jfreds

So they’re watermarking requests according to your environment variables and maybe changing a string format if you’re in a certain time zone? Am I missing something here? Where’s the five alarm fire?

The worst case of these were the few companies that set up token leaderboards, which is perhaps the dumbest way possible to encourage learning how to use LLMs well

My company does something dumber now. A leaderboard of how many lines of code you shipped, weighted by how complex they were (assigned by a heuristic). You can imagine the incentive this creates. I wish we just measured tokens

I’m confused: is it just markdown files in git? Or does the hybrid graph+semantic layer matter? If the latter is true, the title is just clickbait right

I tried semaglutide and while it was effective for losing weight, it made working out impossible (felt exhausted and very sick as soon as hr went up) and it made hangovers awful. Is retatrutide any different?

To be fair this was a thing before AI as well…

I’ve seen loads of very bad/wrong code written by humans

I agree in part, but I think AI does meaningfully make it harder for leadership to detect their bullshit.

The author of stringzilla, Ash Vardanian, has a bunch of really cool talks/lectures/demos. Highly recommend checking him out if you like that kind of thing :)

I don’t think OP’s point has anything to do with AI companions.

The big benefit of moving compute to edge devices is to distribute the inference load on the grid. Powering and cooling phones is a lot easier than powering and cooling a datacenter

My issue with this is that a simple design can set you up for failure if you don’t foresee and account for future requirements.

Every abstraction adds some complexity. So maybe the PoC skips all abstractions. Then we need to add a variant to something. Well, a single if/else is simpler than an abstract base class with two concrete implementations. Adding the 3rd as another if clause is simpler than refactoring all of them to an ABC structure. And so on.

“Simple” is relative. Investing in a little complexity now can save your ass later. Weighing this decision takes skill and experience

I am seeing this at my work right now. They are about to start using token consumption as _part_ of the performance review process. Obviously this is a coarse and problematic proxy for productivity.

OTOH, it’s an attempt to address a real problem. There are people who are in fact falling behind (I’m talking literally editing code in notepad), and we can either let them get PIPped eventually, or try to bring them along. There is a real “activation energy” required to learning new tools, and some people need an excuse/permission. Not saying that token count is a GOOD signal, but I haven’t heard many better ideas

Very bad hire. I’ve gently said as much to my manager and skip. But for some reason hiring is hard and firing is hard, and we’re a small team, so I’ve been told to just lower my standards. Yeah, I know

AI pull request descriptions are my current pet peeve. The ones I have seen are verbose and filled with meaningless fluff words (“optimized”, “performant” for what? In terms of what?), they leak details about the CoT that didn’t make it into the final solution (“removed the SQLite implementation” what SQLite implementation? There isn’t one on main…), and are devoid of context about _why_ the work is even being done, what alternatives were considered etc.

My first round of code review has become a back and forth with the original author just asking them questions about their description, before I even bother to look at code. At first I decided I’d just be a stick in the mud and juniors would learn to get it right the first time, but it turns out I’m just burning myself out due to spite instead.

I burnt out helping a junior on my team for the past few months. It was just terribly obvious she was feeding my responses directly into a chatbot to fix instead of actually understanding the issue. I can’t really even blame her, there isn’t much incentive to actually learn

I think this is well put. Cohesive philosophy, even if flawed, is a lot easier to work with than a patchwork of out-of-context “best practices” strewn together by an LLM

Literally yes

Context management, plan mode versus agent mode, skills vs system prompt, all make a huge difference and all take some time to build intuition around.

Not all that hard to learn, but waiting for things to settle down assumes things are going to settle down. Are they? When?

Strawman. It’s entirely possible for two things to be true at once: border laws are worth enforcing, and the current approach of flooding ICE with untrained goons explicitly targeted with white supremacist recruiting material is not going to end well

You kinda can, just don’t make a Twitter account, don’t buy teslas, don’t use grok. Tell your friends

The crux is in the “and it gets implemented” part. Teams have a limited bandwidth, so what gets implemented absolutely is a zero sum game, that’s why backlog prioritization exists. In order for your idea to get implemented, you have to advocate for it, and convince others to do the same. Writing great code and delivering useful side-projects can make you a 2-5x engineer. If you want to be a 10x engineer, you have to scale your impact beyond what you can do alone

Edit: maybe your great idea is actually something that you can implement on your own, such as a test suite or a tool. You still need to change other peoples behavior. You need to convince people to try it.

good work speaks for itself

I disagree. Ideas don’t speak, and work doesn’t speak. People do. Being a 10x engineer isn’t just about having great ideas, it’s about having great impact.

Sometimes I hear ICs say with some pride that “I’m not interested in playing office politics”. I promise you they will lose out to the engineers who are able to self advocate, and coalition-build.

Inviting a reporter from the Atlantic to your signal chat where you coordinate military plans has nothing to do with government being too big

This is true. What are the implications of that?

I get this at work, frequently.

“Oh, cursor wrote that.”

If it made it into your pull request, YOU wrote it, and it it’ll be part of your performance review. Cursor doesn’t have a performance review. Simple as

automated review tools that publish comments without human review are not allowed

This seems like a curious choice. At my company we have both Gemini and cursor (I’m not sure which model under the hood on that) review agents available. Both frequently raise legitimate points. Im sure they’re abusable, I just haven’t seen it

Citation? Not saying you’re wrong but my time in defense left me very much with the opposite opinion (radar target acquisitions had to be approved by a human, always)

Your points about misquotes and language troubles are very valid and interesting. But a word of caution on your prompt: you’re asking a lot of the word “analyze” here; if the LLM responded that the thread had 15 comments by 10 unique authors, and a total of 2000 characters, I would classify that as a completely satisfactory answer (assuming the figures were correct) based on the query

Yeah this seems like an obvious solution, which axon ought to be on board with since it protects them.

When juniors use the excuse “oh Claude wrote that” in a PR, I tell them if the PR has their name on it, they wrote it - and their PRs are part of their performance review. This is no different

Short answer is yes. Not only is it more token-friendly and potentially lower latency, it also prevents weird context issues like forgetting Rules, compacting your conversation and missing relevant details, etc.

Having gone though exactly this exercise recently comparing a homegrown vector db against Qdrant, I’m wholeheartedly in agreement that getting a working solution FAST, and then spending a decent amount of time interrogating it (with help of LLM), is my favorite learning pattern