Should We Make It Open Source - CourtneyR.dev's 2nd Brain

Should We Make It Open Source? ![rw-book-cover](https://img.youtube.com/vi/laEzOCgtK6c/maxresdefault.jpg) 🔗 **Source:** [Should We Make It Open Source?](https://www.youtube.com/watch?v=laEzOCgtK6c) > [!note]+ Document Notes > tags: #open-source #startup-strategy #ai-coding #business-model > author: Ahmed > source: YouTube > reference: https://www.youtube.com/watch?v=laEzOCgtK6c > date created: 2025-01-12 14:30 > > # 🏷️ Title: Should we make it Open Source? ($5M Seed SF Startup / OSS Decision) >  ## 🧰 Action Inbox - [H1] Evaluate taste model architecture for open source readiness — technical foundation #ai-coding - [H2] Define community-driven monetization strategy beyond API costs — business model #startup-strategy - [H3] Plan security hardening process for open source release — risk mitigation #open-source ## 🧭 Table of Contents - [[#📝 Summary]] - [[#📖 Main Content]] - [[#🗒️ Notes]] - [[#📚 References]] - [[#🗣️ Paraphrase]] - [[#📰 Newsletter]] - [[#✅ Action Items]] - [[#🧪 Hypotheses]] - [[#📊 Methodology]] - [[#❓ Questions]] - [[#🔮 Future Research]] - [[#💡 Implications]] - [[#🧠 Learning Styles Assessment]] - [[#📊 Accessibility Assessment]] - [[#🌍 DEIB Assessment]] - [[#✅ Logical Validity]] - [[#🛡️ DARVO Analysis]] - [[#🎯 Conclusion]] ### 🔗 References - [Command Code Platform](https://commandcode.com) - [Ahmed's GitHub Profile](https://github.com/ahmadawais) ## 📝 Summary - Ahmed, founder of Command Code, debates open sourcing their $5M funded AI coding agent that learns developer preferences through a "taste model" - Company offers an AI coding agent that observes coding patterns and generates auto-updated skills in markdown format - Key tensions: building hackable community tools vs. protecting competitive advantages and monetization - Core business model relies on subsidized API access and optimization rather than selling code itself - Decision impacts community building, recruitment, security posture, and competitive positioning ## 📖 Main Content Ahmed presents the strategic dilemma facing Command Code, an AI coding agent startup that recently raised $5 million in seed funding. The company's core innovation is a "taste model" - not an LLM, but a system that observes how developers code and generates auto-updated skills in markdown format to personalize AI assistance. The taste model addresses a key problem in AI-assisted coding: current tools generate statistically average code that doesn't reflect individual developer preferences or established patterns. Command Code's approach conditions every prompt on learned developer preferences, making the output feel more personal and reducing constant corrections. Ahmed brings significant open source credibility, with 25+ years contributing to projects from WordPress to NASA's Mars Ingenuity helicopter mission, plus 1,000+ GitHub repositories. His "Shades of Purple" theme has 2+ million installs across 30 platforms, demonstrating successful open source community building. The monetization strategy centers on providing optimized API access rather than selling code. Through R&D, they offer the same open source models at 50% cost reduction, with plans starting at $1/month for $10 worth of credits. The key insight: after building with premium models for a week to establish taste, developers can switch to open models while maintaining premium-quality output, reducing API costs by 25x. ## 🗒️ Notes #### ⏱️ [00:00](https://youtu.be/laEzOCgtK6c?t=0) — [H1] Introducing the open source decision dilemma - Key idea: Founder with fiduciary duties weighing community vs. commercial interests - Evidence: "$5 million from amazing investors, I want to make sure we are building a monetized and profitable company" #### ⏱️ [02:15](https://youtu.be/laEzOCgtK6c?t=135) — [H2] Ahmed's open source credentials and background - Key idea: Deep open source experience spanning major projects and platforms - Evidence: "250 PRs merged... NASA Mars Ingenuity helicopter mission... 1,000 GitHub repositories" #### ⏱️ [04:30](https://youtu.be/laEzOCgtK6c?t=270) — [H3] Command Code's taste model innovation - Key idea: Non-LLM system that learns coding preferences and generates auto-updated skills - Evidence: "It sees your accepts, rejects and how you're editing code and builds autogenerated skill" #### ⏱️ [07:45](https://youtu.be/laEzOCgtK6c?t=465) — [H1] The AI coding paradox and personalization need - Key idea: Current AI writes statistical average code that doesn't feel personal - Evidence: "It doesn't feel like it's your code... your preferences, your micro decisions are not there" #### ⏱️ [11:20](https://youtu.be/laEzOCgtK6c?t=680) — [H2] Pro: Making Command Code hackable for community - Key idea: Open sourcing enables developers to build extensions without starting from scratch - Evidence: "You get the best tools, the best system prompt... then you create plugins, extensions" #### ⏱️ [13:45](https://youtu.be/laEzOCgtK6c?t=825) — [H3] Pro: Community building and distribution beyond imagination - Key idea: Shades of Purple theme success demonstrates potential scale of open source adoption - Evidence: "2 million active installs... 