Category Archives: AI / ML

What do we mean by “Artificial Intelligence”?

The future is unwritten.

Original Joe strummer mural on the wall of the Niagara Bar in the East Village, NYC.  Memorializes Joe Strummer (1952-2002) and quotes "the future is unwritten" and "know your rights"

Following on from my last post, and triggered by conversations I have had on the subject since then, it occurred to me that a lot of the confusion around “AI” is that almost everyone has a different understanding of the term. Which, of course, makes serious assessment of the subject difficult.

So, let’s define some parameters:

Broadly speaking, Artificial Intelligence is the ability of machines (computers) to simulate the processes usually associated with cognition or human intelligence. This is where the famous Turing Test comes into play – can a machine/computer respond to questions in such a way that the interrogator is unaware that the other party is not human?

However, a broader definition of AI encompasses the abilities to “learn, read, write, create, and analyze”. I think this is more valuable in terms of scope, because it is closer the common understanding of what is popularly termed “AI” today. So let’s break those tasks down a little:

  • Learn – machine learning (ML) is a subset of artificial intelligence. All ML is AI, but not all AI is ML (although most use it). ML is (broadly) statistical analysis on steroids – calculating and weighing patterns and relationships in large data sets. You need to input good data and you need to train the model on that existing data, and then test and refine against other subsets of the data. ML is great at pattern recognition within data and for images and text but it is very susceptible to the correlation = causation fallacy.
  • Read – machines don’t “read”, they have data input to them. However, in this sense, the task refers to ingesting large amounts of submitted content and breaking it down into sections, paragraphs, etc., discarding filler words or data noise, calculating relationships, tracking usage frequencies, etc. This capability is mature because it lies behind full text indexing and search which has been around for decades, but it is still far from perfect.
  • Write – again, machines don’t “write” but they can create somewhat novel assemblages of text (or images) based on statistical rules derived from their input data. This may be a simulacrum of human writing or it may be a word salad (or visual equivalent). Chat GPT and Claude AI are large language models (LLM) that output text based on input prompts and very large data sets based on analyses of a huge corpus of training information. This is where a lot of the hype around “AI” has been focussed in the past 6-9 months.
  • Create – creation overlaps with “write”. The models can present novel output based on their training data and rule sets but is this “creation”? That’s an epistemological discussion that I’m not qualified to judge, but I would point out that while machines can (and do) find relationships between data that humans have not, they are constrained by their training data and cannot “create” anything that has not been submitted to them as input. They can, and do, create new things from old components but currently there is no way for them to create something wholly original.
  • Analyze – this is the part that machines are really good at; and the area where I believe the greatest strides will be made. Humans have been wonderful at collecting data over the past millennia, but there are limits to the ability to retain enough to be able to draw interdisciplinary conclusions. It has been claimed that Sir Isaac Newton in the late 1600s and early 1700s was the last polymath able to be conversant in all aspects of human knowledge, and even then that was probably an exaggeration. Today we generate data way faster than anyone or any organization can track and AI will certainly help fund relationships between disparate aspects of human knowledge. Of course, this is where hubris creeps in – for instance, will we generate more CO2 from running massive GPU stacks and data stores trying to solve climate change? Will all the assembled data of human knowledge be used to manipulate and sell people things?

So, to return to the original question – what is AI? It’s a term that encompasses machine learning, large language models, advanced statistics, novel data collection and organization, natural language processing, and many other tools, approaches, and capabilities. I don’t think it’s productive to buy, sell, worry about, or legislate AI without being more precise in your terms.

  • Will “arm-wavy” AI solve all my business or science problems? No, it will not, but machine learning, natural language processing, and analysis of your internal documentation may provide actionable insights.
  • Will AI cure cancer or solve the climate crisis? No it will not, but the tools that are part of AI may generate novel approaches for research that have been overlooked in the past which could lead to these breakthroughs.
  • Will AI replace my job? In the short and medium term it is possible that some jobs will be replaced by AI processes, but care and feeding of those models will also generate new jobs. Of course, as is so often the case, the skill profiles of the replaced and replacees will be quite different, so this does merit public discussion.
  • Will AI make Skynet1 self aware and lead to the creation of killer robots that can travel back through time to destroy humanity’s last hope? Well, that depends on whether we let Cyberdyne Systems drive our defense allocations – that’s definitely a public policy question.

