Centaurs of Loving Grace

In February of 1996, world champion chess grandmaster Gary Kasparov sat down to fight a robot. Across from him was Deep Blue—a bespoke refrigerator-sized behemoth, stuffed with dozens of chess-specific computer chips. Purpose-built by IBM's best nerds to break Kasparov's heart, the machine had been through its paces beating up high school chess champions and fine-tuning on other grandmasters.

Kasparov promptly crushed it 4–2.

The thing about these machines though (and the wizards who make them), is that they don't stop learning. One year and a generation of chip updates later, Deep Blue got its revenge. It defeated Kasparov 3.5–2.5 (though I'll say I don't know how you win half a chess, and I refuse to learn).

After the loss, Kasparov accused his opponents of cheating. He was described as "crushed and cowed" by Wired magazine. One can empathize. To spend your life building skill and mastery, only to have it replaced by some code-writing dork and his pet wardrobe stuffed with CPUs?

I'm doing a PhD in molecular biology at USC. I spent the first few years of that PhD learning as much code as I could. In my opinion, I'm decent in R, pretty good in Python, and dangerous in Bash (not in the cool way, genuinely a threat to the safety of most computers). I'm no Kasparov. But even so, at times, watching an incoming generation of students vibe code nearly everything necessary makes me wonder what the point was. (Editor's note: Not students in our lab—we'd n e v e r allow it. But you know, students...)

This isn't isolated to biologists writing amateurish code. I think you can actually see this across the spectrum of LLM intervention. It's not lost on me that seemingly the biggest yuck factor we've all seen across the board is AI-generated images (I won't call it AI art and you can't make me). It would appear that the yuck factor of new AI scales pretty directly with the human effort required to build the skill that the AI is replacing. Put another way, the more effort we invested in learning something, the more we resent seeing AI do it effortlessly.

We all hate seeing something do a thing easily that we worked hard to learn. But with AI, I think this isn't a simple fact of jealousy. It's a fact of fear. Lost in the capitalist hustle to not starve is the fact that people derive immense meaning from the work they do. There was work in the garden.

When you look at the AI vision laid out by futurist CEOs of big tech companies, it's easy to feel like that meaning has come under a grave and terrible threat, easy to feel crushed and cowed. So what are we to do?

Well, this brings me to Centaur and why I'm writing it (and why it's named for mythical horse men).

In February of 1996, Kasparov seemed to hold the world back from an AI future. In May of 1997, Kasparov fought the brain's last stand and lost.

In 2005, two amateur chess players with a couple laptops and a new kind of chess program beat not only a handful of grandmasters, but a handful of grandmasters aided by top chess programs and advanced machines.

Our once-defeated Kasparov then had an interesting epiphany—he posited the idea that an average human working harmoniously and effectively with an average AI system is superior to either an advanced AI or human working alone (or working together with a bad process).

Put more plainly, you slap a muscular human torso on a beautiful horse body, and you've got a heck of a thing.

So, how do we build this gorgeous equi-sapien? Well, let's start with the legs.

There is a pervasive and insidious attempt right now by some AI companies to make AI seem like something akin to magic, speaking in openly religious terms of the systems they are designing (tough talk from a guy who referenced the Garden of Eden earlier, but hey, it's my blog).

These systems are not religion. They are not magic. They are vectors in multiple dimensions computing the distance between the words Germany and bratwurst. They are math. They are beautiful, incredible math — but they are math. Sufficiently advanced technology is not indistinguishable from magic, primarily because it super isn't magic.

This cloud of magical chaff has made it difficult to see the forest through the GPTs. Two things can be true at once. These systems can be a step-change advance in human history akin to the printing press and the internet. They are also not a path to something that should replace humans.

They should not replace us because they do not think like us. This is detrimental to certain tasks. An example: Right now, as I write this, I can recall being 3 years old and stepping on a wasp in my Nona and Papa's backyard. That memory instantly connects to being 11 at summer camp, surrounded by wasps at the lunch tables. At 11, I was remembering my mom's story about biting into a sandwich with a bee on it and getting stung on her lip. I connected these memories and moved to eat somewhere else. Sixteen years later, I can tell you this story.

This isn't an uncommon or unique human ability. Memory is something all healthy homo sapiens share. As of writing, if you take the most advanced models in the world from the best research labs, and you feed them 10,000 tokens (words or parts of words in normal people speak), they get confused and starts to perform worse. ChatGPT can't really keep a novella in its head without losing its way. I can recall decades of personal and networked memories in real time and turn them in my mind to illustrate a completely different point about context length in LLMs, context-rot free.

"Well Daniel, you just sound like a luddite."

First of all, you're a luddite. Second, your posture sucks, sit up straight. Third, I'm getting to the optimism, relax.

As I said, these models should not replace us because they do not think like us. However, these models are a step-change advance in human history because they do not think like us. I've been (as of this point in draft one), hacking at this blog post for a couple hours. LLMs (depending on the model) write at "10-100x the speed at which humans can do these things," meaning they could dump out an equivalent amount of blog slop in 20 seconds or less. They're slamming out working code at speeds that would make an average 5th year grad student blush. They can read at a superhuman clip. They are incredibly powerful. Centaur is going to be very clear and optimistic about that power.

But, Centaur (and I hope this is clear by now) isn't going to be a blog about how to use AI to replace us. It's also, critically, not going to be a blog about how to be an AI-free 1996 Kasparov, beating the machines on raw human intellect. That ship has sailed.

This blog is about how to be two dorks in 2005 using three normal PCs to beat the grandmasters and the advanced systems both. This blog is about using your human brain to outsmart the horses, and your incredible hooves to outrun the humans.

Housekeeping if you're still here:

• Please throw your email in the box to get Centaur as soon as new posts go up. Be sure to check your inbox to confirm signing up (and your spam folder in case you didn't get a confirmation email).
• I'm using this blog as an excuse to try to learn as much as possible and grow in how I use AI. As such, I think the next post or two will be about AI and how I'm thinking about using it and going about using it (probably with a focus on current strengths of the systems like coding and analysis). I'm also thinking I want to do a post about how I'm protecting my brain from it and what it means to cultivate human strengths that the current AI systems don't have as easily. After that, I want to do one post at least about the environmental impact (I don't know enough about it to say more than that and frankly neither do you so please don't leave a comment about it). In between I'll probably talk about bio foundation models, but that's not a promise (it's a threat). Stay tuned!
• Also, I write this blog to learn as much as I write it to document my thoughts and (hopfully) teach. I'm by no means an expert. If you see anything in here that is wrong (or even that could just be improved in accuracy), please don't hesitate to reach out!!
• If you're interested in all the ways AI has already made its way into Centaur a (nearly) fully transparent log is available here and will be updated every post