The Unreliable Computer Revolution

For decades, our mental model of computers has been of them being the reference object of total reliability — give a computer the same inputs 1000 times and you'll get the same output 1000 times. Anything else is a bug. A good computer was a Swiss watch made of solid bedrock. They were a way for us to offload our repetitive or boring chores, exactly the chores we needed done fast and accurately, like calculations and data processing.

But that's not true of modern GenAI (Generative AI) systems like ChatGPT, Dall-E, and MidJourney. They are incredibly useful tools that can imagine and reimagine using free association and lateral thinking. I asked ChatGPT to give me ways that a proton was like a sneeze, and in seconds, it had come up with an answer that would have taken me an afternoon at least to match.

So we for the first time in history have computers that can think outside of their box – but here's the kicker – it's also a box that dreams, imagines, and even “hallucinates”.

The double-edged sword of generative AI

The concept of "unreliable computers" might initially strike as an oxymoron. These systems have incredible power to generate new content, ideas, and solutions in the text, images, audio and code - with video and 3D rapidly on the rise. But with that creative capability comes a tendency to “hallucinate”—to make up facts and information that has no basis in reality. Hallucination and generation are two sides of the same coin. A feature, not a bug, even according to Sam Altman of OpenAI. If they were too accurate, they would just repeat the training data they’d seen before. Creativity requires a little crazy. Crazy means unreliability.?

Did you know that ChatGPT can just barely count? Up until a few weeks ago if you gave it some text and asked it to count the words, it failed every time. If you asked it today to generate a 300-word bio of a famous person, it still fails (I just got 331 words). It's a computer system that's bad with numbers.

Its image-generating sister model, Dall-E, is quite hilariously bad, but still claims confidently that it isn't:

I recently posted on LinkedIn about the Microsoft Future of Work Report that said "Knowledge workers are 37% faster, 40% higher quality but ~20% less accurate" using ChatGPT. I still can’t read that without laughing out loud it’s so bizarre. How did they separate quality from accuracy?

Finally, one of the most successful and powerful tech brands in the world sends me an endless firehose of instability and error rate emails.

Reframing how we see computers

GenAI turns the whole notion of computing on its head. It was like waking up to find your dependable accountant had morphed overnight into a whimsical artist, splashing paint across our digital landscape.?

The challenge today lies in integrating this "wacky artist" of a computer into our organizations and workflows in a way that understands its challenges and leverages its strengths. An artist can change the world, even if they’re not the best at spreadsheets.?

To try to set the right context, I often say imagine you were just given a team of 10,000 interns fresh out of school. You could get an incredible amount done with them... but can you trust them? Not without checks and balances, you can't.?

It's this delicate dance of leveraging their brute force power to create, to innovate, while also building processes that account for the reality of how they work.

So what do we do?

Where are the business processes that could benefit from having an unreliable computer generate ideas and content? I've encouraged clients to find those niches in their business processes where a sprinkle of unpredictability isn't just tolerated – it's welcome. Tasks like these are all creative tasks that are perfect matches for this technology:

• Creation and drafting
• Design and planning
• Metadata modelling (at the exploration phase)
• Persona simulations (to drive requirements of all kinds)
• Brainstorming?
• Journey mapping and experience design

Use AI to give yourself a promotion

So our 10,000 interns effectively puts you in charge of virtual team so you can promote yourself up a level to team leader. You design the processes, templates, and guardrails that will get the best out of the whole team.

Many in the industry are comparing this to Daniel Kahneman's Systems 1 and 2 thinking from one of my all-time favourite books “Thinking Fast and Slow”. In it, he discusses how our minds use two types of processing.?

• System 1 is our “Fast system”: It is very much like a generative AI: extremely fast, with access to huge amounts of our unconscious bank of aggregate memories, but often irrational and instinctual. It is the thinking used driving a car (once you’re an experienced driver), by a virtuoso violinist at play, an elite athlete in a flow-state of competition, and also sadly, a racist making a snap irrational judgement. It is a workhorse system designed for speed. It decides which way to dodge when we’re being attacked by a lion or on the receiving end of a spear. Quality is absolutely job 2.
• System 2 is our “Slow system”: It brain system where we do reflection, analysis, and strategy. Here we mull over options and compare attributes. Ironically, our conscious reflective mind has far worse access to our accumulated memories. When you’re making a complex strategic decision, you can’t remember the tens of thousands of experiences, significant and trivial, which are actually shaping your thoughts. You can just remember the main key narrative points and use them to make an executive conclusion.

This executive/worker relationship is a nice metaphor for the human/AI relationship: You, or the humans on your team, are the executives, leading, checking, and guiding your team of fast-but-flakey interns.

