Why Did They Build This - CourtneyR.dev's 2nd Brain

Why Did They Build This? ![rw-book-cover](https://img.youtube.com/vi/RnxTTF8x-24/maxresdefault.jpg) 🔗 **Source:** [Why Did They Build This?](https://www.youtube.com/watch?v=RnxTTF8x-24) > [!note]+ Document Notes > tags: [programming-languages, ai-agents, language-design, vercel-zero, developer-tooling] > author: The Primagen > source: YouTube > reference: https://www.youtube.com/watch?v=RnxTTF8x-24 > date created: {{date:YYYY-MM-DD}} {{time:HH:mm}} > > # 🏷️ Title: Why Did They Build This? >  ## 🧭 Table of Contents - [📝 Summary](#-summary) - [📖 Main Content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action Items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future Research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning Styles Assessment](#-learning-styles-assessment) - [📊 Accessibility Assessment](#-accessibility-assessment) - [🌍 DEIB Assessment](#-deib-assessment) - [✅ Logical Validity](#-logical-validity) - [🛡️ DARVO Analysis](#-darvo-analysis) - [🎯 Conclusion](#-conclusion) ## 📝 Summary - Vercel released Zero, a new programming language claiming to be "made for agents" - The Primagen analyzes Zero and finds it remarkably similar to existing languages like Go and Zig - Key criticism: no evidence that Zero is actually better for AI agents than existing languages - The language lacks novel features and may actually be worse for agents due to limited training data - Questions the fundamental premise of designing programming languages specifically for LLMs ## 📖 Main Content Vercel recently announced Zero, billing it as "the programming language for agents." The language promises features like explicit capabilities, JSON diagnostics, type-safe fixes, and agent-first design. However, upon closer examination, Zero appears to offer little innovation beyond existing systems programming languages. ### Language Design Analysis Zero's syntax and features closely mirror existing languages: - Error handling similar to Zig's `check` mechanism - Memory operations reminiscent of Zig's explicit memory management - Standard library approach similar to Go's comprehensive built-ins - Type system with "raises" annotations for error propagation The claimed "agent-first" benefits are unclear: - **Explicit capabilities**: Most programming languages already have explicit capabilities - **JSON diagnostics**: JSON formatting doesn't inherently improve LLM comprehension - **Type-safe fixes**: Standard feature of typed languages - **Deterministic tooling**: Unclear what makes this different from existing deterministic tools ### Agent Suitability Concerns Several factors suggest Zero may actually be worse for AI agents: - **Limited training data**: As a new language, LLMs have no existing examples to learn from - **Verbose syntax**: More tokens required compared to established languages - **Context pollution**: Need to provide language documentation reduces available context window Established languages like TypeScript and Rust offer better agent experiences due to extensive training examples and mature toolchains. ## 🗒️ Notes #### ⏱️ [00:00:00](https://www.youtube.com/watch?v=RnxTTF8x-24?t=0) Initial Reaction to Zero - Programming language enthusiast's confusion about Zero's purpose - Marketing claims about "made for agents" unclear #### ⏱️ [00:02:30](https://www.youtube.com/watch?v=RnxTTF8x-24?t=150) Language Goals Analysis - Agent-first learnability through small, regular language surface - Standard library depth for common capabilities - Claims sound identical to Go's design philosophy #### ⏱️ [00:04:45](https://www.youtube.com/watch?v=RnxTTF8x-24?t=285) Syntax Examination - No GitHub syntax highlighting available yet - Error handling with `check` keyword similar to Zig - Mixed consistency in error handling patterns #### ⏱️ [00:07:20](https://www.youtube.com/watch?v=RnxTTF8x-24?t=440) Implementation Quality Issues - 15 nearly identical vector implementations in codebase - Repeated code patterns across different modules - Backend implementations with inconsistent patterns #### ⏱️ [00:10:15](https://www.youtube.com/watch?v=RnxTTF8x-24?t=615) Agent-First Claims Debunked - No clear advantages for AI agents over existing languages - Verbose syntax increases token usage - Novel language requires more context explanation #### ⏱️ [00:12:30](https://www.youtube.com/watch?v=RnxTTF8x-24?t=750) Community Response - Developers questioning lack of backing research - Concerns about ecosystem and training data - Skepticism about claimed benefits #### ⏱️ [00:14:00](https://www.youtube.com/watch?v=RnxTTF8x-24?t=840) Ginger Bill's Critique - Language design for LLMs compared to "horoscopes" - No understanding of what to optimize for in AI models - Question why new language would perform better than existing ones ## 📚 References - [Zero Programming Language Repository](https://github.com/vercel/zero) - [Vercel Zero Announcement](https://vercel.com/blog/zero) - [Zig Programming Language](https://ziglang.org/) - [Go Programming Language](https://golang.org/) ## 🗣️ Paraphrase A experienced developer examines Vercel's new Zero programming language, which claims to be designed specifically for AI agents. Despite marketing promises of agent-first features, the analysis reveals Zero to be largely derivative of existing languages like Go and Zig, with no clear advantages for AI development. The verbose syntax, lack of training data, and absence of novel features actually make it less suitable for agents than established alternatives like TypeScript or Rust. ## 📰 Newsletter 🚀 **New Language Alert**: Vercel just dropped Zero, claiming it's "made for agents." But is it really? Our deep dive reveals it looks suspiciously like Go wearing a Zig costume. While we're all for innovation in developer tooling, the evidence for agent-first benefits remains thin. Maybe stick with TypeScript for your AI projects until we see some actual benchmarks? 🤖💭 ## ✅ Action Items #### DevRel - Research and document actual performance differences between languages for LLM code generation - Create educational content comparing established vs. novel languages for AI development - Develop guidelines for evaluating "AI-first" developer tools claims #### Community - Encourage evidence-based evaluation of new programming languages - Foster discussion about what truly makes a language suitable for AI agents - Share experiences with different languages in AI development contexts ## 🧪 Hypotheses - New programming languages designed for AI agents offer no measurable advantages over existing languages with extensive training data - Verbose syntax in "agent-first" languages actually reduces AI effectiveness by consuming more tokens - Language familiarity in training data is the primary factor determining AI code generation quality - Marketing claims about AI-optimized tools often lack empirical backing - Established languages with strong type systems (TypeScript, Rust) remain superior for AI-assisted development ### 📊 Methodology - Comparative analysis of code generation quality across different programming languages - Token usage studies comparing Zero syntax to equivalent Go/Zig implementations - Survey of AI model performance on established vs. novel programming languages - Empirical testing of LLM code generation accuracy with varying amounts of language documentation in context ### ❓ Questions - What empirical evidence supports claims of "agent-first" language design? - How does training data volume affect LLM performance across programming languages? - What language features genuinely improve AI code generation quality? - Why create a new language instead of improving tooling for existing ones? ### 🔮 Future Research - Longitudinal study of Zero adoption and AI agent performance - Analysis of language design principles that actually benefit AI systems - Investigation of whether specialized AI programming languages emerge as a viable category - Research into optimal programming language characteristics for human-AI collaboration ### 💡 Implications - **Developer Education**: Need for critical evaluation skills when assessing AI-focused tools - **Language Design**: Question whether AI optimization requires fundamentally new languages - **Tool Selection**: Established languages may remain preferable for AI development - **Marketing Transparency**: Claims about AI-optimized tools require empirical validation ### 🧠 Learning Styles Assessment Reading level estimate: Flesch-Kincaid Grade 10-12 #### 👀 Visual learners - Present: Code syntax comparisons, repository structure screenshots - Improve: Side-by-side language feature diagrams, visual token usage comparisons #### 🔊 Aural learners - Present: Video format with spoken analysis and commentary - Improve: Podcast version with key technical concepts explained verbally #### 🗣️ Verbal learners - Present: Clear explanations of programming concepts, analogies to existing languages - Improve: Glossary of technical terms, more analogies comparing language features #### 🧩 Kinesthetic learners - Present: Code examples that can be compiled and tested - Improve: Interactive playground for comparing Zero vs. other languages #### 🧠 Logical learners - Present: Systematic comparison of language features, logical critique structure - Improve: Feature comparison matrices, decision trees for language selection #### 🧑‍🤝‍🧑 Social learners - Present: Community responses and developer discussions - Improve: Forum for developers to share experiences with different AI-coding languages #### 