It's quite bizarre. The idea of each "row" of a spreadsheet being used to compute one particular thing seems like a deep and fundamental misunderstanding of spreadsheets, and a two-tier hierarchy (at the top level, each s-expression gets its own cell; below that a cell can contain an arbitrarily nested s-expression) seems like a deep and fundamental misunderstanding of s-expressions.
It is a bit bizarre, but it's also a simplified demo. I might have spent too little time to fully grasp what exactly is being demoed, but it looks like it could be easily extended to support cell references (including ranges), to change the evaluation strategy, and so on. Though it might then lose the property the author seems to care about, ie., parallelizability of the computation.
So yes, it is bizarre, but it probably is like that for a reason, since making it less bizarre would be trivial (and Emacs' calc and org-mode packages (the latter is even referenced in the post), among others, already do this). We just failed to immediately see that reason, I guess :)
The link describes increment/decrement as “one of the enhancements we made” (emphasis mine). The other things rattled off there aren’t necessarily “enhancements”, so I don’t think it’s clear that Mazor is giving an exhaustive list of changes or enhancements.
What's amusing and/or interesting is that the stuff under "Built to Be Written by Agents" also would make it easier to write by hand, I imagine. It's just better debug diagnostics; nothing really to do with agentic development specifically.
Maybe that's just marketing, but it's really intriguing how human best practices are also agent best practices; how, maybe, we've reached a point where even devs aren't recognizing agent best practices as human best practices anymore; and how suddenly best practices are being prioritized in a way they often weren't - humans could/would/were expected to suck it up and deal with poor devex, but agents simply can't. A strange sort of victory for us folks who cared about it.
A friend of mine who works a lot with AI commented that working with AI was making him develop like a grownup for the first time. Turns out that all the best practices really do carry over, but the consequence of not doing it is greater with agents.
AI has forced me to write out more planning documents. Not waterfall strategic planning, but tactical planning such as describing the desired behavior of a script before I write it, or pre-registering a data analysis and my expected findings in writing before I start querying data -- that goes in a PLAN.md within whatever folder I'm working on at the moment.
I now keep a systematic todo list that I call TASKS.md, which groups tasks into topical sections and every task has an ID, such as MD3 or EV7.
I also keep an append-only-ish journal of results, findings, unfinished thoughts, rejected ideas, etc.
All of that goes right into my project Git repo. It works great for both AI agents and humans.
But perhaps more interesting than my new workflows, is that I actually learn things from the AI. The idea to start pre-registering all my little analyses, not just the big tasks, was a Claude-ism that I've since adopted even when I'm not using AI.
I do TODO.md instead of TASKS.md but what's the benefit of giving each task an ID? Just easier for you to refer to a specific task? I usually just paste the whole line if I need to refer to it, but given it a short ID might not be a bad idea.
Yeah, the main benefit is being able to refer to specific items rather than ambiguously referring to patterns that might be misinterpreted by the model.
Yes, that. And -- in keeping with the theme -- it's just as useful for me, the human, as it is for the model. I can even cross-link them to Jira items, so I can say things like "task EV5 corresponds to Jira subtask DS-7491 within story DS-7389".
Also this is in the context of data science, so I often have a big and ever-growing list of secondary analyses I could run, and each one ends up with a small pile of source code. This actually lets me organize my code around the IDs, so I can have a directory like `analysis/EV7/` with its own little PLAN.md and README and some .sql and .py files and a .ipynb to pull together the results (and/or a small AI-generated data explorer app). And then later if I need to refer to those results I can write a comment like "parameter derived from EV7 results" or whatever.
git logs are a bit of a pain to search through and also perhaps not the place to insert long instructions, I do remember looking at some repo and just browsing the list made me quickly find what was intended during development.
waiting for the cargo lock is something else: cargo was not designed for multiple simultaneous entities developing in parallel
my take is, in the current state of the ecosystem, you should use https://github.com/kunobi-ninja/kache and make agente work on separate worktrees,each with their own target directory
usually separate target dirs means multi-gigabyte artifacts for each worktrees (and waiting for the same code to compile again). but not with kache, it builds only once and reduplicate files (with reflinks if your filesystem supports it)
there's also some projects for live reloading rust, and some specific tips about faster builds during iterative development (the most relevant being, use the mold linker)
The way Rust compile times can sneak up on you as your project grows is not to be underestimated. More so if you're running multiple agents in parallel.
