AI Is Amazing at Some Tasks — and Still Behind at Others

 Writing, coding, creating images, and editing video all take humans a long time. AI is already dramatically faster than us at some of this work - but not all of it. Understanding why reveals a lot about where AI is actually heading.

Where AI Wins: Text and Code

AI is much faster than humans at writing and programming. Part of the reason is how it processes information: text and code are raw data, and AI can move through that data more efficiently than a human reading and typing it out manually.

Where AI Still Struggles: Visual Processing

Processing images and video is a different story. It's costly, slow, and far less mature than text-based AI work - for a few overlapping reasons.

Humans are built for this. We process visual information constantly, every waking hour, and we do it at very low cost to us. For the vast majority of cases, it currently isn't worth handing visual tasks over to AI - humans are already highly efficient at them.

Training data is uneven. AI has been trained on huge amounts of images, and image and video generation has improved a lot as a result. But that's different from AI being trained on how to use the websites and programs we use every day. Every time it navigates a new interface, it's often doing so visually for the first time, even where plenty of documentation exists. That will likely improve with time.

Most tools weren't built for AI. They were built for human visual input. When AI was purely web-based, it was limited - reliant on humans to feed it information a piece at a time. Moving onto the desktop freed it up considerably, especially for text and code, since it no longer had to wait on us to relay everything manually. But anything graphical has remained a struggle.

What Might Change

• Better CLI and text-based tooling - more ways for AI to work through interfaces via commands rather than visuals
• Restrictions elsewhere - some websites or platforms may start locking AI out rather than opening up
• Cheaper visual processing - costs coming down to the point where speed stops being a barrier
• More training data - enough exposure to close the gap with text and code

Until AI reaches that point on cost and training, humans are likely to remain better at these visual tasks - and it makes more sense to point AI at the work it's already efficient at, rather than force it into tasks it isn't ready for yet.

The Deeper Reason: Hardware, Not Just Training

There's a more fundamental issue underneath all of this. Transistors are much faster than neurons - it's part of why a calculator will always outpace a human at arithmetic. But current AI models run as a virtual network of knowledge, simulated in software, rather than the physical, hardwired connections our brains use. For tasks that lean on that kind of architecture, this makes AI less efficient by design.

Better techniques may narrow that gap over time. But without them, closing it fully may require more than smarter software - it may take a fundamental redesign of AI hardware to truly compete with the brain.

What's the Opposite of a Duck

 

You might quickly brush it off, a duck doesn't have an opposite. Or is a duck just too complex to have an obvious opposite?

What makes dead the opposite of alive? What makes hot the opposite of cold?

What's the opposite of a boat, is it a car, a plane, or even a submarine? Can you only have opposites for things with two opposing features, or could you have multiple opposites?

Of course, it depends on how you wish to define "opposites."

But maybe you could define boats, cars, planes, and submarines as being just as opposed to each other as black and white. Or you can call these something else instead of opposites, if you don't like the idea of having multiple opposites.

Male and female are opposites, until you imagine a species that needs a third gender, and now the opposite has turned into a tri-opposite.

 

Opposite Categories

You can have simple concepts, normally judged on one axis, where the opposites sit at either end of the extreme — hot/cold, light/dark.

You can have an opposite based on a binary choice — male/female. In these cases, even a complex concept can have an opposite, because the cultural context makes it obvious what metric we're measuring them against.

You can get situational opposites too. If you went to collect an animal and the choice was a duck or a pig, you'd pick the duck, and the opposite is the pig. But if the choice was a duck or a goose, the opposite is now the goose. These opposites change based on the situation rather than being inherently opposite.

 

What's the opposite of 2? Minus 2. But what about a culture that hasn't discovered the concept of negative numbers — do they simply not have an opposite concept for numbers? Are there opposite concepts out there we haven't come across yet?

What's the opposite of blue? Depending on which colour wheel you pick, it could be one of a few different answers. Sometimes an opposite is based on the metric you're measuring by, rather than something inherent to the item itself.

-------------------------------------------

Amendment: what a language model actually thinks the opposite of a duck is

A while after writing this, I got curious whether any of it could actually be tested. Language models represent words as vectors — long lists of numbers — and the direction between two vectors is supposed to capture their relationship. Famously, if you take king, subtract man, add woman, you land near queen. So I wondered: is there one "opposite" direction the way there's apparently a "gender" direction? Or is opposite-ness a looser, messier idea than that, the way this whole post suspected?

I pulled real word vectors and ran the actual arithmetic rather than guessing. A few things came out.

Antonyms sit next to each other, not across from each other. Hot and cold, up and down, black and white, alive and dead — every pair I tried had a positive similarity, several of them quite high. Nothing was anywhere near "opposite" in the geometric sense. That makes sense once you think about it: hot and cold appear in almost identical sentences ("it was so ___", "the water felt ___"). The model can't tell they mean opposite things, only that they're used the same way. So in this kind of model, "opposite" was never a distance. It's a specific direction between two words that otherwise live right next to each other.

And it isn't one direction — it's dozens of unrelated ones. I checked whether the hot→cold direction pointed the same way as the alive→dead direction, or the good→bad direction, or any of the others. They don't. Almost every pair came back close to zero correlation with every other pair. There's no single "opposite-ness" axis running through the space that every antonym pair sits on. Each pair has quietly learned its own private direction, and those directions mostly have nothing to do with each other.

Gender was the one partial exception. Man→woman lined up reasonably well with boy→girl, sister→brother, aunt→uncle — not perfectly, but there's a real shared pattern for human-family words. It fell apart completely for animals, though: duck→drake and goose→gander had essentially zero relationship to the human gender direction. The model hasn't learned "maleness" as one transferable idea. It's learned "maleness, for people" and "maleness, for waterfowl" as two entirely separate, unrelated things.

So, what is the opposite of a duck? I tried it three ways. Duck minus drake plus goose lands on "bird" — not a specific answer, just a shrug toward the general category. Searching the entire vocabulary for whichever word sits furthest from duck turns up meaningless noise — obscure tokens with no real content, because "furthest away" just finds junk, not opposition. And duck's actual nearest neighbours are goose, bird, rabbit, pigeon — its neighbourhood, not anything resembling an opposite. There's no encoded opposite of a duck. Not because the model failed to find one — because there isn't one to find, exactly as this post guessed at the start.

Boat, car, plane, submarine held up too. Each one sits inside its own tight cluster: boat with sailboat, dinghy, canoe; car with vehicle, suv, truck; plane with jet, aircraft, helicopter; submarine with frigate, battleship, navy. None of them oriented against boat on some shared "vehicle axis" — they're just different categories, each coherent on its own terms. Which is another way of saying: boat doesn't have one opposite. It has several equally reasonable "different from a boat"s, depending on which feature you're grabbing onto — sea versus land versus air versus military. The model, in its own numerical way, ended up agreeing that opposites are as situational as the post argued.