🧘‍♀️ Solitary learners - Present: In-depth technical analysis suitable for individual study - Improve: Self-assessment questions about language design principles #### 🌿 Naturalistic learners - Present: Evolution of programming languages as ecosystem development - Improve: Connections to broader patterns in technology adoption and tool evolution ### 📊 Accessibility Assessment #### Screen reader semantics - Present: Clear heading structure, descriptive text - Improve: Alt text for code screenshots, more descriptive timestamps #### Keyboard navigation - Present: Video player with standard keyboard controls - Improve: Transcript with jump-to-time functionality #### Contrast and readability - Present: High contrast video with clear audio - Improve: Transcript formatting with better visual hierarchy #### Media and captions - Present: Full transcript provided - Improve: Synchronized captions in video, downloadable transcript ### 🌍 DEIB Assessment #### Cultural breadth - Present: Universal programming concepts accessible globally - Improve: Examples from diverse programming communities, international perspectives #### Inclusive language - Present: Technical focus without exclusionary jargon - Improve: Define technical terms more consistently, avoid gaming culture references #### Neurodiversity support - Present: Clear structure with distinct sections - Improve: More frequent section summaries, consistent formatting patterns #### Representation - Present: Focus on technical merit over personality - Improve: Include diverse voices in programming language evaluation ### ✅ Logical Validity #### Strong reasoning - Comparative analysis using concrete examples from Zero's codebase - Evidence-based critique focusing on observable language characteristics - Clear logical progression from claims to evidence to conclusions #### Fallacies - **Appeal to authority**: Citing Ginger Bill's opinion without additional evidence - should include more expert perspectives - **Hasty generalization**: Concluding Zero offers no benefits based on limited examination - needs more comprehensive analysis #### Tone of voice - Conversational and skeptical, with humor and personal reactions - Technical authority demonstrated through language feature analysis - Constructive criticism rather than dismissive rejection ### 🛡️ DARVO Analysis No DARVO patterns detected. The analysis focuses on technical critique of a programming language rather than interpersonal conflict or accusation scenarios. ## 🎯 Conclusion Zero represents an interesting attempt at creating an AI-optimized programming language, but fails to demonstrate clear advantages over existing alternatives. The language's similarity to Go and Zig, combined with verbose syntax and lack of training data, suggests that established languages remain superior for AI-assisted development. The broader question of whether specialized "agent-first" languages are necessary remains unanswered, pending empirical research on what truly benefits AI code generation. ## 🧱 Refactor Appendix ### 📚 Literature Split Pad ### Zero offers no clear advantages over existing programming languages > [!quote] [Why Did They Build This?](https://www.youtube.com/watch?v=RnxTTF8x-24) by [[The Primagen]] via [[YouTube]]. Published 2024. Accessed {{date:YYYY-MM-DD}}. > "There's nothing here that's interesting. There's nothing here that I see as being agent first." Analysis of Zero's syntax reveals strong similarity to existing languages without novel agent-optimized features. [view highlight](https://www.youtube.com/watch?v=RnxTTF8x-24#t=285) #xref #articleseed ID:rw_001 ### Novel programming languages face adoption challenges with AI systems > [!quote] [Why Did They Build This?](https://www.youtube.com/watch?v=RnxTTF8x-24) > "A brand new language that no LLM has ever heard about or knows how to write. Brilliant." Limited training data for new languages creates context pollution and reduces AI effectiveness compared to established alternatives. [view highlight](https://www.youtube.com/watch?v=RnxTTF8x-24#t=615) #xref ID:rw_002 ### Language design for AI agents lacks empirical foundation > [!quote] [Why Did They Build This?](https://www.youtube.com/watch?v=RnxTTF8x-24) > "People who are designing languages for LLMs are doing the equivalent of horoscopes." Claims about AI-optimized programming languages often lack scientific backing or clear optimization targets. [view highlight](https://www.youtube.com/watch?v=RnxTTF8x-24#t=840) #xref ID:rw_003 --- ### 🔖 Atomic Split Pad (Zettel Candidates) #### Training data volume determines AI programming language effectiveness - Scope: LLM code generation across different programming languages - Evidence from TypeScript vs. Zero comparison at 10:15 - Links: [[AI-Code-Generation]] [[Language-Adoption-Patterns]] #### Verbose syntax reduces AI agent efficiency through token consumption - Scope: Context window limitations in LLM programming tasks - Evidence from Zero syntax analysis showing increased verbosity - Links: [[Token-Optimization]] [[AI-Context-Windows]] #### Marketing claims about AI-optimized tools require empirical validation - Scope: Developer tool evaluation and evidence-based assessment - Pattern observed across multiple "AI-first" tool launches - Links: [[Tool-Evaluation-Methods]] [[Evidence-Based-Development]] --- ### ✏️ Fleeting Queue (Extract by Selection) - Research comparative studies of programming language effectiveness with LLMs — context/video analysis #fleeting #inbox - Investigate what language features actually benefit AI code generation — context/empirical research gap #fleeting #inbox - Document patterns in "AI-first" tool marketing vs. actual capabilities — context/critical evaluation #fleeting #inbox --- ## Full Transcript [00:00] You know, I've always loved programming [00:01] languages. I have looked into many and [00:04] I've found love in so many different [00:06] flavors. And so when today I saw a brand [00:09] new one dropping from Vercel, you could [00:12] imagine I was a bit curious. I was [00:14] piqued, okay? My hams, they were salted. [00:17] I was ready. Introducing Zero, the [00:19] programming language for agents. Now, [00:22] I'm not going to lie to you. I [00:24] immediately don't know what that means. [00:26] I always thought English was the new [00:27] programming language for 2026. AM I [00:30] RIGHT, BOYS? I wanted a system language [00:33] that was faster, smaller, and easier for [00:35] agents to use and repair. Explicit [00:38] capabilities, JSON diagnostics, type [00:41] safe fixes, made for agents on day zero. [00:46] I'm just going to be real with you guys. [00:48] I effectively don't know what any of [00:50] that means. Now, I've been programming a [00:52] long time and it's kind of embarrassing. [00:54] I I I I think I can understand what JSON [00:57] diagnostics mean, but explicit [01:00] capabilities? What doesn't that mean? I [01:02] I think a lot of programming languages [01:04] have explicit capabilities. Type safe [01:07] fixes? There's I assume that means you [01:09] have a type safe language. And no, this [01:12] is not some sort of April Fools' joke. [01:14] It's right here. There's an entire [01:15] repository filled with zero examples. I [01:19] looked a little bit at the compiler and [01:21] I'm not a compiler engineer, but yikes. [01:24] I was able to get this bad boy to [01:26] compile and I got a hello from zero. So, [01:29] we actually just have to talk about this [01:30] because honestly, this has to be one of [01:33] the times in my life in which I am just [01:36] utterly confused. I don't I I truly [01:38] don't understand the purpose of this at [01:40] all. But first, [01:42] the bag, of course. Hey, you. Do you [01:45] like getting something better for less [01:47] money? Well, let me tell you about [01:48] today's sponsor. Blacksmith is the [01:50] fastest way to run your GitHub actions. [01:53] Two x faster job runs, four x faster on [01:56] caches, and 40x faster on Docker builds, [02:00] slashing costs up to 75% like that is an [02:04] obvious win. Blacksmith is already [02:06] trusted by over 3,500 companies [02:08] including Ashby, Veed, Clerk, and [02:11] Mitrefinch. Check the link in the [02:13] description to see how you can save a [02:14] boatload of money and have an excellent [02:16] action runner. Okay, welcome back. What [02:19] Zero is aiming for? Agent-first [02:22] learnability. A small regular language [02:24] surface that agents can pick up quickly [02:27] from examples, docs, and compiler [02:29] feedback. Standard library depth. Common [02:31] capabilities should live in documented [02:33] coherent library APIs instead of [02:35] scattered dependency stacks. Okay, I I [02:37] like that. That sounds good. [02:40] Deterministic tooling. I have never [02:43] heard in my lifetime of undeterministic [02:46] tooling. I have to say, this is a first [02:48] for me. Direct developer experience. [02:51] Checking, running, formatting, [02:52] inspecting, and repairing code should be [02:54] fast, copyable, and scriptable. I [02:56] believe that is pretty much programming [02:59] in any language but Rust. Regularity [03:01] over syntax. Prefer one obvious way to [03:04] express most things even when that makes [03:06] code more explicit than a human might [03:08] choose in another language. You know [03:10] what? You know what all these mean, [03:12] right? You know you know what he's [03:13] saying, right? Call Go. Like bro's [03:16] describing Go almost word for word. A, [03:19] Go's extremely limited in what you can [03:22] do. B, there's pretty much just one way [03:24] to do anything. Three, it's extremely [03:27] explicit. It is very very simple. You [03:30] got an error? You got an if error does [03:32] not equal nil handling case. Oh, you [03:35] want a complete tool set in which the [03:37] standard library effectively covers most [03:39] uses of the language itself? Again, [03:42] that's Go. Deterministic tools. Bro, [03:45] still don't know what that means but Go [03:47] Go is just jam-packed with all of it, [03:49] okay? Okay, so this is the actual [03:51] language that they're doing. Notice that [03:52] there's no syntax highlighting. Hey, we [03:54] haven't got there yet. GitHub, come on. [03:56] Can you get with it? This is a new [03:57] language that's been out for days. I [03:58] kind of expect you to already have some [04:00] syntax highlighters on the server able [04:02] to do this. Also, no fallback syntax [04:05] highlighting? [04:06] You can see right here that it does some [04:08] sort of reading from the environment. [04:09] When you read from the environment or [04:11] you get something from an array, you [04:12] have like first.has, that meaning this [04:15] thing is not null, and then doing a [04:17] standard memory equal first value to [04:19] write that means it's doing a memory [04:21] compare, a mem compare. This is like [04:23] straight out of Zig's playbook. Second, [04:26] when you write out to a file for [04:28] whatever reason, this writing out to a [04:30] file just somehow works, but when you [04:33] print out, then you have to do a check. [04:36] Like, what I mean, writing to a file [04:40] definitely can fail, but writing out to [04:42] the CLI, I mean, that can also fail, [04:45] too, but I just [04:47] What is this language design? Also, [04:49] check. I assume this is just effectively [04:50] if you get an error, it returns it, and [04:53] you can kind of tell that it returns it [04:54] because raises must say, "Hey, this [04:56] function can raise." And then this must [04:58] be the return type, which means you [05:00] effectively have some sort of colored [05:01] function, meaning that when you call [05:03] something that can raise an error, you [05:06] yourself probably will end up raising an [05:08] error, and then you kind of have this [05:09] raises goes everywhere. When I look at [05:12] this language, you guys can see all of [05:14] this. There's nothing here that's [05:16] interesting. There's nothing here that I [05:17] see as being agent first. This just [05:19] looks like Zig, which seems a lot kind [05:22] of almost antithetical to what they're [05:24] saying right here. Like, if you want [05:26] something that always is more explicit, [05:29] then do pray tell why you're just [05:31] tossing errors all over the place. This [05:33] kind of feels like an if error does not [05:35] equal nil situation brewing, but that's [05:37] not happening. Like, for me, this is [05:39] where I genuinely struggle. I don't get [05:42] why this exists. I don't see how we got [05:46] anything better here. I was actually [05:48] super curious about what does this [05:50] language look like? What does it agents [05:52] first language look like? I don't really [05:55] get that. Maybe it has something to do [05:57] with how the language is really small, [05:59] so it saves on tokens. Well, that's not [06:01] the case. [06:02] Maybe there's something that I can't [06:04] conceive of because my human brain is [06:06] too small and pattern matching in some [06:08] higher dimension that I will just never [06:10] see because I can't see around corners. [06:13] Nope. It's just plain old Zig with a [06:16] little bit of extra and some renaming. [06:18] Now, one day it may finally actually get [06:19] to where it needs to be, which obviously [06:21] is Go. That's where it's heading at [06:23] least based on its description. But [06:25] until then, this is what it looks like. [06:27] Also, just as a short nitpick, void not [06:30] just being a keyword but instead a type, [06:32] I'm triggered. Like honestly, I'm I'm [06:34] triggered AF, okay? That is just Okay, [06:37] hey, this is a crime against humanity. [06:39] But enough about what I think, what do [06:41] people think? Language with zero [06:43] representation in the current training [06:45] data set, bold move. Wow, this might be [06:47] in the training data in 2 years. Then it [06:49] may be usable. I'm assuming the Arch 64 [06:51] back end intentionally does nothing, but [06:53] what's the point of the project in [06:55] general? Is there a design doc available [06:57] anywhere to read? Many of the decisions [06:59] make absolutely zero sense to me. [07:01] Interesting, but how do you even know [07:04] it's