I’m a huge fan of The Wire (one of my all-time favorite shows), and I loved how you have captured those perpetual “juking the stats” moments in Baltimore PD (where everybody scrambles to make the numbers look good instead of actually doing good).
Regarding the AI part (which is my field of work), I wanted to add a small note about the recent o3 AI developments you mention, especially around the debate of “teaching to the test” versus real breakthroughs.
Yes, we can always wonder whether an AI has simply been “trained” to ace a particular benchmark—much like students can be “taught to the test” in school. But the truly exciting part about o3 is its inference-time chain-of-thought reasoning.
Being able to dynamically write and refine its own code as part of the reasoning process (which I suppose is what this new model is doing under the hood) suggests we’re inching closer to a world where AI can explore solutions beyond what was explicitly laid out in training. It’s a little like seeing McNulty in the detail room, working through leads step-by-step; only now, the detective has superhuman coding abilities and almost infinite knowledge.
One of the best indicators of o3’s potential is its performance in Frontier math, jumping from 2% with o1 to 25% under o3, is an undeniable zero-to-one leap which makes me think we are in the beginning of AGI (We have to think about AGI not as a singular point in time but as a process) or very, very close to it.
Again, thanks for the great read—I’m excited to see what next year brings us!
Thanks for this fantastic piece Luis!
I’m a huge fan of The Wire (one of my all-time favorite shows), and I loved how you have captured those perpetual “juking the stats” moments in Baltimore PD (where everybody scrambles to make the numbers look good instead of actually doing good).
Regarding the AI part (which is my field of work), I wanted to add a small note about the recent o3 AI developments you mention, especially around the debate of “teaching to the test” versus real breakthroughs.
Yes, we can always wonder whether an AI has simply been “trained” to ace a particular benchmark—much like students can be “taught to the test” in school. But the truly exciting part about o3 is its inference-time chain-of-thought reasoning.
Being able to dynamically write and refine its own code as part of the reasoning process (which I suppose is what this new model is doing under the hood) suggests we’re inching closer to a world where AI can explore solutions beyond what was explicitly laid out in training. It’s a little like seeing McNulty in the detail room, working through leads step-by-step; only now, the detective has superhuman coding abilities and almost infinite knowledge.
One of the best indicators of o3’s potential is its performance in Frontier math, jumping from 2% with o1 to 25% under o3, is an undeniable zero-to-one leap which makes me think we are in the beginning of AGI (We have to think about AGI not as a singular point in time but as a process) or very, very close to it.
Again, thanks for the great read—I’m excited to see what next year brings us!
Acta non verba.
Excellent post!
Thank You 🙏
Excellent piece !