5 million developers use... amazing five by five reviews" #### ⏱️ [16:10](https://youtu.be/laEzOCgtK6c?t=970) — [H2] Pro: Solving recruitment challenges through open source contributions - Key idea: Open source contributions provide better hiring signal than text-heavy applications - Evidence: "10-20 PRs already merged, you are an excellent candidate. It's a win-win" #### ⏱️ [17:30](https://youtu.be/laEzOCgtK6c?t=1050) — [H1] Monetization strategy: Product and API platform, not code - Key idea: Business model based on subsidized, optimized API access rather than selling software - Evidence: "10 times discounted plan... we can serve same open source models for $10 instead of $20" #### ⏱️ [20:15](https://youtu.be/laEzOCgtK6c?t=1215) — [H3] Key insight: Taste enables open models to match premium quality - Key idea: After building taste with premium models, can switch to cheaper open models - Evidence: "Open models matching quality of premium models... reduce API costs by 25 times" #### ⏱️ [22:45](https://youtu.be/laEzOCgtK6c?t=1365) — [H2] Con: Copycat concerns and competitive disadvantage - Key idea: Previous experience with rapid copying of innovative features and architecture - Evidence: "Week later, Anthropic came up with... exactly what we are building... 100,000 skills" #### ⏱️ [25:30](https://youtu.be/laEzOCgtK6c?t=1530) — [H3] Reframing copycat risk: Inevitable regardless of open source status - Key idea: With AI assistance, serious competitors will reverse engineer regardless - Evidence: "Anyone serious enough with funding can explain any engineering idea and rebuild it" #### ⏱️ [27:00](https://youtu.be/laEzOCgtK6c?t=1620) — [H2] Security considerations: Short-term vulnerability vs. long-term hardening - Key idea: Open source exposes bugs quickly but enables faster community fixes - Evidence: "WordPress was able to fix so many security bugs that would have never been caught" #### ⏱️ [29:15](https://youtu.be/laEzOCgtK6c?t=1755) — [H1] Conclusion: Leaning toward open source despite risks - Key idea: Community building and distribution benefits likely outweigh competitive risks - Evidence: "Pros are going to outweigh cons... build it together... I've done it before" ## 📚 References - [Command Code Documentation](https://docs.commandcode.com) - [Shades of Purple Theme](https://github.com/ahmadawais/shades-of-purple-vscode) - [Skills vs Taste Blog Post](https://commandcode.com/blog/skills-vs-taste) ## 🗣️ Paraphrase - A well-funded AI startup founder with deep open source roots faces the classic tension between community-driven development and commercial sustainability - The company's core innovation addresses AI coding's personalization problem through learned developer preferences rather than generic statistical outputs - The business model centers on optimized infrastructure and API cost reduction rather than proprietary software licensing ## 📰 Newsletter **Open Source Strategy for AI Startups: A $5M Decision** Command Code founder Ahmed shares his real-time decision process about open sourcing their AI coding agent. The startup's "taste model" learns developer preferences to generate more personalized code suggestions - addressing the core problem that AI-generated code feels generic. Key insight: Their monetization doesn't depend on closed source code, but on optimized API infrastructure that reduces costs 10-25x. After building coding preferences with premium models, developers can switch to cheaper open models while maintaining quality. This case study illuminates how AI-first companies might embrace open source strategies differently than traditional software businesses. The debate touches recruitment, community building, security hardening, and competitive positioning. **Takeaway**: When your value lies in optimization and personalization rather than algorithms, open sourcing the core platform might accelerate rather than threaten business growth. Resources: [Learn WordPress](https://learn.wordpress.org/), [Developer Docs](https://developer.wordpress.org/), [WP-CLI](https://wp-cli.org/) ## ✅ Action Items #### DevRel - Developer advocacy: Create taste model integration workshops for major IDEs - Education: Develop curriculum around personalized AI coding workflows - Developer experience: Document API optimization techniques and cost reduction methods - Community management: Establish contribution guidelines for taste model improvements #### OSPO - Policy & Governance: Draft open source licensing strategy for AI training data and models - License Compliance: Review taste model training data for licensing compatibility - Security Management: Implement security review process for community contributions - Community Engagement: Establish maintainer funding for core taste model development #### Community - