What if AI stopped getting better?
A new report says AI might be plateauing.
What if AI stops getting better? And the best generative AI models that humanity will ever get is what we already have today.
For today's Unfiltered Friday, I want to talk about a topic that is surely the antithesis of pro-AI supporters: What if AI doesn't keep getting better forever?
The AI scaling law
The last 24 months have been an incredible whirlwind of progress as new AI models were released one after another.
- New models were either demonstrably better, or
- Delivered similar performance with smaller models.
It seemed the party would never end, as everyone banked on the "scaling law" - the idea that training with more data and compute power equals better AI models.
And that it will scale forever.
This led to massive data centres, then data centre campuses, and finally multiple campuses working together to scale up AI training further.
Is AI plateauing?
A new report over the weekend suggested that the party might already have ended.
- Unnamed OpenAI researchers said the next version of ChatGPT is showing a smaller performance jump than with GPT-3 and GPT-4.
- And on certain tasks, the upcoming model "isn't reliably better than its predecessor."
The issue appears to be the lack of original, quality data.
This is a non-trivial problem to overcome, as using AI-created data for training causes 'model collapse'. And yes, they are coming for your data.
Limitations of today's AI
I've been using AI since ChatGPT came out. Today, I use the paid version of the following daily:
- Type ai.
- Claude.
- ChatGPT.
- Perplexity.
If I were to sum up my successes and facepalm moments using AI for writing, it would be these:
- The great averaging machine - Left to itself, AI will say nothing of note, averaging everything to mediocrity.
- Hallucination prone - AI will always find ways to hallucinate, even with explicit prompts and sandboxed environments.
So, AI is far from perfect.
The world has already changed
Some who are deeply worried about AI might welcome the notion that AI might never get better.
However, I will also say that the world is already irrevocably changed by AI - imperfections and all.
And even if AI never gets any better, I expect some jobs to disappear and others to change. It just hasn't reached everyone - yet.
The question is: What are you doing about it?