launched a Responsible AI Challenge where entrepreneurs and AI builders creating ethical and holistic AI can win up to $25,000. They also released a research paper on AI Transparency in Practice in conjunction with ThoughtWorks.

The Ada Lovelace Institution published a discussion paper analyzing the EU's 2021 AI Act and its defined technical standards. The report identifies a regulatory gap between the EU's standardization policy (the process by which the European Union develops and adopts consistent technical standards) and the AI Act, raising concerns about the protection of fundamental rights. 

The paper states:

“If neither the legislative text of the AI Act nor standards clarify how to comply with the AI Act’s essential requirements for fundamental rights and other public interests, AI designers may not implement them effectively, leaving the public unprotected.”

It suggests policy strategies to increase civil society participation, and enhance democratic control over essential requirements, ultimately reinforcing the AI governance framework.

Politics, Policy, and Law

With AI often being rushed to production, we've seen governments and lawmakers worldwide both embrace and grapple with AI.

At the end of last month, Italy's data-protection regulator banned ChatGPT in the country, citing privacy concerns. The ban followed an outage that exposed users' conversations and payment information on March 20th. The regulator said "the mass collection and storage of personal data for the purpose of 'training' the algorithms underlying the operation of the platform" had no legal basis and also touched on the fact that there aren't any tools in place to verify the age of the platform's users. They gave OpenAI 20 days to respond to how it plans to address their concerns. Other regulators like the Irish Data Protection commission and the consumer advocacy group BEUC have echoed similar sentiments. ChatGPT is already blocked in several countries, such as China, North Korea, and Russia (although there are other geopolitical complexities to consider in these cases). 

The UK Chancellor, Jeremy Hunt, announced that the government will launch an "AI sandbox" to encourage research into artificial intelligence. With aims for the country to become a "science and technology superpower" the government plans to launch a £2.5bn into quantum computing. However, the country's legislation on AI has been scrutinized for not ensuring that the use of AI is safe and ethical in the long term.

The U.S. Chamber of Commerce published an Artificial Intelligence Commission Report last month. They claim that "over the next 10 to 20 years, virtually every business and government agency will use AI. " and highlight issues around potential harms to individual rights. Promoting responsible AI is a top priority for the U.S. government, but with the speed at which AI is progressing, it won't be an easy feat.

Last month, the U.S. Copyright Office also officialized a landmark authorship policy regarding generative AI. The policy stipulates the grounds under which AI-generated work can be copyrighted; "whether the 'work' is basically one of human authorship, with the computer [or other device] merely being an assisting instrument_,_ or whether the traditional elements of authorship in the work (literary, artistic, or musical expression or elements of selection, arrangement, etc.) were actually conceived and executed not by man but by a machine." Although there is bound to be much deliberation over the degree to which AI is used as an "assistive instrument" or not, it's a great starting point.

The Romanian Prime Minister unveiled an AI "adviser", telling him what people think in real-time. Named Ion (the Romanian equivalent of John), the AI system "will use technology and artificial intelligence to capture opinions in society" using "data publicly available on social networks," according to a government document detailing the project. Autonomous social media monitoring used by a head of state could make some feel rather uneasy, especially considering the abovementioned biases. 

As in most cases, policy needs to catch up to technological advancement. What's more interesting is how governments utilize AI technology in day-to-day work, such as in the case of the Romanian Prime Minister.   

Pharma, Healthcare, and Life Sciences

We know that AI will continue to disrupt almost every industry, but the beginning of AI within Pharma, Healthcare, and Life Sciences is particularly promising. 

Last month we saw a research article by Lin and colleagues describing the creation of a transformer language model capable of predicting hundreds of millions of protein structures. This groundbreaking work can help us better understand life itself and develop new applications in medicine, agriculture, and many other fields. This represents a potential advancement over AlphaFold, a model which has predicted the shape of 200 million proteins. To put both of these in perspective, we'd been able to map approximately 100 thousand proteins in the preceding 50 years of study.

The National Institute for Health and Care Research (NIHR) is contributing over £1 million pounds to develop AI to cherry-pick organs for transplants. The new deep learning, computer vision model known as Organ Quality Assessment (OrQA) "will be trained using thousands of images of human organs to assess images of donor organs more effectively than what the human eye can see." OrQA aims to allow surgeons to take a picture of an organ, upload it to OrQA and receive immediate feedback on how it should best be used. It is estimated the technology could help up to 200 more patients receive kidney transplants, and 100 more receive liver transplants every year in the UK.

