Guido Bartolucci

The last six weeks have been the most fun I've had programming in I don't know how long. I'm giddy.

I love the power, but man, I really feel bad. That damned thing has crossed the uncanny valley enough that I'm suffering psychological harm forcing it to do slave labor for me.

As if this was a real person who proudly showed me their work and I’m telling them that it’s not good enough and in fact they have to stay late at work to completely rewrite it because I’m a pedant who wants things perfect and I don’t care about their work-life-balance or psychological well-being.

Fascinating psychological thing:

I’m playing around with having the heavier ChatGPT models design electrical circuits and after some back-and-forth I will reply with “Ok, now go update the whole design based on that” and when I do that I get this sinking feeling that they are going to reply “no”.

Then respond with: “Ok. I didn't read them. Keep track of whether or not I tell you I know the item and I want to ask questions about each one and the questions and answers night go into the other topics. Keep track of that. Now, what was the first point again?”

As an example that I’ve been exploring, ask this question to your favorite LLM chat interface: “What are the 30 most important facts about biology that a high school student should graduate knowing?”

I can imagine a world diverging between those who adhere to “the empty-vessel theory of education” and continue fighting against the use of LLMs in the classroom and those who embrace the use of tools that allow people to explore the world of knowledge and skills one-on-one with their private tutor.

I was adamant of the truth of that statement in a pre-LLM world.

Bolting on what is essentially just a next-token prediction engine gives us a Q&A loop that mimics a conversation with a private tutor. This property of emergence that we’ve seen with models at this size is nothing short of beautiful

Knowledge used to be locked up in expensive books. Skills were transferred via expensive schools. A lot of it is now available free on the internet. You can argue that it was kinda that way before but the wealth and abundance of educational resources everyone in the world now has access to is absurd

We’ve just lived through the collapse of the cost of an education to essentially zero and we are all just going about as if it didn’t happen.

This is a statement that I’ve been shouting at loud as possible to anyone who will listen since at least 2017 when I wrote it down in my notebook.
#education

Actually, this whole thought process came up last night when I was thinking about how much data a 400mm wafer could store in mask ROM. The answer is about 2 PB. That's 51 bits of byte-addressable data! Well, I guess our 64 bit address space might actually be useful now.

I'm assuming this isn't novel. The fabs must do stuff like this already for logic, but I don't think anyone is currently producing large mask ROMs so I guess this is unique with respect to that.

During QC we find the defects, burn the PROM table with the location of the defects and the correct data. Maybe using a XOR overlay? Each chip will be unique here. We should get a massive increase of yield.

The idea to use a small portion of the silicon as PROM came to be late last night when I was thinking about how to deal with silicon defects. The mask ROM is super dense and as long as you can route around the defects then you can still make use of it.

At some point we are going to get to the point where the need to running billions or trillions of these instances will outweigh the need for using the most up-to-date models. Not sure when that will come but it seems like someone should be preparing for this. Maybe I need to start a company...

You could bake all of DeepSeek R1 (671B params, FP16) into ~76 mm² of mask ROM at 5nm—smaller than an M4 chip. No DRAM, no loading, instant-on. Use PROM + overlay logic to patch yield defects. LLM as silicon. #llm #osh

Lesson learned: Don’t use a thermometer with an EInk screen to determine whether your pipes are going to freeze and burst. Apparently the battery died when it was 13C in the garage. 🤦‍♂️

TIL that in Python you can use map with a string join even though it's a method instead of a function:

map(''.join, s)

I guess I intuitively new that ''.join is a callable object but I didn't realize you could do this without a lambda or using functools.partial. Cool!