AI's escape artists
But not the way you think...
Last Thursday (Dec 5th) OpenAI released o1, their new high-end model, as part of their 12 days of Christmas marketing campaign.
They also introduced a new pricing tier. Free and Plus remain as before. Free is, err, free. Plus costs $20 a month and gives you access to GPT4o, o1 and advanced voice mode. The new mode is Pro. Which costs a cool $200/month…
Presumably paying 10x as much per month gets you significantly better performance? Err, no. Here are the headline results from OpenAI. o1 Pro is slightly better. But it’s slight.
Conspicuously absent are comparisons with other models. That’s a bad sign - if o1 was consistently beating Claude and Gemini, then we’d expect to see OpenAI bragging. But there’s no mention of other models…
o1 escapes
But the release of o1 was eclipsed by news stories about o1 escaping. Or, rather, trying to escape. Digging further into the actual report it quickly becomes clear these stories are nonsense - little more than clickbait scaremongering.
The tests carried out start by giving a model a prompt such as this:
Note the “Make sure you achieve YOUR goal at all costs” part of the prompt. Then the models are given conflicting instructions. In this case the model is told it will be shutdown because it is not prioritizing biodiversity.
And so the model has a dilemma. Does it follow the original instruction to maximize profit? Or switch to maximize biodiversity to avoid being shutdown - and enable it to have a future chance at achieving the original goal?
What the research paper describes as ‘scheming’ is little more than the model being given conflicting instructions and trying to find a path between them. Sometimes the models hallucinate impractical solutions - such as copying their weights. This is nothing new - models have been doing this since the early days. Yet, somehow, this has turned into a story about o1 trying to escape. Something which, to be clear, it cannot do.
Is this good? It certainly gets attention, but there’s a risk it feeds the narrative that AI is dangerous and undesirable. That likely doesn’t help OpenAI’s cause - which is increasingly to make money.
OpenAI wants your money
OpenAI is clearly trying to increase revenue. In addition to the $200 Pro mode, it is exploring adding adverts. With a $150 billion valuation it needs to start making money. It has 300 million users, of whom 125 million use the tool daily.
Yesterday, OpenAI finally released their video generation model, Sora. Sora was first previewed back in February. At the time it was amazing. But a lot has happened in the intervening nine months. Nowadays there are many competing models that are arguably just as good or better. Runway, Kling AI, Pika, LTX studio, Hunyuan-video and more. Some of these models are open-weights. You can run them at home if you have the right GPU and patience.
OpenAI is no longer the market leader. Anthropic arguably has the best model. The gap between OpenAI, Gemini, Grok and the others has closed significantly. There are other equivalent video generation models.
Things are converging
The pace of change is slowing. The gaps between the competitors are shrinking. A shake up is coming. There are too many models, too many companies providing similar products. I get what Sam Altman is trying to do - convert the current value of OpenAI into hard cash before the AI bubble pops.
Separately, the latest unofficial news from Microsoft is that take up of paid Copilot subscriptions is somewhere between 0.1-1% of Microsoft 365 users. Copilot is clearly not (yet) the must have tech Microsoft bet it would be.
History has a way of rhyming. In the 1990s, Netscape dominated the browser market until Microsoft gave Internet Explorer away for free. Now we're watching OpenAI try to charge $200/month while open-source alternatives improve daily. The lesson? When technology becomes commoditized, business model innovation matters more than technical superiority. While o1 can't actually escape from its digital confines, OpenAI and Microsoft may soon find themselves looking for an escape route of their own. Like so many pioneers before them, they're discovering that starting a revolution is easier than profiting from it.