NOTE: the picture is of the original and best Joe Strummer memorial mural on the wall of the Niagara bar at 7th and A in the East Village. It was painted by Dr Revolt in 2003. It was unforgivably removed and replaced by a “cleaner” version in 2013 after the bar was renovated. Same artist, different vibe.

The full quote from Joe is “(a)nd so now I’d like to say – people can change anything they want to. And that means everything in the world. People are running about following their little tracks – I am one of them. But we’ve all got to stop just following our own little mouse trail. People can do anything – this is something that I’m beginning to learn. People are out there doing bad things to each other. That’s because they’ve been dehumanised. It’s time to take the humanity back into the center of the ring and follow that for a time. Greed, it ain’t going anywhere. They should have that in a big billboard across Times Square. Without people you’re nothing. That’s my spiel. The future is unwritten

  1. Can we talk about the NSA making a surveillance program after the Terminator antagonist? Is this horribly tone-deaf or is it some kind of inside joke? ↩︎

Are we in the hype phase of AI?

The entire tech industry has embraced the “AI” label in the past few months, but how real are the offerings in the marketplace today, and who will reap the benefits of these AI functions and capabilities in many of the tech tools we all use?

AI, ML, LLM and related terms have been emerging in many different areas of tech for the past few years. At Oracle for Research, we funded a lot of AI projects – including use of AI to triage accident victims based on X ray images of long bone fractures, use of ML to interpret three dimensional posture analysis based on the inputs from a smart watch (trained on exercise videos on YouTube), AI assisted molecular modeling for drug screening; and a project for which I was proud to be a co-author on a conference presentation using AI to map agricultural land use in Nigeria from satellite photos. In fact, we sponsored so many AI and ML workloads that I had a weekly meeting with the GPU team to determine where in the world was best to run these workloads to minimize impacts on paying customers.

It’s clear that the impacts of AI and ML in many enterprise systems will be large and I see Microsoft, Apple, Oracle, Google, and others making enormous investments to add these capabilities to consumer and enterprise products. This afternoon I was able to take a photo of a plant in my garden, and the ML integration with the iPhone camera was able to tell me immediately what the pant was and gave me a set of informational links on how best to care for it.

I’ve been using ChatGPT for help on scripting and coding too – it’s great at suggesting R and Bash prompts based on what I have already done – and then I can test whether it’s correct in RStudio immediately. The success rate is not 100%, but it’s pretty good – and more efficient (although probably not as good for my learning) than the countless google searches for suggestions I would have otherwise used.

Realistically, though, how is AI going to impact most of the businesses and organizations that I have spent the past 20 years working with around the world? AI and ML might transform how things are done in Palo Alto, Seattle, Austin, and Cambridge but are they really going to make a big difference for that international steel distributor I worked with? The one that had 35 different ERP systems with no shared data model, data dictionary, or documented processes (and yet was still a billion dollar company). Or the truck parts manufacturer in Indiana with facilities in five countries who didn’t use cloud resources because they weren’t sure if it was a fad? How about the US Federal department that oversees a substantial part of the GDP of the nation – where their managers vaguely waved their arms about “AI” transforming their (non-documented) processes. How, I asked, were they going to train models when they didn’t actually collect data on processes and performance today?

I don’t mean to be a downer, and I think the capabilities of AI and ML can, and will, transform many aspects of our lives but I do worry that most of the people who are the technology’s biggest advocates have no idea how exactly the vast majority of their users (organizations and end-users) work day to day. Most companies and organizations in North America, Europe, and APAC haven’t even mastered and deployed search yet. Employees spend substantial parts of their work weeks looking for things that exist – and many of the largest tech firms are in this situation, not just mom and pop businesses.

The process of transforming most organizations and enterprises around the world to data driven practices – which will then provide data that can be used to train models – is still underway and has been for many years. The general purpose LLMs will be great for fettling language in press releases, and the pattern matching models will be great for sorting and tagging my photos, but true, transformative change to the way that organizations work based on AI insights tailored to their specific needs and trained on their data will be much further away.