But frankly I don't mind it as much with agents as I did when handwriting the code. It's on the agent to endure the slower feedback loop.
It's really "fun" how people suddenly started writing short and concise tutorials on how to use a tool or a library without too much yapping, and somehow instead of being at the beginning of the docs, it's buried deep in "skill.md" file, not intended to be seen by humans.
In my experience the agents actually read the AGENTS.md. I used to put info in the README.md but only noticed uptake when I put it in AGENTS.md instead.
Sure, it can be marketing, but I think it's more that it has become evident that the best way to get agents to write code you can trust, at a scale that actually improves productivity, usually means adopting and enforcing best practices at a scale that the industry silently really wasn't doing at all.
I've spent the last few years banging on a drum about teams installing linters, static analyzers, and similar in Laravel teams. I get nods of agreement from seniors but little willpower from managers, who just want to see the number of commits line go up.
Surprise surprise, now that I am in charge of a project and have enforced these standards, the team has done incredibly well with even just partial AI adoption in their work - because the guard rails are there and agents don't get tired of hitting them.
Have you looked at Docker's home page recently? The title is now "Secure Sandboxes for AI Agents". There is some insanity going around, rebranding basic tools as "for agents". Docker is just secure sandboxes for everyone, from that tagline you'd think they were somehow specifically made for agents.
The idea that humans never needed a concise summary of where everything is, but then agents came along and suddenly it's really useful, really bugs me. But here we are, and I'm occasionally reading AGENTS.md files like some sort of weirdo. I'm not an AI, why am I being made to read this?
On the other hand, Nextjs has been adding debug info that is feasible for agents to parse, but are basically inscrutable to human developers with limited context widows.
This post is referring to Victoria, Canada. The website, Orcas Airband, has radio traffic from a bunch of stations around the Salish Sea (Victoria, Seattle, Vancouver, Nanaimo...); orcas are like a mascot of this part of the world.
AFAICT, while there are many, many places named after Victoria, none of them are in California.
Woah! This is a fascinating TIL. I live in the area and assumed "Orcas" (Island) was pronounced like and was named after the whales. Kind of an incredible coincidence, actually.
I have no love for Nintendo's litigiousness, but the reason to believe they'd leave you alone if you refrained from hacking their latest games and consoles is that they've left alone developers who refrained from hacking their latest games and consoles. Emulators for everything before the Switch are still widely available and untouched by Nintendo. Even Yuzu, the Switch emulator Nintendo lawfare'd out of existence, lasted six years - almost entirely contemporary to the hardware - until they flew too close to the sun by soliciting donations to the tune of hundreds-of-thousands per year and encouraging piracy of TotK at launch.
Citra (3DS) shut down with Yuzu. So not exactly. And Ryujinx did not take donations nor encourage piracy of TotK at launch, yet met a similar fate.
Regardless, for previous generations included, work on emulators and hacking consoles happened when the console was still current, of course.
You can’t start the reverse engineering work only after the console is superseded, by then it is too late for preservation, especially with live service games.
Citra shut down with Yuzu because they were owned by the same company that
It's true that Nintendo hasn't always sicced its lawyers on emulator developers for current gen - I'm fondly remembering the GBA era, when emulators based on leak dev kits started popping up before the console was even released - although that really only hurts matheusmoreira's point all the more; they're clearly OK with giving devs a fair amount of leeway. As for why that didn't apply to Ryujinx, it's hard to say, though my best guess is Nintendo was trying to get ahead of the "Yuzu alternative tears of the kingdom" Google searches.
This is a minor point, but I'll mention there really weren't any real NES or SNES emulators during their lifetime and yet they're among the most well-preserved consoles. Although sadly this did mean we lost quite a bit of Satellaview content, speaking of live services.
100%. Although large handwritten software code bases are often indecipherable to the average human being (actually, almost always indecipherable to the average human being, but often indecipherable even to skilled software engineers too); sometimes due to inherent complexities, sometimes due to Enterprise-isms.
The point isn't some people not understanding whatever stuff with limited effects, it's no people understanding anything important.
Mathematics is about describing fundamental reality, so you can change it. When corporations own it exclusively, they can do whatever they want with it.
> The point isn't some people not understanding whatever stuff with limited effects, it's no people understanding anything important.
Using an LLM doesn't mean "I must write this code in a way that I cannot understand it". I'd encourage using the LLM to write code in a way that you can understand.
> When corporations own it exclusively, they can do whatever they want with it.
This is an argument in favour of open-weight models.
reply