With prolific advancements in the field, the FDA also issued a re­quest for in­for­ma­tion on  Artificial Intelligence in Drug Manufacturing last month alongside a discussion paper containing eight questions. Some of the questions include:

• What types of AI applications do you envision being used in pharmaceutical manufacturing? 
• Are there additional aspects of the current regulatory framework (e.g., aspects not listed above) that may affect the implementation of AI in drug manufacturing and should be considered by FDA? 
• Would guidance in the area of AI in drug manufacturing be beneficial? If so, what aspects of AI technology should be considered?

From these questions, CDER and the FDA hope to consider the application of "its risk-based regulatory framework to the use of AI technologies in drug manufacturing." The deadline for submitting responses is May 1st.

Finance

Finance has been using AI at scale for longer than almost any other industry. As early as the 1990s, AI has been used to estimate credit risk, support portfolio management, and automate trading. Early comments from the industry on the future of AI in Finance, such as those from the deVere Group CEO, seem to indicate more of the same but with increased sophistication and scale. He said,

"AI chatbots and virtual assistants can help financial institutions offer personalized customer service, 24/7, and respond to client queries in real-time. It could also help financial institutions discover fraudulent activities by analyzing large amounts of data in real-time and identifying unusual behaviour trends. As such, this will help financial institutions make better and faster decisions by analyzing facts and figures and providing insights into potential opportunities or risks." 

After offering ChatGPT subscriptions in mid-February, OpenAI announced that Stripe will be powering ChatGPT Plus payments, and that GPT-4 will be used to build AI tools for Stripe. Stripe has already employed AI to help manage fraud and increase conversion rates, but they have also explored how GPT-4 can streamline operations. One of the first enhancements announced is GPT-powered Stripe Docs, allowing developers to ask questions about Stripe documentation and get relevant answers.

Security

Let's start with some good news. The US Transportation Security Administration (TSA) is using AI to reduce unnecessary pat-downs. Designed to ease the passenger airport process for transgender and nonbinary travelers, the AI algorithm for body scanners has been rolling out for the past few months. The TSA attributes the updated algorithm to the shrinking number of pat-downs reported. Previous technology relied heavily on gender binary to determine whether passengers could be hiding contraband. After recognizing a trend of false alarms, an algorithm update was deployed to over 1000 Advanced Imaging Technology units "to significantly reduce false alarms" says R. Carter Langston, a TSA spokesperson.

Less good news is the emergence of AI-powered fraud. A couple in Houston claim they have been scammed out of $5000 after thieves used AI to clone their son's voice. The couple reports that phone scammers impersonated their son and falsely claimed to have been in a car accident involving a pregnant woman.

OpenAI's CEO touched on the risks of the platform during an ABC News interview, saying, "I'm particularly worried that these models could be used for large-scale disinformation. Now that they’re getting better at writing computer code, [they] could be used for offensive cyber-attacks." 

The UK National Cyber Security Centre: published a blog outlining the risks of ChatGPT and large language models. The blog highlights concerns that LLMs might "learn" from your prompts and offer that information to others who query related things. It also touches on the risk of companies with less robust privacy policies acquiring LLMs and taking a different approach to user privacy. Posted six days before ChatGPT malfunctioned, it highlighted the risk of queries stored online being "hacked, leaked, or more likely accidentally made publicly accessible." 

We've seen how AI can be leveraged for malicious purposes, taking AI tools and adapting them to deceive and gain access to external systems. We've also seen how malfunctions can cause security issues, such as leaking private information. What's more worrying is the potential for bad actors and attackers to infiltrate the inner workings of AI systems themselves.

Oprea A and Vassilev A produced a paper called Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations for the US National Institute of Standards and Technology last month. It introduces and expands on various concepts that attackers and bad actors could use to compromise AI systems, such as; availability breakdown, integrity violations, and privacy compromise. The paper's goal is to "create a standard terminology for adversarial attacks on ML that unifies existing work."

https://twitter.com/benmschmidt/status/1635692487258800128

Adversarial Machine Learning will undoubtedly be a critical field as we enter the age of AI, especially following the ChatGPT outage/malfunction on March 20th.

In Conclusion

Despite the length of this article, it is by no means an exhaustive roundup. Frankly, at the pace AI technologies are advancing, there's too much to cover. Although we've highlighted more of the bleaker aspects of AI from last month, here at Janea, we truly believe that AI can create a positive and long-lasting impact on the world. Having seen and worked on AI projects that continue to revolutionize the way we live (such as PyTorch), we're confident that the benefits of AI far outweigh the challenges. As we continue to build on our AI practice, we focus on responsible development, careful consideration of ethical implications, and collaboration between diverse experts to help develop AI technologies that empower humanity.

To learn more about our work in Artificial Intelligence, contact us here.