Frederick Brooks in his essay "No Silver Bullet" (included in the collection Mythical Man Month) talked about the conventions of software development and I recall had called for taking an iterative approach to software development similar to what I had followed for the Automunge project, I went into a little more detail about that in my 2019 essay of the same name:
https://medium.com/automunge/no-silver-bullet-95c77bc4bde1
A few comments:
- It has long been known in other settings that a small number of points can impact performance of different conventions, this could perhaps be considered a validation of relevance towards the largest scales
- I wonder if the reverse could be considered true, if such a small scale of data included in a training corpus can impact the model performance in a negative direction, could that same amount of data impact a model in the positive direction?
- I think this is suggestive that there remains benefit to more authoritative sources of data aggregators, like respected publishers, journals, libraries, whereby inclusion of data in such more respected repositories can be considered validation of reliability for training.
I hereby present Book 7 of the essay series from the Diaries of John Henry, "Inevitable". I used to publish these every year, took a little more time on this book featuring a self-authored collection of essays published over last three years around subjects like AI, technology policy, investing, and entrepreneurship. One of inspirations for the journey arose from essays of Paul Graham so sharing to the Y-Combinator community as a courtesy. No language models were used or harmed in the making of these essays (unless otherwise disclosed). Cheers!
- Nicholas Teague, 9/17/2025
Part of the problem of this form of benchmarking is that in some domains we wouldn't only be interested in the percent of times that an error channel is successfully mitigated, we would also be interested in the distribution of types of errors for cases where an error channel isn't successfully mitigated. The paper appears to be silent on that matter.
Thank God someone other than Scott Aaronson is publishing this kind of content. Shetl Optimized is too much of a vanity project to serve as a mainstream resource.
Joplin really laid the stage for the emergence of jazz music. The ragtime conventions are deceptively simple, they carry a kind of counter balance between the cleffs almost resembling polka at times, but the rhythmic styles often are more subtle with off beat progressions. The song structures progressing through chapters of themes are particularly interesting and lacking from most modern popular music.
The book is not just about the ideas, it is about the tapestry of how it was presented. It was a new form of literature. The whole document was its own strange loop.
Hi there, author here, this collection marks the sixth year of writing essays on themes like machine learning, entrepreneurship, and quantum computing. Through the project we've published a python library, had workshop papers at research conferences like ICLR and NeurIPS, and yeah just had a lot of fun. Some thanks owed to the Hacker News feed for fueling us with diversions over the years. Never quite made it into YC but you know what I think the essays may have turned out better as a result. Cheers.
You can't evaluate rationality in context of small number of samples of a fat tailed distribution with wide uncertainty bands. Besides, startups can be a kind of success even in failure based on opportunities that arise as a result.
The funny thing is that entrepreneurship doesn't even have to require dropping out of school. A three month summer break is easily enough time to start digging into a niche and validating the market. It probably won't get to sustainability, but by the end of a summer of focus you may at least know if there is something there worth pursuing. A better investment of time than backpacking across Europe.
They cited some of LeCun's solo work from the 80's on related matters, note that he also collaborated with Geoffrey Hinton on a related paper from NIPS 89 in which they used random perturbations to derive a gradient signal. One of the limitations of that approach was that it required each layer to have fewer units than the preceding one.
I cited this in a recent paper because it was representative of different cases for stochastic injections in neural networks. Interesting to see that similar lines of inquiry are continuing to this day.
The root enabling factor was not only the reinforcement learning, it was the simulation. Put differently, any environment that you can simulate within sufficient accuracy, you can derive robust and efficient control by reinforcement learning.
The cool thing about math applications is just how easy it would be to generate synthetic data. That these large language models haven't attempted to supplement their gigabytes+ scale data sets with such is an oversight.
- design certification treated as "just another 737" which ignored fundamental deviations in hardware and software, complicated by safety certification by manufacturer instead of public employee
- hardware design including a "dynamic instability" in which
airplane approaching an aerodynamic stall had a tendency to go further into the stall due to lift produced by the oversized engines at high angles of attack, which was intended to be mitigated with software
- omission of using multiple inputs, including the opposite angle-of-attack sensor, in the computer's determination of an impending stall
- a changed philosophy about human/machine interaction from humans winning a battle of the wills every time to computers winning a battle of wills in cases of envelope protection
The final item is perhaps the one most fundamental to other cases of safety critical human machine interaction designs. If we are signing over agency to machines for envelope protection, that means we need in advance to understand every potential edge case scenario where that envelope may be mis-framed. Such comprehensive foresight in some environments may be intractable. For the 737 this was exasperated by the presence of an inherent source of instability originating from hardware design.
Based on wikipedia it appears the original definition of perceptron was associated with the specific case of a neuron with a heavy-side step function activation output (which maps aggregated weights to either 0 or 1). I've generally adhered to referring to neural networks nodes in my writings as "neurons", which I believe is a catchall for any type of activation output, although have seen others use the term "perceptron" in the same usage in modern practice.
The biggest asset of Facebook for now is the facebook platform, perhaps a more durable asset will be the social graph sourced by the facebook platform. What would it mean for a durable social graph commercialization model? I'm not sure, but at a minimum it would need to be exportable and portable between platforms and ecosystems.