There’s More to AI Than Just Agents – A Perspective from Martin Rückert, Chief AI Officer at Tallence AG

If you follow the media coverage of AI these days, you could easily get the impression that AI is an entity of almost human-level intelligence - able to think independently, operate fully autonomously, and eagerly fulfill our every wish. While today it exists only on our screens and talks to us through our speakers, many believe it will soon manifest in our homes and physical environments, ready to perform almost any task we choose to automate.

Unfortunately, we’re fooled by the impressive generative capabilities of the technology behind these “AIs” - the enormously large models that have reached unprecedented levels of quality.

To us humans, being able to write and speak eloquently is the ultimate sign of intelligence. Evolution shaped us so that the ability to understand questions and respond convincingly became the test that separates the dumb caveman from the genius scientist. Intelligence is an evolutionary advantage.

However, the “AIs” we work with today are, first and foremost, perfect word-generation machines. They’re so good at producing text - and, by extension, text-based conversations - that this perfection tricks us into associating them with human-level intelligence. The term “hallucinations” - often used when talking about large language models - is not really a bug; it’s by design. The core concept of a language model is to generate plausible language, not to answer questions accurately or reason intelligently.

Companies are flocking to these models that generate text (and now other types of data like video and audio) because they’re hyped by investors and, often, with lots of handcrafted prompt chaining - which is common these days - they produce surprisingly good results, as long as your life doesn’t depend on the result.

For the first time in human history, we can see the dawn of the ultimate intelligent machine - one that could take over everything we consider work and perform tasks we’d prefer a machine to handle. Since every value creation process centers around humans and machines - and machines typically improve the cost/value ratio - a borderline intelligent machine is every economist’s dream: create any product or service at near-zero marginal cost. This is why investors throw billions at Ilya Sutskever, ex-Chief Scientist at OpenAI, now running his own venture “SSI - Safe Super Intelligence,” even though it has no product, not even an early demo.

Whether we will achieve artificial general intelligence (AGI) or even artificial superintelligence (ASI) remains unclear. It’s also debatable whether human-like intelligence is what we should want in a machine. I would argue that we won’t replicate human intelligence in a machine because machines don’t need to mimic our full existence - and since human intelligence is tied to our physical being, it represents just one form of intelligence useful for humans. However, I’d argue we can certainly achieve superhuman performance in all sorts of narrow tasks - and that’s strategically where the industry should focus its efforts.

Since current AI is not made for full human-level performance, here’s a simple rule of thumb for using AI practically and sustainably:

1. When the task is dull and repetitive but not narrow, use AI to do the work - but involve a human to check correctness.

(Examples: document understanding, answering customer support calls, transferring complex information from one system to another, booking meetings.)

1. When the task is dull and narrow, let AI handle it autonomously.

(Examples: document OCR, boilerplate code modules, image classification, audio transcription, industrial defect detection, finding a free slot in calendars, finding similar artifacts, etc.)

1. When the task is complex and not narrow, don’t rely on AI alone.

(Examples: high-value lead generation, writing quality code based on a rough product idea, negotiating contracts, running a company, etc.)

Before diving into sub-discussions based on that list, let me close this blog with my recommendation for the coming months (you never know…): today’s popular AI models are nowhere near as human-level intelligent as we might think, but companies are investing as they learn what can - and cannot - be done reliably. What we see is companies discovering all the limitations as they use popular AI models, then implementing fixes and guardrails. Fundamentally, though, expecting perfection is hopeless for many reasons. So, let’s continue investing in tasks where perfection can be achieved - and trust me, there are countless such tasks. But don’t build products or services that just ride the hype wave of so-called AGI or ASI and then fail in the realities of everyday life - it’s disappointing for everyone, including customers and investors.

While some may think current model architectures are the be-all and end-all, that’s not actually the case. For some tasks, new or currently unpopular model architectures could be much better suited - and these models and system architectures are already being tested. I’m very optimistic that these new approaches will deliver far greater levels of autonomy than today’s models.

So, stay informed. Talk to your AI teams and experts - they can help you develop sustainable AI strategies your customers will love - even after the demo is over.

Yours,

Martin