Ensuring quality, not chaos: Best of both worlds

The key is learning to combine the strengths of both old and new technology. As I said before, “Traditional software can handle your accounts beautifully and perfectly. It cannot do what GenAI does and vice versa."

We need to bring together the structure and reliability of traditional computing with the unpredictable creativity of modern AI. When we learn how to build these complementary strengths into our teams and workflows, we'll see the true potential of this technology revolution.

The first way to get good results from GenAI is good, consistent, structured prompting. We’ve already given away the RAUX Guide with some examples of how to do this.?

Beyond good prompting practices, there’s also pairing and integrating Gen AI with it’s older, more mature cousins. Although GenAI has co-opted the entire term “AI”, the other types of AI we’ve been using for years are stable, and complement the “youthful exuberance” of GenAI.?

For example:?

• Autocomplete: Love it or hate it, you probably use it every day. From text messaging to emails, computers are finishing our sentences and nudging us towards what we could do next. This was in fact the first mass-market example of generative AI at work.
• Knowledge representation and modelling: The often weird and wonderful worlds of taxonomy, ontology, knowledge graph (KG) databases, and structured content are more important than ever, to give GenAI the structure and rules it needs to understand specific domains like pharma, finance, and manufacturing. Hallucinations may still occur, but their numbers drop drastically when backed up by the structured input of a KG, reducing the GenAI's reliance on their non-existent intuition.
• Autotaggers and categorisers: Tools that make sense of our unstructured data by leveraging things like taxonomies and ontologies to put the tags on content that humans didn’t tag in the first place. They work on text, video, and almost anything else that’s digital, and again, work best if you’ve properly structured their reference models and training data.
• Machine translation: Once a joke, now machines can do quite a decent job of translation between many languages, and these features are getting integrated in daily life across everything from your email inbox to AirBnB reviews and chats between owners and guests.?

All these are AIs that have been proven and stabilised over the years and can be used in workflows to support GenAIs to make more reliable solutions for the enterprise.?

Driving AI quality with content

The process works the other way as well. AIs that are trained on high-quality content outperform generic GenAI by a huge margin. “High-quality” means

• well-written
• well-structured, and?
• very importantly, well-tagged.

Sébastien Bubec - Sr. Principal Research Manager in the Machine Learning Foundations group at Microsoft Research (MSR), said that structuring content to a point of what he called “Textbook quality” gave order-of-magnitude improvements to GenAI accuracy:

“Before working on this kind of work I was focusing on optimisation and improving the architecture. I worked on this for a few years and we could get 2%, 3% improvement. Small, around the edges, is nice but it's tiny.

Suddenly when we focused around the data and focused on really trying to craft data in a way that's more digestible by the LLM at training time, suddenly we saw these incredible 1000x gains. I think it's massive and really pointing to where the gold is. The gold is in the data.”

7 years ago, before I had ever heard of OpenAI, I was working with Barclays, one of Europe’s largest banks. Their business infrastructure lead told me that “of all the efforts we made to optimise the success rates of the AI in our Chatbots, the main thing that moved the needle was improving the taxonomy tagging of the source content.”

A study by data.world showed that structured database improved performance on natural language business questions by huge margins. Sometimes success was still less than 50%, but that was up from 0% to begin with.

Moving forward with AI-aligned workflows

As we navigate this new landscape, it's crucial to remain adaptive and open-minded. The integration of generative AI into our digital ecosystem represents a shift towards a more dynamic, creative partnership with technology. It challenges us to rethink our roles and processes, to imagine a future where humans and AI collaborate in unprecedented ways.

The future will be about finding harmony between the structured and the unpredictable, the accountants and the interior decorators of our digital worlds, or for meganerds like me: the Newtonian and the quantum.?

The journey into the world of generative AI is about finding the sweet spot where creativity and unpredictability intersect with our needs and aspirations. The question isn't just about what AI can do for us, but how we can evolve our mental models alongside it, to unlock the vast new possibilities.

But watch this space...

All that said, that is the current state of AI. This is the fastest-moving tech that humans have ever seen. As fast as the computer revolution was, AI is moving about 5-7 times faster. This is an acceleration rate that the human mind (at least this human mind and most of the ones I've talked to about it) struggle to comprehend.

Apparently, Google has in the works systems that will reduce hallucination drastically. I have no doubt OpenAI has the same. The ideal world being one where we have a "creativity dial", which we can turn up when we want creativity, and down when we want accuracy.

It's not yet on the open market, and I still hold firm on the idea that any intelligence, artificial or not, will always do better with better inputs than worse inputs. But in any case, we're in for a wild ride.

Get in touch if you want to discuss how Urbina Consulting can help you make the most of AI in your business to maximise creativity, reliability, and ideally - both.