Floating Octothorpe

Thoughts on AI

It's been a minute since I last added anything to this site... well six years to be precise. Since then technology, and more specifically AI, has changed significantly. Large Language Models (LLMs) have gone from an academic topic to being the zeitgeist of recent years.

Unlike a lot of posts on this site, this post is going to speculate on the future; specifically where I think AI is going in the next few years. However, like other posts on this site, this post is entirely my personal opinion; and, like most future predictions, will likely look foolish in hindsight. With that said, here is why I think the adoption of AI is going to be impactful; as well as slower, messier and more constrained by economics, accountability, and institutions than many people currently expect.

Selling Shovels

"During a gold rush, sell shovels"

Back in 2017, google introduced transformer architecture to artificial neural networks with their paper Attention Is All You Need. This opened the door for heavy parallelisation during model training, something GPUs are well suited to. Since then, Nvidia has rapidly grown as a company, overtaking Apple and Microsoft to become the biggest company globally by market capitalisation. Personally I can't see this changing any time soon, while many AI business models are questionable at best, Nvidia are currently upstream of most other companies and selling to a captive market. Longer term I think there are three main things that could change this dynamic:

  1. The demand for AI workloads falls dramatically
  2. New approaches to training supplant or diminish the need for GPUs
  3. Competing companies (likely AMD) catch up or surpass Nvidia

Until then Nvidia will continue to demonstrate that selling infrastructure is a more reliable way to make profit than developing AI itself. Longer term I think the competition will catch up, but chip development is a slow process.

Predictions

AI and Accountability

I came across an old quote from an IBM training manual a couple of years ago:

"A computer can never be held accountable, therefore a computer must never make a management decision."

The original quote is from 1979, and to me is strikingly relevant today.

The capability of LLMs, diffusion models, and generative artificial intelligence has undeniably made huge advances in recent years. Do I think this is going to profoundly change many professions? Absolutely. However, do I think this is going to lead to mass redundancies? No, at least not for the foreseeable future. Ultimately you can't hold an LLM accountable. Mata v. Avianca Inc. is an early example of this, at some point in the chain, a human is still accountable for the decision being made.

Self driving cars are another interesting case in point. The DARPA Grand Challenge started in 2004 as a competition to encourage research and development in autonomous vehicles. This did just that, and within a few years there was increased optimism that autonomous vehicles would become commonplace. CGP Grey's video Traffic Has a Perfect Solution, Humans Won't Use It, is an interesting snapshot in time reflecting this. While fundamentally the video does highlight real problems and solutions provided by autonomous vehicles, it's too easy to come away with the impression that everything is going to change in the next few years; I certainly felt that way when I originally watched the video. With the benefit of hindsight progress has certainly been slower than I would have thought 10 years ago. This has not been helped by companies like Uber who were sharply criticized after the first fatal collision involving a self-driving car. With that said, progress has been made and Waymo now operate a robotaxi service in a handful of U.S. states, and plan to continue expanding their service.

Fundamentally, AI being used as a form of automation is inevitable when there are clear economic incentives, however progress is often slower than people first anticipate; and issues around accountability are at the heart of many of the hurdles facing adoption.

Predictions

Consciousness and Containment

While I think there are legitimate concerns around how AI will impact job markets, I'm much less sold on it becoming an existential threat to humanity. On 30 May 2023 a Statement on AI Risk was published:

Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war.

Among the list of signatories are many notable figures including Sam Altman, Bill Gates, and Dario Amodei. Personally I feel the statement is too alarmist and overplays the capability of a hypothetical future AI. Typically the logic behind such statements is we will eventually produce an artificial general intelligence (AGI), which can subsequently accelerate AI development and produce a superintelligence that may or may not be aligned to humanity. At the extreme you end up with beliefs like Singularitarianism. While definitely interesting ideas to muse over, they very quickly step over the line of reality and into science fiction.

Another plot point that is increasingly crossing over from science fiction is artificial consciousness. Chatbots in particular are easy to anthropomorphise, something that dates back over 60 years to early chatbots like ELIZA. Modern LLMs are undoubtedly more useful, but are they any more conscious? Skipping over what constitutes consciousness, which is itself a contentious subject, the general consensus is no. Again there are interesting questions about what a sentient machine would mean, and while they are philosophically interesting; I don't see any compelling evidence that modern AI systems are approaching anything that would be broadly recognised as consciousness.

Predictions

Corporations and centralisation

One trend that has been clear over the past few years is an increase in X as a service business models. With this has come an increasing reliance on centralised services and consumer subscription models; AI is no exception to this trend. A common phrase from ~10 years ago was:

The Cloud Is Just Someone Else’s Computer

While fundamentally true, for workloads related to AI models most people do not have hardware capable of running (or training) them. It is certainly possible to run local models with projects like Ollama, however consumer hardware with sufficient RAM is frequently a limiting factor. This is only exacerbated by demand outstripping supply as new data centres are built.

The reality for most users today is AI models are only accessible remotely through services like ChatGPT, Claude, and Gemini. This brings with it a number of concerns around privacy, control, and censorship. There are some niche services like Confer aiming to address these problems, but for most users relying on large vendor subscriptions for access to AI models will be an inevitability.

Predictions

Source Code and Security

"Given enough eyeballs, all bugs are shallow."

Software development has already been impacted significantly by the widespread adoption of LLMs, thanks in part to the widespread availability of free and open-source code used as training data. This has opened up a number of interesting questions around copyright, security, and open-source software development.

Copyright has been a contentious topic for AI, significant volumes of copyrighted material are frequently used in training data sets and in some instances can be fully reproduced. However for software development, big companies like Microsoft have put their weight behind LLMs with their Copilot Copyright Commitment. Consequently the threat of legal action isn't a big concern for many companies adopting LLMs to generate code.

The biggest driving factor behind the adoption of LLMs in software development will ultimately be efficiency. Metr have done some interesting research looking at this. Counter intuitively the study found open-source developers were actually more efficient when not using AI despite perceiving the opposite. They have since done followup research showing a small reversal in this trend, but there is still a lot of uncertainty around how much efficiency is gained using LLMs. Despite this, many companies are looking to reduce their developer workforce, recently Atlassian laid off ~1600 employees stating:

Our approach is not “AI replaces people”. But it would be disingenuous to pretend AI doesn’t change the mix of skills we need or the number of roles required in certain areas. It does.

Not to be left out, LLMs are also beginning to have a big impact on security research. With the recent announcement of Claude Mythos, there has been a lot of speculation; fueled by statements like "we view this as a watershed moment for security", and limited access to the model to "critical industry partners and open-source developers". Unsurprisingly this has generated headline news, which in many cases is overly alarmist. With that in mind, I've found Daniel Stenberg's recent post "Mythos finds a curl vulnerability" to be a much more level headed take on where using AI models to find vulnerabilities currently is.

Predictions

Time Will Tell

So how confident am I in my predictions? Not very...

Predicting the future is a notoriously difficult task, and I'm sure I'll look back and laugh at some of the predictions above. With that said, the one thing I am confident about is watching the next few years unfold is going to be interesting.