May & June 2026 Link Roundup

Published: June 03, 2026

Updated: June 18, 2026

Updated June 18th: Added more links. (Hyaaaaaaa!)

I found this article recently when doing a search for the origin of the aphorism "code is a liability". It's a very relevant warning for any team that works with any significant quantity of LLM generated code.

A good engineering culture with emphasis on strong fundamentals and good design is as important now as ever. I recently, briefly, inherited maintenance of a codebase that demonstrated firsthand what incautiously accepting LLM suggestions (and merging them) can result in - things like a 13,500 line service class that wasn't even a year old. Much of this code base ended up very difficult to understand, extend or maintain. And to the inevitable "let the LLM maintain it" apologists, that literally did not work; Claude may have written most of the code in that 13,500 line file, but as soon as I asked it to suggest any changes, it would consistently crash the conversation with out of context errors. (Going back a message to try again hardly helps if it was your first message!)

All that to say, while LLMs can be useful, they certainly still require oversight and human judgement to write actually useful - and usable - code that will enable you to ship features at a good velocity, with low errors, over time.

On the flip side, one of the earlier projects I was involved in in my career had a very strong focus on good engineering principles. That codebase, while still containing a variety of sections and moving parts, was incredibly easy to extend and pivot on features. We did such a great job that the client extended our contract in order to train up junior developers to continue the trend; the project as a whole was a major success story for the client.

Shipping lines of code as fast as possible is a really poor metric.

Normalization of Deviance in AI

This video demonstrates a much more mature approach to using an LLM to write code than what we often seen thrown around. Pocock highlights how important it still is to have good fundamentals and architecture in your project, because failing to do so will leave the LLM lost and confused as much as it does human developers.

This article got recommended to me on LinkedIn recently. The author goes into detail at length about how the design of LLMs predisposes them to behaviors such as always finding errors (whether there are any there to find or not), and discusses how to avoid falling for the trap. Great read.

One of my favorite quotes: "The core issue is that LLMs are trained to produce plausible text, not to produce correct text. Plausibility and correctness are highly correlated in many domains, which is why LLMs are so impressive in general use. But in domains where correctness has a precise, objective definition, such as code, mathematics, or formal logic, the gap between plausibility and correctness can be significant and consequential."

The Infinite Bug Factory: Why Large Language Models Never Run Out of Errors to Find, and What You Can Do About It

Wonderful talk by Brandon Sanderson about what art really is and what it means. I love his conclusion that WE are the art, the end result is just proof of the work that was done.

In this article for the Atlantic, Ted Chiang explains why Anthropic's claims about Claude "showing signs of consciousness" (and any other similar claims about LLMs) are at best hype and at worst dishonest. He gives a few good thought experiments to demonstrate the pitfall that people generally fall into when imagining that they are, and goes on with some more in depth examination of what it would mean, ethically, if Claude were conscious (and how we can tell from Anthropic's behavior that the company almost certainly does not genuinely believe this).

No, Artifical Intelligence is Not Conscious

Gary Marcus has some great insights and gathers some really interesting data about LLMs (and has been active talking about the research of AI since long before these generative models were introduced to the general public).

In A Knockout Blow For LLMs?, Marcus addresses recent research by Apple that demonstrates that LLMs, including the new "LRM" reasoning models, cannot figure out how to complete a Tower of Hanoi puzzle with greater than 7 disks, and do not in fact "reason" about it in a logical way. That is, even when provided with the correct algorithm and how to do it, they are not capable of following the steps logically from one to the next. Meanwhile, this is trivially solved by a simple program. There are additional examples in the paper which Marcus does not discuss in detail, but excerpts a graph showing similarly poor results across a selection of models on a variety of tasks. He discusses the implications for claims of "reasoning" in some detail.

In Slop, productivity, and why the AI-fueled world is going nowhere mighty fast, Marcus goes through some interesting graphs and numbers about the relative increases in available content in a variety of mediums since the introduction of LLMs, vs whether the increased content is actually valuable in any way. (Some of the graphs have numbers to back up the latter discussion, others don't.) He also suggests that a lot of the perceived usefulness of LLM generation will vanish when prices are closer in line with costs, which I also think is quite likely.

Dave at You Suck At Programming asks "what does it mean to get left behind" and explores some possible answers for the question, and also analyzes the framing of "left behind" in terms of having goals. He then discusses his opinions on what not using LLMs means for each situation, with a general conclusion of "not at all"; the only potential exception being a job situation where a coworker is accomplishing tasks significantly more quickly with the same quality. (This hypothetical, in my opinion, is completely unrealistic both in timing and in quality.)

A bit long, in Anthropic/OpenAI may be spending more than $1000 for every $100 you pay them Wierda explores some best-guess estimates on what LLM use actually costs the providers vs what users are currently paying. It doesn't look good for IPOs when they have to actually show a profit.

This is a bit of a long video (~54 minutes) but Jevon put together some great information here. The first 20 minutes or so are pretty technical about the mechanics of LLMs and how they work relative to how adding more GPUs scales them, but later in the video he also gets into a lot of interesting data about electrical requirements, including some rough estimates of what it would mean for construction of both data centers and the power grid to scale enough to "replace all white collar workers"; some practical considerations about hallucination, accountability, and legal requirements for various jobs; and some research published by Anthropic itself about the actual effects of LLM availability on unemployment rates.

This post is tagged: