You use an NVIDIA GeForce RTX GPU and an Apple Mac Mini with unified memory for running open-source LLMs locally. You realise that systems without enough GPUs, which cannot run LLMs, stop making sense to you. Your interfaces to the world are digital. Your devices are digital. You are digital. You replaced writing on paper with writing on a reMarkable tablet, saying that writing on the device’s surface feels just like writing on paper. You accumulated three variants of Kindle readers and three variants of iPads. You grew with every device launch. You don’t let go of your AirPods (the second set) even after realising that they don’t work well with your Windows GPU box and SkullSaints mini-PC. You didn’t bother to check whether the AirPods worked with your Linux mini-PC, as you never made time to do so. Your Apple Watch and iPhone track your daily movements, sleep patterns, and conversations.
The 27-inch iMac is already outdated and a relic of a bygone era. You wanted to use a 27-inch desk monitor. You needed something portable, so you bought a portable monitor to go with your various mini-PCs. Don’t get me started on your list of wireless keyboards, mice, and digital pencils. A smart ring tracks your sleep. Its worktime is your sleep time. You program the next best product and launch it in the market, curiously tracking every new registration.
We have spoken about digital have-nots on many occasions. They must cover a longer distance now, from non-digital to digital and then from digital to AI. Queues at government ration shops continue to be longer than those outside stores launching the next version of the smartphone. It’s unfair to expect someone to build a machine learning model while they don’t know where today’s dinner will come from. Ironically enough, building an ML model is one way to earn dinner. We frequently take a laptop, a desktop, a phone, a pair of earphones, and an internet connection for granted. Ask those who don’t have them.
In the race to become the best data scientist, build the next best LLM, launch the next useful digital product, earn that promotion, secure this year’s hike, are we leaving half of the world behind? While we are taught to score the top marks in a class, shouldn’t we also be taught to discuss a plan for the bottom half of the class? Can AI help?
The uncomfortable truth is that scale in AI has largely been built on abundance of data, compute, and connectivity. But inclusion demands designing for scarcity. Low-bandwidth models, offline-first systems, speech interfaces in native languages, and hardware-aware optimisation are no longer “nice-to-have” research problems. The next breakthrough is not just a larger model, but a more accessible one. A 7 billion-parameter open-source large language model that runs reliably on a second-hand CPU in a rural school may create more societal value than a trillion-parameter model locked behind APIs.
There is also a behavioural gap. If the datasets we curate, the products we design, and the incentives we chase continue to centre already-digitised users, the divide will compound quietly but irreversibly. The question is no longer whether AI can help, but whether we are willing to redirect enough attention to make it help where it matters most. Increasing coverage of your AI system within the digitised world is the low-hanging fruit. Go and get it. However, we must not lose focus from the fruit hanging high: reaching those who have not seen the bytes.
The future of AI will not be judged solely by intelligence, but by distribution. Not by how well it performs in ideal settings, but by how meaningfully it operates in constrained ones. Bytes built the system. Bites will test its purpose.
Disclaimer
Views expressed above are the author’s own.