better for agents? Any backing [07:05] results? Better than what? All existing [07:08] programming languages? How is it for [07:10] humans? How do agents deal with the [07:12] relatively tiny code example size for [07:14] zero? What about the ecosystem? A brand [07:17] new language that no LLM has ever heard [07:19] about or knows how to write. Brilliant. [07:22] Zero is amazing. I want to try this [07:25] programming language for agents. Bro [07:27] just called this amazing and then goes I [07:29] want to try. Like you can't [07:31] You can't call this amazing. Now, as far [07:34] as it actually goes and more of like my [07:36] opinion on the technical side, I was [07:38] pretty alarmed to see this tweet right [07:40] here. I don't know, I found it funny [07:41] that there's 15 basic identical vector [07:44] implementations. He is correct. You can [07:46] see right here length, capacity, and [07:49] then field type items. It does the same [07:51] thing right here with ex- expression. [07:53] And then there's a whole bunch of other [07:54] ones type args that go on. There's [07:57] literally the same code over and over [07:59] again handling the push logic, being [08:02] able to push into each one of these. [08:05] It's so good. Don't get me started on [08:07] the back end because they're each having [08:08] this nice different way in which it's [08:10] handling big endian and different [08:11] operations, but they're all [08:13] approximately named the same. If you [08:14] scroll down, you'll see something like [08:16] this one A64 diag in which it checks to [08:19] see if this pointer does exist. And then [08:21] in the next implementation when you go [08:23] down, the elf diag, oh that always [08:25] exists. This is truly kind of like an [08:28] alarming project for me in my personal [08:31] opinion. Now, I don't know what it takes [08:33] to build a programming language. I've [08:34] built some toy ones. I did, I built a [08:36] compiler in college as one of the most [08:38] informative times of my life. Uh but [08:40] when it comes down to actually building [08:42] your own that you're actually going to [08:43] support with lots of lots of people in [08:45] the really like strong think through [08:47] that you have to do on language [08:49] ergonomics and design goals, I've never [08:51] done that. And this seems well, it seems [08:54] very interesting. In fact, when I told [08:56] the Teeg about this, he said, "Oh yeah, [08:59] I read that earlier. I'm designing my [09:00] own language too right now." Which is [09:02] just even I [09:04] >> [laughter] [09:04] >> I didn't even know how to respond to [09:06] that. Of course, it was OCaml-based [09:07] classic Teeg OCaml. But nonetheless, [09:10] like I just don't understand the purpose [09:12] of this. The language doesn't offer [09:14] anything that at least appears to be [09:15] interesting. Is the value all in the [09:17] compiler output? If that was the case, [09:19] why not just do something for [09:20] TypeScript, an extremely well-understood [09:24] environment? If you're concerned about [09:26] dependencies, you could have like the [09:27] TypeScript toolkit in which bootstraps [09:30] your project with everything you need to [09:31] write a program. And then guess what? [09:33] You're at Versel, so it's just [09:34] automatically going to work anyways. [09:36] Like I just truly don't understand what [09:39] the heck is this thing even for? Here's [09:41] the thing, I'm not trying to make fun of [09:43] Chris Tate or anything. It's just that I [09:44] I truly don't get this. Now, let me [09:46] explain myself why I keep saying that. [09:48] One, there's nothing interesting at [09:50] least as far as I can tell from a [09:52] language perspective. It just looks like [09:53] another C clone with some basic errors [09:56] as values and some conveniences around [09:59] auto magically returning out like that's [10:01] in Odin, that's in Zig. I wish it was in [10:04] Go, but Go it's not in you. It's just [10:06] that dog ain't in you, Go. The thing [10:07] about the language is that as far as I [10:08] can tell there's nothing agent first [10:11] about it. The examples given are [10:13] extremely verbose and so a bunch of [10:16] tokens means context windows get [10:18] polluted faster. So, that can't be it. [10:20] Also, since it's a novel language, the [10:23] context again is going to be polluted [10:25] faster by giving it the kind of [10:26] pre-required information for it to even [10:28] generate correct output. So, in other [10:31] words, it's doesn't appear to be agent [10:34] friendly at all. And the thing is is [10:35] like [10:36] what makes TypeScript good for agents is [10:39] that there's 1 billion examples of [10:41] TypeScript. And