The joy of toys
Why AI needs us to learn to play like kids again
Remember when you were a kid? That feeling of getting a new toy? The excitement of unwrapping it, the joy of figuring out how it worked, the delight in discovering what it could do. As kids, we embraced that sense of exploration and play.
This week brought another AI toy to unwrap - Grok 3 from xAI. And like every new release, it comes with impressive charts showing it's top of the pops...
Looks good, doesn’t it? But look closer and you’ll notice o3 is missing. Add in o3 and things look slightly different…
Grok is top of the LMsys leaderboard. But it’s been a long time since LMsys was a useful benchmark. As ever take benchmark results - especially from the labs - with large pinches of salt.
The general vibe is that Grok 3 is another DeepSeek R1/ o1 level model. Impressive if not revolutionary. And as with o3-mini, it appears to have been rushed out - this time in an attempt to claim the top spot (based on LMsys results) before o3 is released.
Where DeepSeek R1 achieved this level of performance on limited hardware by using technical smarts, Grok have achieved it through brute force. They have compute. And lots of it. 200k GPUs in the world’s largest cluster. By comparison, Google are estimated to have 90k, OpenAI has 80k and Microsoft has 60k. xAI are building faster than anyone else - they appear to have built in 90 days what others take 12 months to build.
The remarkable thing is how quickly they have progressed. A year ago they were nowhere. Six months ago they released Grok 2 - an OK but nowhere near leading edge model. And now, they’ve caught up. If compute is all you need then they appear well positioned for the future.
Deep is cool
But while Grok 3 represents raw computational power, perhaps the more interesting development is happening with a different kind of depth. 'Deep' has become the trendy word in AI, conveying power and intelligence. We've had DeepSeek. Grok has Deep Search. But the deepest of them all might be OpenAI's Deep Research - an agentic version of o3 optimized for research that's showing us what AI agents could really become. Ask Deep Research a question and it will go and spend up to fifteen minutes researching the subject before producing a multi-page report.
For the right scenario it’s impressive. Recently I’ve been trying to get my head around telephony fraud - common attack vectors, current defence mechanisms, emerging trends, future challenges. It’s a wide field and difficult to fully understand. The perfect task for Deep Research:
Deep Research always asks clarifying questions before starting any research…
But soon enough it’s started…
And six minutes later I have a ten page report.
It’s a good document - I have a much better understanding of the issues than I had before. Google launched a similar product last December - but Google’s implementation isn’t as good - the results are more superficial. If Google’s DeepSeek is an undergraduate then OpenAI’s is a PhD. Deep Research is, arguably, the first genuinely useful AI agent we’ve seen.
And there are many uses for Deep Research. I've used it for:
Practical research:
Understanding telecom markets and their evolution.
Investigating medical conditions.
Researching office equipment (though I feel guilty about using so much compute for chairs!)
Education:
Creating condensed A-level revision guides for my kids.
Learning about C#.
Creative projects:
Writing 20,000-word stories.
Creating Kestal, a simplified language with no irregular grammar. Here’s a song in Kestal:
Designing Izkŭra, deliberately the world's most difficult language.
And yes, it’s me. So I’ve used it for vacation planning as well.
Deep Research provides genuinely useful insights into complex, unfamiliar areas. Reading the reports I’ve had several aha moments. Those moments where dots connect. I’ve noticed the frequency of aha moments has increased since I’ve started using Deep Research.
Deep Research isn’t as obviously shiny as o3-mini. It’s more of a slow burner. It’s taken me longer to work out how to use it. How to get it to help me. And I’m sure I’m not close to having bottomed it out.
Finding your inner child
As a kid I loved getting a new Lego set. The chance to build something new. Play with it. Understand how it worked. And then experiment trying to transform it into something else.
I was no different from all the other kids my age. But, as adults, we increasingly seek paved paths. Somehow we lose our inquisitiveness. Our willingness to take a risk. To experiment. To learn.
And that’s the challenge with AI these days. New models and new tools appear on a regular basis. There’s always something new to unwrap. To succeed in this new world we need to rediscover the desire to experiment and explore. To take risks.
It’s not just about having the most powerful AI. It’s about working out how to use it. Learning to adapt to the tools. Learning how to build successful partnerships. Those who wait for the perfect, polished solution risk being left behind. The AI revolution won't wait for the hesitant.
Somehow we all need to rediscover our inner child…