Alex Becker's Blog

A branching tech tree of the Transformer: the path of innovations from 2012 to 2017 - AlexNet, Word2vec, Attention, ResNet and more - drawn as glowing nodes on a skill-tree diagram, each one unlocking the next on the way to the architecture behind modern…

How I've come to terms with LLM non-determinism in software and abandoned the comforting lies we've been telling ourselves about traditional software.

The feeling of poverty amidst wealth stems from less opportunity to realize surplus.

How Steve Sailer wormed his way into Silicon Valley's hivemind through Slate Star Codex.

Claude writes all my code now, but the bigger shift is in *what* code I produce.

Prompt injection requires training LLMs with robust "untrusted content" primitives, viewing it as a solvable cybersecurity problem distinct from jailbreaking.

It is in startups' interest to pay employees cash and save their equity to sell to VCs.

Silicon Valley backs Trump for "bold action," because America's problems require it.

Why do rich Americans feel poor?

I built a new prompt injection benchmark for AI agents because the existing ones are not great.

USENIX Security '25 roundup on LLM security topics: prompt injection, data leaks, CAPTCHA bypasses, and model misbehavior.

The Chinese Room argument is flawed because it ignores the practical speed and scale of computation, misguiding our intuition about AI intelligence.

Prompt injection, where untrusted input overrides LLM system instructions, is a hard problem with no reliable defenses yet.

LLMs display "generalized intelligence," yet research lacks a theoretical explanation for their capabilities, how they work, or their inherent limitations, and there is no clear path to fixing this.

AI's economic future is debated.

Writing stuff down about the Transformer architecture and training methods to solidify my own understanding.

Democrats lost due to inflation & immigration.

Israel holds nearly all power in the conflict.

Satoshi's Bitcoin, intended for freedom, has failed to realize Finney's ambitions and instead empowered global crime, scams, and human trafficking.

AI will transform work, replacing some jobs & enhancing others, but its crucial impact lies in the "signaling problem," where AI-generated content devalues human competence.

"Nuclear" bombs aren't all equal.

How to project SaaS user growth?

Blockchains promise "trustless" transactions, but they actually require new forms of trust in value and network security, making them vulnerable to major attacks.

Electronic Health Records (EHRs) harm patient care & raise costs due to bad design, lack of interoperability, and government mandates.

Debugging a front-end Firefox bug: A double-click race condition caused `mousedown` to overwrite an element's reference before `mouseup` could remove it.

Mathematicians often fail to share the human intuitions behind their work, leading to a profound loss of understanding, exemplified by Mirzakhani's death.

Fight back against pervasive surveillance & exploitative algorithms: lie about your data to disrupt models that stereotype, entrench poverty, and reinforce corporate power.

We often think in terms of the mean even when this is ill-defined.

Violent protests against Milo Yiannopoulos ironically did his work for him.

Grey goo scenarios obscure the more imminent and likely risks: military AI, inequality-fueling tech, and other powerful tools society already struggles to control.

An individual vote has never tipped an important election, so why vote?

Designers often misuse opacity for text, leading to unreadable stacking effects and CSS bugs like new stacking contexts.

Why is HTTPS adoption so slow?

This post argues Django is dangerous due to fundamental flaws in data integrity and dynamic templating that worsen as projects scale.