AI bipolarity
Summary
I’ve returned to work from a personal tragedy and ever since, I’ve had to balance my AI enthusiasm with my AI scepticism. My AI bipolarity has led me to reflect on the state of the AI bubble, how useful AI really is and the risks of a “just AI it” attitude.
Since my last blog post, I took a break from writing. I didn’t intend for the break to be so long, since I’d only planned a three-week break. But midway through that break, I had to deal with a personal tragedy - one that I’m still recovering from. And while I’m a creature of habits, breaking out of them has given me time to reflect on work, life, and their intersections.
I realise that while my job requires me to be an AI evangelist, I’m also an AI sceptic. Call me AI bipolar if you will. I use AI every day in some way or another. AI will help me clean up this article when I finish my draft. It’ll help me build a thumbnail for socials. That banner you see above is AI-generated. I may even use AI assistance to draft my LinkedIn post, though I fear it’ll sound like every other LinkedIn post out there. But yeah, I use AI all the time these days. At work. At home. Everywhere.
And then there’s the other Sumeet. The other Sumeet fears that the AI bubble will burst and that the whole revolution is a house of cards. Surely it’s a bit perverse if Nvidia gives OpenAI a load of money, so OpenAI can buy Nvidia’s chips? Yes, AI can do a few impressive things. Still, if all these services are a product of an industry bubble that defies the basic rules of SaaS profitability, I wonder how much any business should rely on such services.
Beyond impressive demos
There are also limits to generative AI’s impressiveness. In theory, this is how custom application development could work:
Step 1: Fire up Gemini and ask it to generate a prompt that you can provide to UXPilot to design the screens of the application you want to build.
Step 2: Paste the prompt from Gemini into UXPilot, and watch it magically build your screens.
Step 3: Provide the screens to a vibe-coding tool like Bolt or Replit and bring your application to life.
Step 4: Release your new application into the wild!
It’s incredible to see these new creative processes in action, and I understand why these approaches to knowledge work excite executives about the potential of AI. The trouble is that most executives aren’t hands-on. They don’t always notice that processes like the one I’ve described are great for prototypes and demos, but are often useless for production-grade software.
In fact, a lot of AI is useless for professional outputs. Take Gemini’s Nano Banana release, for example. I returned from my break to news that Gemini was now going to make graphic designers redundant. I’ve given Nano Banana a serious college try, and I can tell you that my best word to describe it is “amusing”. If someone believes that 1024 x 1024 pixel outputs will make graphic designers obsolete, then they don’t understand the first thing about design. Oh, and don’t tell me that my prompting sucks. I don’t want to spend two hours developing prompts to do something I can finish in seconds, using Photoshop!
The need for fundamental skills
Which brings me to the point that most people don’t recognise about AI tooling. We keep hearing news about how AI will eliminate one role or another and how it’ll eliminate entry-level jobs. If you can’t critically examine AI outputs and tweak what it spits out, AI isn’t going to make you more effective! You’ll spend most of your time writing prompts that AI processes in a non-deterministic fashion, and you’ll often frustrate yourself with the stochastic nature of these tools.
Frankly, the most valuable work for AI is to be an assistant for people who are already skilled at their job.
If you’re a skilful designer, UXPilot can help you generate initial concepts that you can fine-tune.
If you’re a skilful content creator on YouTube, Nano Banana and Midjourney can help you create a passable thumbnail - something that can be the equivalent of an intellectual root canal, otherwise.
If you curate a newsletter, AI can help with your SEO, plagiarism checks and spelling and grammar checks.
If you are an experienced programmer, AI can be an in-context assistant—tracing root causes for bugs, generating boilerplate code, and helping you build iteratively.
Each of the above jobs I mentioned is about assisting high-end professionals who engage in deep work. But when executives don’t practice deep work, I fear they don’t see through the flashy demos from AI startups. Meanwhile, corporate foot soldiers suffer through layoffs or, at the very least, unreliable tools that produce slop.
The trouble with “perpetual betas”
In an earlier post, I described the idea of AI as “perpetual beta” software. To this day, I haven’t seen one AI tool that works perfectly. Be it Canva, Gemini, Adobe’s Creative Suite, Replit, Midjourney, good old ChatGPT or even my enterprise favourite, Glean. Stochastic software is software with a personality. And when I want to get a job done, I like boring, predictable functionality, and not the erratic, schizophrenic personality disorders that AI can exhibit.
In short, I want my tools to work reliably. But the AI arms race has shifted everyone’s attention from the core, which must work reliably, to the fringe features that may differentiate one tool from another. Product teams are adding AI features at a breakneck speed, disregarding the user experience, which suffers from a lack of cohesiveness and stability.
Product value lies at the core, but AI makes us push at the fringes
Even as an avid, enthusiastic AI user, I’m most disappointed with how executive enthusiasm about AI further erodes the opportunity for deep work. AI can do things fast, no doubt. But speed and productivity aren’t the same things. Some work needs a cognitive struggle. You can’t “just AI it”.
Sure, AI can orchestrate a series of web searches to build a research report for you. Can you also delegate to AI the work of making sense of that research report in your context? AI is effective at generating concepts, but shouldn’t we invest cognitive effort to evaluate competing ideas, build a creative direction for an ad campaign, develop the visual language for an application, or strategise for a new product capability? Just because code is now a commodity that AI can spit out, do we also throw engineering out of the window? Are coding and engineering the same thing? Are mockups and design thinking the same thing?
The “need for speed” across the industry, combined with capitalism’s lust for cheap labour, is killing deep work, one prompt at a time. Line up a few AI dominos and then watch them work their magic. That’s it, eh? It’s this zero-thought-short-circuiting of that gritty creative process that troubles me the most. Until the AI bubble bursts or the gold rush ends, I don’t see an end in sight to this frenzied approach to knowledge work.