so generally speaking, [10:44] it kind of knows what right looks like [10:46] even when it produces horribly ugly [10:48] code. Okay, whatever. I [10:50] at least it can produce things generally [10:51] that work. Lot of heavy lifting in the [10:54] word work there. And then there's Rust. [10:56] It also has a gazillion examples, but it [10:58] also has the borrow checker which tends [11:00] to make it so that it has to keep the [11:01] LLM even more honest because so much of [11:04] it is even more constrained. Like the [11:06] amount of programs you can write with [11:08] Rust is greatly reduced comparatively to [11:10] say something like TypeScript. And so [11:13] those both offer something unique. I [11:15] don't know what this thing offers. And [11:17] even more so, I was genuinely excited to [11:20] understand what is a programming [11:21] language for agents because hey, I could [11:23] be ignorant in some areas. I actually [11:24] wanted to see like is there a [11:26] fundamentally different take [11:29] on what it means to write a program that [11:31] is not optimized for say human input or [11:34] is there an entire different branch or [11:35] family of languages that just seem to be [11:38] better? I'm surprised it's not Lisp, [11:39] right? I thought Lisp was the greatest [11:41] language of all time. The data is the [11:43] program. But I instead it's it's just [11:46] well, it's pretty much just C that [11:47] borrows a lot of stuff from Zig and [11:50] raise is kind of looks like throws from [11:52] Java. Check is clearly try from uh [11:55] what's it called? Zig. The only thing [11:57] that I think I guess they're going for [11:59] that would be unique is better error [12:01] messaging, but I feel like that's [12:02] something you could contribute to or [12:04] fork off already existing tools to say [12:06] in TypeScript or in Rust, but Rust [12:08] already has pretty decent tools and [12:12] pretty decent output. So that can't be [12:14] it. TypeScript has pretty decent errors. [12:16] So that can be like I just don't [12:18] understand it. And by the way, JSON [12:21] isn't some sort of magical unlock to [12:23] make LLMs understand stuff better. Like [12:25] I can type barely coherent English and [12:28] the thing can understand what I'm [12:30] saying. It's not like typing it in JSON [12:33] somehow makes it that much better. I can [12:35] hand it crazy C++ template errors and [12:38] it's going to spit out for me where the [12:40] template has gone wrong. It's rather [12:42] impressive. It operates on languages. [12:45] It's not somehow confined to HTML or [12:47] JSON despite what you've heard. So I [12:49] truly don't get like I I just don't. I'm [12:53] so curious and I was so excited. Please, [12:55] if you understand, please just explain [12:57] to me, okay? I'm old, okay? My fossils [12:59] are creaking. I'm clearly not that [13:01] smart. I'm not a language enthusiast. [13:03] I'm just a language using enthusiast [13:05] that has programmed in many, many [13:06] languages. I think Ginger Bill said it [13:09] best, the creator of Odin of course. I [13:11] think people who are designing languages [13:13] for LLMs are doing the equivalent of [13:15] horoscopes. They have no idea what they [13:18] are optimizing for since they don't [13:20] actually understand how these models [13:21] work. Why would their new language fair [13:24] better than any existing language? And [13:27] right now I'm completely inclined to [13:28] believe him. I There is absolutely no [13:31] reason for us to think that somehow [13:34] there is a language out there that is [13:35] just dramatically better for agents than [13:38] what we currently have. And if there was [13:40] a language that was dramatically better, [13:42] I assume it would be dramatically [13:44] different from any language we have [13:46] today. Okay, that's it. That's all I had [13:48] to say. I know that was There was some [13:50] strong app in me. The name is the [13:52] Primagen.  📄 **Full text:** ![[Full Document Contents/Articles/Why Did They Build This.md|Why Did They Build This? by YouTube]] ## Highlights - You know, I've always loved programming languages. I have looked into many and I've found love in so many different flavors. And so when today I saw a brand new one dropping from Vercel, you could imagine I was a bit curious. I was piqued, okay? My hams, they were salted. I was ready. Introducing Zero, the programming language for agents. Now, I'm not going to lie to you. I immediately don't know what that means. I always thought English was the new programming language for 2026. AM I ([View Highlight](https://read.readwise.io/read/01ks86fw9cjrvg9eksdetx79as)) - Tags: action/topic/action items, action/h3 - effectively