diff --git a/posts/llms_dont_make_sense.md b/posts/llms_dont_make_sense.md
index 57c6733..ce0a552 100644
--- a/posts/llms_dont_make_sense.md
+++ b/posts/llms_dont_make_sense.md
@@ -12,7 +12,7 @@ computer. The data? Whatever is sent over the network...
 
 You may be an optimist. Fair. I am not though. I simply cannot see how the 
 network will become a general purpose computer from looking at network data. And 
-for that reason, I don't see how LLMs will learn to be AGI by looking a text.
+for that reason, I don't see how LLMs will learn to be AGI by looking at text.
 
 ## The Human Experience
 
@@ -62,26 +62,25 @@ Reinforcement learning (RL) was king for much of 2000s. Before AlexNet's big run
 and the eventual transformer hegemony, RL was the coolest kid of the block. And 
 I think, it will outlive deep conv nets and transformers.
 
-Value systems are an integral part of decision making. I find the smartest 
-someone is (not academically but on a more philosophical benchmark), the more 
-self-consistent their value system tends to be. The decisions we are heavily 
-influenced by multiple factors like emotion and what we value.
+Value systems are an integral part of decision making. Economics captures this 
+part of us, at least partially (and why I think it and game theory are important 
+for machine learning research). You may like fruit A more than fruit B (or 
+you're like me who doesn't really eat fruit) but after giving you _x_ fruit A, 
+you will probably want fruit B more...
 
-Economics captures this part of us, at least partially. You may like fruit A 
-more than fruit B (or you're like me who doesn't really eat fruit) but after 
-giving you _x_ fruit A, you will probably want fruit B...
-
-And this is important. Really important. OpenAI knows that their is a base 
+And this is important. Really important. OpenAI knows that there is a base 
 entropy in language modeling. You know that too. Think about it. You probably 
 can guess what someone you're close to would say if you text them. But would 
 that change if they were more stressed? What if they normally like your texts 
 but today they got hammered by a bunch of people asking them the same thing? 
 
-How your values changed (as influenced by many factors like emotion) affect how 
-you interact with the environment, including communication, which can not be 
-predicted on average, like a LLM tries to do.
+I find the smartest someone is, philosophically not academically, the more 
+self-consistent their value system tends to be. The decisions we make are 
+heavily influenced by multiple factors like emotion our values. How your values 
+changed affect how you interact with the environment, including communication, 
+which can not be predicted on average, like a LLM tries to do.
 
-I don't think we want an average-perdicting model, we want AGI, something I hope 
+I don't think we want an average-predicting model, we want AGI, something I hope 
 is a bit better than the average person.
 
 ## Architecture is King
@@ -93,7 +92,7 @@ is a bit better than the average person.
 OpenAI has very succesfully [psyoped](https://en.wikipedia.org/wiki/Psychological_operations_(United_States)) 
 us into thinking scaling was king. More parameters, more compute, more data, more
 more more. But, I think they're wrong. Transformers can scale really well. And 
-so we do scale them. But if you think transformers is the finally architecture 
+so we do scale them. But if you think transformers are the final architecture 
 then go off and spend your trillions on training your single model.
 
 Transformers have one flaw. They're not the perfect model. Okay, maybe that's