The Articulate Intern; A Quick Ramp Up On Ramping Up Quickly; Snitches Get Stitches

Two long-form pieces and one “heads up” resource to start this sweltering week (at least on the U.S. left coast and northeast).

---

TL;DR

(This is an LLM/GPT-generated summary of today’s Drop using SmolLM3-3B-8bit via MLX and a custom prompt.)

(So, it “decided” to not follow instructions and made named links vs. use raw URLs. I changed nothing config- or prompt-wise. Gotta love non-deterministic computing.)

• Jimmy Hartzell critiques LLMs as unreliable assistants, emphasizing their reliance on human oversight and their role as language mimics rather than true programmers, as detailed in LLMs in Programming.
• Ian Ireland explains SpiderMonkey’s four-tier execution model, including a baseline interpreter generated at startup, seamless tier transitions, and inline caches for dynamic languages, as discussed in A Quick Ramp-Up On Ramping Up Quickly.
• The Bloomberg article highlights Flock’s AI system, which uses travel patterns to flag individuals for police investigation, raising concerns about lack of transparency, potential bias, and mass surveillance, as reported in Surveillance Company Flock Now Uses AI to Report Us to Police.

---

The Articulate Intern

Photo by David Valentine on Unsplash

In “LLMs in Programming, Jimmy Hartzell uses a long-form storytelling format to walk us on a journey where, at first, we may think we’re about to learn how to evaluate an LLM system⁺ in a programming context, but we soon learn that is not the true intent. Hartzell merely uses said premise to discuss nature of”AI” tools, their actual capabilities versus the hype surrounding them, and how we should think about their role in programming and society more broadly.

He first shows how LLM systems work in programming by demonstrating that Claude needs existing tools like the Rust compiler to function; meaning, it’s more like a translator than a true programmer. This connects to his historical view that LLM systems represent the next step in computer interfaces, not artificial intelligence. We’ve moved from punch cards to command lines to graphical interfaces, and now to natural language interfaces.

This leads him to question what “intelligence” actually means. He argues that the Rust compiler shows more real intelligence than Claude because it truly understands program correctness, while Claude just mimics language patterns. Based on this, he suggests we should treat LLM systems as unreliable assistants that need human oversight, not as replacements for human workers.

These insights support his broader concerns about the “AI” hype cycle, and we walks us through specific concerns regarding both the financial bubble around “AI” investment and the social harm from people trusting these tools too much for advice, companionship, or critical decisions.

I realize we’ve all encountered a plethora of commentary on LLM systems, of late, but I believe Jimmy’s piece is still worth your time and also worth all of us sharing to folks with less “AI” agency.

————

⁺ I made a decision over the weekend to only use “LLM” when referring specifically to a model; we aren’t interacting with LLMs online, we’re working in a very complex and sprawling LLM-centric application.

---

A Quick Ramp Up On Ramping Up Quickly

“A quick ramp-up on ramping up quickly” is a talk about how Mozilla’s SpiderMonkey JavaScript engine uses an unusual four-tier execution system, with two of those tiers being interpreters.

Ian Ireland walks us through a classic problem JavaScript engines face: you need fast startup but also fast execution. The usual solution is tiered compilation: first, start with a slow interpreter that can run code immediately, then move hot code to increasingly sophisticated compilers that take more time but produce faster code.

SpiderMonkey’s twist is their “baseline interpreter.” Instead of writing it in C++ like a normal interpreter, they generate it at startup using the same code generation infrastructure as their baseline compiler. So, they JIT-compile an interpreter.

This sounds odd but gives them several advantages. They can reuse all the complex logic for handling JavaScript’s dynamic behavior (like how addition might be math or string concatenation depending on the types). Both tiers use the same stack layout, so transitioning between interpreted and compiled code is seamless. And they get to use inline caches in the interpreter, which is unusual.

Inline caches are an optimization technique borrowed from Smalltalk. The idea is that dynamic languages are usually boring in practice. For example, if you add two integers once, you’ll probably add integers again next time. So you can cache the fast path for common cases while still handling edge cases correctly.

The baseline interpreter performs about halfway between a traditional interpreter and their baseline compiler, which is pretty good. On-page-load benchmarks they saw up to minor but real improvements.

The broader point is about iterative problem solving. Each solution creates new constraints that drive further innovation. The imaginary website keeps complaining about different aspects starting with speed, then compile time, then memory usage. Each complaint pushes the engine design forward.

I share the talk as it is a solid example of questioning assumptions, and it’ll provide some brief moments away from all that’s going on around us IRL.

---

Snitches Get Stitches

For whatever reason I failed to read this heads up back in late July (it was in both the Inoreader and Raindrop queues, but I guess I decided to not doom-read for a couple days). I found it due to a new post by Bloomberg.

Flock is a police surveillance company that tracks license plates (U.S.) nationwide.

Well, it was that.

It’s now an “AI” system that actively tells police who to investigate based on travel patterns. The company’s algorithms now flag people for “suspicious” movement across states, track who travels together, and identify patterns the “AI” thinks indicate criminal activity. There’s no transparency about how these decisions get made, no oversight of a private company making these calls, and obvious potential for bias. The ACLU, and others, are concerned because it’s mass surveillance that targets folks before they’ve done anything wrong, purely based on algorithmic suspicion.

Essentially, now, [y]our driving habits can now get us flagged as persons of interest.

---

FIN

Remember, you can follow and interact with the full text of The Daily Drop’s free posts on:

• 🐘 Mastodon via @dailydrop.hrbrmstr.dev@dailydrop.hrbrmstr.dev
• 🦋 Bluesky via https://bsky.app/profile/dailydrop.hrbrmstr.dev.web.brid.gy

☮️