mbbutler

False-positive rate is so high with Mythos according to friends and other reporting I have seen.

The original Mythos release used ASan to filter false-positives so it was able to maintain a good FPR, but when Mythos moves into domains that don't have a readily available oracle to help filter hits, the result is a deluge of false bullshit.

Newport's book on this topic is terribly underwhelming, and I say that as someone who has really enjoyed his other books, blog posts, and YouTube videos. Most of the anecdotes he gives to support his "Career Capital over Passion" have an ambiguous directionality of the causal arrow. For example, the fact that the happiest admin assistants have been at the job the longest does not mean that getting good at being an AA makes you happy with the work. There is an equally plausible explanation that enjoying the work of an AA makes one likely to stay in the job longer. Most of his examples in the book are like this.

The place where I think Newport flounders in this area is that, in order to get "So Good They Can't Ignore You", you actually have to be able to put in the time and effort to get good. And the vast majority of people do not possess the self-control and willpower to force themselves to do something they dislike to the level that is necessary to achieve said mastery.

Yeah, LLMs suck at named lifetimes. The number of times I have seen Claude reach for indices and clones instead of just using proper named lifetimes is too many to count at this point. Not great for high-performance code!

No, you don't understand. LLMs will never be capable of knowing what questions to ask, only how to ask the questions. /s

It would be helpful to add in some cases that do not contain any vulnerabilities to assess false-positive rate as well.

In my personal experience, the rate at which Claude Code produces suboptimal Rust is way higher than 1%.

But Eon's tagline is "Solving brain emulation as an engineering sprint, not a decades-long research program"! How could they have ever gone wrong?

I especially love the quadratic fit, chosen with no justification, that brings the US within the uncertainty envelope in the second plot. Also notice how much work the Mexico and USA data points are doing to the previous linear model fit. Oh, my high leverage data point can't be an outlier because it's within uncertainty when I fit the data with the potential outlier included. This is basic linear model validation stuff.

> Has a hard to explain fixation on doing things a certain way, e.g. always wants to use panics on errors (panic!, unreachable!, .expect etc) or wants to do type erasure with Box<dyn Any> as if that was the most idiomatic and desirable way of doing things

Yes! I see this constantly. I have a Rust guide that Claude adheres to maybe 50% of the time. It also loves to allocate despite my guide having a whole section about different ways to avoid allocations.

Two use-cases recently where Claude sucked for me:

1. Performance-critical code to featurize byte slices for use in a ML model. Claude kept trying to take multiple passes over the slice when the featurization can obviously be done in one. After I finally got it to do the featurization in one pass it was double-counting some bytes but not others (double counting all of them would have been fine since the feature vector gets normalized). Overall it was just very frustrating because this should have been straight-forward and instead it was dogshit.

2. Performance-critical code that iterates over lines of text and possibly applies transformations, similar to sed. Claude kept trying to allocate new Strings inside of the hot-loop for lines that were not transformed. When I told it to use Cow<'a, str> instead so that the untransformed lines, which make up the majority of processed lines, would not need a new allocation, Claude completely fucked up the named lifetimes. Importantly, my CLAUDE.md already tells Claude to use copy-on-write types to reduce allocations whenever possible. The agent just ignored it, which is _the_ issue with LLMs: they're non-deterministic and any guidance you provide is ultimately just a suggestion.

It's not just assuming that everyone learns the same way. It's assuming that everyone learns the way that all of the research literature on learning claims does not work.

Learning requires active recall/synthesis. Looking at solved examples instead of working them yourself does not suffice in math, physics, chemistry, or CS, but somehow it is supposed to work in this situation?

Machine learning people use "tensor" to just mean an N-dimensional array of numbers. The term is divorced from its meaning in Physics and Mathematics, which caused me some confusion when I started looked at machine learning papers coming from physics.

But the energy transported to Earth from your space power plant still creates waste heat when it is used to do work (and also when it is transported to earth). You cannot beat the second law.

They don't even decouple at high material standards of living. Recent increases to GDP produced emissions too but those new emissions were offset by reductions in emissions of existing industries.

This "decoupling" gets us basically nothing because it's not like we can just stop emissions tomorrow since GDP and emissions are "decoupled".