Lessons From a Decade in AI

When I first encountered neural networks at university, I didn’t realise it would fundamentally shift how I understood programming. I was trained in the classical way: logic, control flow, rules, predictable outputs. You write the exact steps the computer must follow. Nothing more, nothing less.

‍

AI broke that mental model.

‍

Suddenly the challenge was not to hard-code logic, but to show a system enough examples that it could learn the behaviour on its own. It felt like discovering a new paradigm of programming. And it aligned perfectly with two areas I was already deeply interested in:
• psychology
• low-level embedded hardware

‍

AI sat exactly at that intersection.

Why Scaling Neural Networks Was Always Inevitable

By observing how the complexity (number of neurons and, most importantly, the connections between them) in living beings is proportional to intelligence and capability, it stands to reason that:

• you cannot “code” intelligence. It is an emergent quality arising from data, complexity, and learning
• emergence requires a huge amount of compute and memory to support neural networks comparable in size to the human brain
• understanding hardware and software from the ground (transistors) up allowed me to draw reasonable parallels to what might happen if we could scale artificial neural networks

The pattern was intuitive:
Bigger models → more parameters → higher chance of emergent behaviour.

‍

We now see this clearly in today’s large language models. When you grow a model’s size and run it on enough compute, it begins learning capabilities it was never explicitly trained for: new reasoning, new skills, sometimes even new problem-solving strategies.

‍

I wrote about this trend in my bachelor thesis back in 2011, long before AI was a mainstream topic. At the time, it seemed speculative. Today, it is obvious.

‍

This acceleration is why tech companies are building enormous data centres, optimising chips like never before, and pushing model sizes beyond previous limits. We are watching the scaling hypothesis unfold in real time.

Why Compression Matters More Than Ever

Training a massive model is one thing. Deploying is another.

‍

Even the most advanced AI companies do not run their original “full-size” models in production. They are simply too expensive. What they run instead is:

Large model → compressed into a smaller, cheaper, faster model → deployed to users.

This is true for OpenAI, Google, Groq, DeepSeek, everyone.

‍

GPT-5 is a perfect example. The model is not only better than GPT-4; it is dramatically cheaper to run. Half of that progress came from breakthroughs in compression and efficiency.

‍

At OmniShelf, we faced a much harder version of this challenge: How do you make AI run in real time on extremely old devices, with no internet connection?

‍

Years of research (including a European-funded project) led us to new compression architectures that preserve essential behaviours of large models while making them small enough to run on any Android device.

‍

I won’t go into further specifics for now, but this work is what makes our retail execution technology possible.

The Three Biggest Challenges Teams Face When Adding AI to Products

Across the industry, I see companies rushing to “add AI features” without understanding what actually makes an AI project succeed. The problems are almost always the same, and they appear long before deployment.

1. Getting the right data in the right format

Neural networks are only as good as the data you feed them.

‍

Most teams underestimate the difficulty of:
• collecting high-quality data
• labelling it consistently
• cleaning it
• preparing it in a format the model can actually learn from

‍

Great data makes mediocre models look good. Poor data makes even world-class models fail.

2. Generalisation (the most misunderstood success metric)

A model that only works under the exact conditions it was trained on is not useful. A model that generalises (meaning it can perform well in new or unseen environments) is the one you want.

‍

Generalisation is what allows a model trained on one dataset to:
• handle different lighting conditions
• handle different physical layouts
• correctly process new items it has never seen before

‍

This is also where the debate about creativity and reasoning in LLMs emerges. Some researchers see them as “statistical databases.” Others see sparks of genuine creativity. The truth is likely somewhere in between.

‍

We could decompose creativity into:

1. Truly creative genius
   Very rare individuals who advance domains like mathematics, physics, or economics, perhaps 2–5 people per century.
2. General innovation
   Relies on cross-domain knowledge, pattern recognition, analogies, and domain transfer. More incremental, higher volume, lower impact.

‍

What both forms share is the need for an advanced and robust world model that allows a system to think, ideate, simulate, test, and refine. World models exist for niche domains, but a general universal model is still far away.

‍

Current LLMs demonstrate some level of general innovation. They benefit from cross-domain knowledge and pattern understanding. More work is needed in analogies and domain transfer (a current research trend).

3. Accountability and trust

This is the least technical but often the most difficult.

AI systems are probabilistic. They can be right 99% of the time  far better than humans  and still not be guaranteed correct.

‍

When the system makes a mistake, who is responsible?

‍

This challenge appears in scenarios such as:
• loan approvals
• medical analysis
• autonomous driving
• safety checks
• legal decisions

Autonomous driving illustrates this clearly: self-driving systems react faster than humans and statistically cause fewer accidents. Yet if one failure harms someone, the accountability gap becomes immediately visible.

‍

This philosophical and legal uncertainty slows deployment more than any model limitation.

Where AI Is Heading Next

The pace is accelerating, not slowing. As compression improves and hardware becomes more specialised, AI will move increasingly:

from the cloud → to the edge → eventually fully on-device.

This unlocks:
• real-time processing
• full privacy
• offline capabilities
• dramatically lower costs

For companies building real products,  this matters far more than having the “biggest” model.

‍

The winners will not be those who build the largest neural network. The winners will be those who deploy AI efficiently, reliably, and sustainably at scale.

Closing Thoughts

Every week, AI breaks another expectation. But beneath the hype, one thing remains constant: the fundamentals still matter.

‍

Data. Generalisation. Accountability. Efficient deployment.

‍

These are the pillars that determine whether an AI product will succeed or fail.

‍

At OmniShelf, these principles shape how we build, what we optimise for, and how we push the boundaries of what is possible on constrained hardware.

‍