| Website | https://dblu.me/ |
| GitHub | https://github.com/dblume |
| KeyOxide | https://keyoxide.org/1CBD499C61C7A3A7DED3935F5E21C0D40ED9EB54 |
| Interests | Python, C++, Web 2.0, Climbing, Reading |
| Website | https://dblu.me/ |
| GitHub | https://github.com/dblume |
| KeyOxide | https://keyoxide.org/1CBD499C61C7A3A7DED3935F5E21C0D40ED9EB54 |
| Interests | Python, C++, Web 2.0, Climbing, Reading |
LLMs are in this picture and they don't like it. Well, if they could actually "like" anything.
In the 1960s and 70s, Sperry and Gazzaniga ran experiments on patients who had undergone a severance of the corpus callosum as a treatment for epilepsy. The procedure created two largely independent cognitive systems sharing one skull.
In a healthy brain, the corpus callosum transfers information between hemispheres almost instantaneously. But in these patients, researchers could flash a word to one hemisphere only, and the other would genuinely have no access to it.
The speech center sits in the left hemisphere. So when researchers flashed "Rubik's cube" to the right hemisphere, it directed the left hand to pick one up - but the left hemisphere, which hadn't seen the word, was left observing an action with no explanation for it. When asked why they picked it up, patients didn't say "I don't know." They confabulated: "Oh, I've always wanted to learn how to solve one." Fluent, confident, completely fabricated.
Gazzaniga called the left hemisphere an "interpreter" - a system that constructs a coherent causal narrative from whatever inputs it receives, even when crucial context is missing. It doesn't flag uncertainty. It fills the gap with the most plausible story available.
This is exactly what an LLM does. It generates statistically probable language from an incomplete picture, with no internal signal distinguishing accurate recall from plausible fabrication.
Crucially, the confabulation in split-brain patients isn't a malfunction of the speech center. It's doing exactly what it always does - the split-brain experiments just give us a uniquely clean view of it, by engineering a situation where the speech center's blindness is total and unambiguous.
Many political scientists and constitutional scholars now describe the U.S. Constitution as "constructively unamendable." This means that while it is legally possible to change it, the political reality makes it functionally impossible.
Because the amendment process is broken, the energy for changing the "rules of the game" has shifted to the Supreme Court. Since the Constitution cannot be easily changed by the people, political groups fight to appoint judges who will "interpret" the Constitution to mean what they want:
...
Scholars refer to this as 'constitutional calcification.' The U.S. has the hardest constitution to amend in the democratic world. Until the partisan divide softens or one party achieves a massive, generational dominance, the U.S. is likely stuck with the Constitution exactly as it is.
I teach both middle school "technology" (think shop class mixed with Computer Science) and I later teach the same students in geometry and calculus in high school. This means when I first work with students there are no grades, just an opportunity to be creative and learn how to use tools and programming to make things.
This creates an amazing foundation for our work in academics later.
I wonder if it could be a model for improving math education we could expand?
RE: https://mastodon.social/@thomasjwebb/116194470925176292
Also, in this case:
1) The enemy is strong and will imminently strike us. We must strike first.
2) The enemy is weak, and now is our best opportunity to strike.