It’s February 7th. Quite cold outside but apparently not enough to cool down the excitement ChatGPT sparked just before winter. Microsoft is ready to announce Bing Chat, a chatbot built on top of the next-generation model from OpenAI and capable of web search — Google is doomed (for the second time in a matter of three months, no less). Everyone will jump to Bing and Microsoft will eat Google’s search revenue.

Or will it?

It’s March 14th. Flowers are shyly blooming in the northern hemisphere. We have been particularly awaiting a rare one with 8 (or is it 16?) petals. It is about to open up (or maybe not). GPT-4 is out; the most secretive release ever for a language model. But it’s better, no — much better than GPT-3.5. Who wouldn’t pay $20/month for a 100x productivity boost? It’s a bargain.

Well, maybe not.

It’s March 21st. A soft, possibly hallucinated melody has awakened Google from a long hibernation. Bard is music for Sundar Pichai’s ears. But some notes are off-key — a rushed release? It may appear the pincer movement of the two above worked just fine. But Bard is just a tester — the real banger will come down the road with more powerful models. Google is back in the race.

Is it, though?

It’s July 18th. The summerest summer ever. The sun is high; the air is hot; GPUs will go brrr once more. Meta is done wasting time on the Metaverse and announces a widely applauded AI release: An open-source second version of their popular LLaMA model, Llama 2. They’ve done it to give OpenAI, Microsoft, and Google a lesson on how to do things in the open.

Have they really?

Alberto, I think you’re missing a few more — you know, Anthropic’s Claude? What about Perplexity? Character? Inflection’s Pi? AI21’s Jurassic? Cohere’s Xlarge? Mosaic’s MPT?… And don’t you dare forget about the gazillion tons of slightly overpriced and overvalued gift-wrapping paper coming out every week!

Oh boy, is this getting absurd.

This article is a selection from The Algorithmic Bridge, an educational newsletter whose purpose is to bridge the gap between AI, algorithms, and people. It will help you understand the impact AI has in your life and develop the tools to better navigate the future.

Maybe that’s enough

I read a piece by writer Zulie Rane about social media platform saturation inspired by the craze to sign up on Threads (only to sign out a week later). It was timely and fantastically relatable. I loved the intro style — which I borrowed for this article. The headline structure is borrowed too, from another article.

I couldn’t help it. The parallelism is uncanny: the exact same phenomenon happening with social media is happening with generative AI tools. Perhaps this is what we have become as a society, driven by the fear of missing out, the sheer amount of overwhelming information, the never-ending drive to accelerate our careers, or the desperation not to be left behind.

Whatever the case, we can’t help it. The generative AI frenzy is making us unconsciously trade off our mental sanity for the absorbing trap of overabundance. I don’t need so much of it. You don’t either. We actually need almost none of it.

The “but two” part of the headline is merely an opinion. (I was thinking of one tool for writing and another for images. But maybe you want a different one for programming. Or maybe you don’t code. Or maybe you don’t care about AI art instead. But you get the idea.) Yet it reflects a very real sensation that I — and I dare to guess you, too — have since ChatGPT went, all at once, inexplicably, unexpectedly, unprecedentedly, and — yes, thankfully — viral.

The world was unprepared for the immense benefits, the cross-sector threats, and the individually felt — yet collectively-shared — sense of AI fatigue. Influencers, marketers, and grifters don’t cause this annoying feeling. They merely adopt it for leverage; the cause precedes them and is hard to avoid.

That’s why I decided to narrow it down. I focused on what I really wanted and stuck to that. The little extra value I would have gotten otherwise wouldn’t have compensated for the mental cost.

Three reasons to avoid generative AI fatigue

So far this is a purely visceral rant but there’s a rationale underlying my emotions.

One tech trick to rule them all

Despite the amazing scope of generative AI technology, the proven efficacy of the products, and the amount of money flowing from one pair of hands to another (most of it without ever leaving Silicon Valley, though), the truth is they all stem from the same technical groundwork.

This isn’t to deny that on any given day I may want to use Bing’s search bar or Claude’s 100,000-token context window or GPT-4’s reasoning abilities or Bard’s prompt multimodality or Pi’s high emotional sensitivity or Character’s versatile array of personalities… But let’s be honest: Do I really need all of them?

In no time, they will be pretty much indistinguishable. The wealthiest companies will manage to commercialize the best products and all of them will share the same fundamental features. The rest — VC-backed startups and smaller LM-wrapper projects — will have to niche down or die out. The tools we use will come down to personal preference but all of them will come from a tiny bunch of big companies. The same group as always.

Even the presumed gaps in capability (which pre-trained models are better) and behavior (which haven’t been RLHFed to the point of uselessness) are irrelevant for most tasks. The alleged temporary degradation of GPT-4 is merely a by-product of OpenAI iterating in public and will be certainly resolved soon, whatever the cause.

In short: less is more in the soon-to-be-commoditized generative AI industry.

Avoid the trap.

Stay away from the tool shipping mill

But maybe you do care about those little differences.

In that case, I encourage you to try a bunch of tools and you will see that sticking to a couple of them is enough. Why? Because you will feel more naturally drawn to some and not others. As Rane says: “Your skills and interests make you a bad fit for 99% of platforms and a great fit for 1%.” She meant that for social media but it can be perfectly extrapolated to generative AI tools. My three criteria are ability (what am I good at?), preference (what do I like to do?), and activity (what do I need it for?)

For instance, let’s do a brief non-exhaustive overview for work tasks: if you’re an author or creative writer maybe high-temperature low-RLHFed base models are the best for you (e.g., GPT-3 or 3.5). If you’re a SEO content marketer or copywriter maybe tailored wrappers are the best choice (e.g., Jasper, although now ChatGPT, Bard, and Claude would do just fine). If you’re a tech-savvy writer maybe Llama 2 is better to avoid dependency. If you’re a digital artist, Midjourney. If you have coding skills and want higher steerability, Stable Diffusion instead. As a coder, GPT-4 or GitHub Copilot. Data analysis? Code interpreter.

Looking at the tools that are coming out is fine, but I’m more productive if I stick to one or two; having to analyze the news every week to see if the new thing is 0.1% better than the current thing is exhausting and the main ingredient of burnout.

The user-company inherent mismatch

There’s a third reason — rather peripheral to the others — for why it isn’t worth getting generative AI fatigue.

Although the landscape appears to be a web of race-like dynamics, conflicts of interest, and business tensions whose inevitable outcome is competition-driven consumer well-being, all the companies I’ve named above have tight, favorable, and mutually-beneficial relationships whose sole goal is to make us, the users, pay for their products (including, eventually, those which are currently free).

These companies aren’t shipping more and more and more products to provide more and more and more value but to get a portion of the succulent generative AI pie. Which isn’t bad. I mean, who wouldn’t do the same? It’s expected — in no way worse than social media platforms — but worth bearing in mind just in case you thought we are the main beneficiaries of this — we are not. Companies won’t hesitate a single second to take a direction that doesn’t benefit us. They won’t hesitate to cut off access to products and shut down entirely their services if they must.

So, in closing, generative AI tools can be a blessing. They can also be a curse. Life is too short to be chasing all the time after things we don’t really need.

Narrow down. Stay freed. Avoid fatigue.

 
 
Alberto Romero is a freelance writer who focuses on tech and AI. He writes The Algorithmic Bridge, a newsletter that helps non-technical people make sense of news and events on AI. He's also a tech analyst at CambrianAI, where he specializes in large language models.

 
Original. Reposted with permission.