have some sort of colored function, meaning that when you call something that can raise an error, you yourself probably will end up raising an error, and then you kind of have this raises goes everywhere. When I look at this language, you guys can see all of this. There's nothing here that's interesting. There's nothing here that I see as being agent first. This just looks like Zig, which seems a lot kind of almost antithetical to what they're saying right here. Like, if you want something that always is more explicit, then do pray tell why you're just ([View Highlight](https://read.readwise.io/read/01ks5ds0gxyj8qt2x2p94g9qyn)) - Note: Zero offers no clear advantages over existing programming languages - Tags: atom/literature/h1 - just being a keyword but instead a type, I'm triggered. Like honestly, I'm I'm triggered AF, okay? That is just Okay, hey, this is a crime against humanity. But enough about what I think, what do people think? Language with zero representation in the current training data set, bold move. Wow, this might be in the training data in 2 years. Then it may be usable. I'm assuming the Arch 64 back end intentionally does nothing, but what's the point of the project in general? Is there a design doc available anywhere to read? Many of the decisions make absolutely zero sense to me. ([View Highlight](https://read.readwise.io/read/01ks5dssgt3ms2fcek4y70zse9)) - Note: Training data volume determines AI programming language effectiveness - Tags: atom/zettel/z1 - just being a keyword but instead a type, I'm triggered. Like honestly, I'm I'm triggered AF, okay? That is just Okay, hey, this is a crime against humanity. But enough about what I think, what do people think? Language with zero representation in the current training data set, bold move. Wow, this might be in the training data in 2 years. Then it may be usable. I'm assuming the Arch 64 back end intentionally does nothing, but what's the point of the project in general? Is there a design doc available anywhere to read? Many of the decisions make absolutely zero sense to me. ([View Highlight](https://read.readwise.io/read/01ks86g8nmtj8yjb1cqb6knnee)) - Tags: atom/zettel/z1, hypothesis - Interesting, but how do you even know it's better for agents? Any backing results? Better than what? All existing programming languages? How is it for humans? How do agents deal with the relatively tiny code example size for zero? What about the ecosystem? A brand new language that no LLM has ever heard about or knows how to write. Brilliant. Zero is amazing. I want to try this programming language for agents. Bro just called this amazing and then goes I want to try. Like you can't ([View Highlight](https://read.readwise.io/read/01ks5dscy6v6b05sd2q3ghcvx1)) - Note: Novel programming languages face adoption challenges with AI systems - Tags: atom/literature/h2 - Interesting, but how do you even know it's better for agents? Any backing results? Better than what? All existing programming languages? How is it for humans? How do agents deal with the relatively tiny code example size for zero? What about the ecosystem? A brand new language that no LLM has ever heard about or knows how to write. Brilliant. Zero is amazing. I want to try this programming language for agents. Bro just called this amazing and then goes I want to try. Like you can't ([View Highlight](https://read.readwise.io/read/01ks86fpvt7zjp48xgp2qwjqn1)) - Tags: atom/literature/h1 - in Odin, that's in Zig. I wish it was in Go, but Go it's not in you. It's just that dog ain't in you, Go. The thing about the language is that as far as I can tell there's nothing agent first about it. The examples given are extremely verbose and so a bunch of tokens means context windows get polluted faster. So, that can't be it. Also, since it's a novel language, the context again is going to be polluted faster by giving it the kind of pre-required information for it to even generate correct output. So, in other ([View Highlight](https://read.readwise.io/read/01ks86g2p23se2xdg2hkx3sj1m)) - Tags: hypothesis, atom/zettel/z1 - smart. I'm not a language enthusiast. I'm just a language using enthusiast that has programmed in many, many languages. I think Ginger Bill said it best, the creator of Odin of course. I think people who are designing languages for LLMs are doing the equivalent of horoscopes. They have no idea what they are optimizing for since they don't actually understand how these models work. Why would their new language fair better than any existing language? And right now I'm completely inclined to believe him. I There is absolutely no ([View Highlight](https://read.readwise.io/read/01ks5dsfck0ra77tft74tp2v73)) - Note: Language design for AI agents lacks empirical foundation - Tags: atom/literature/h3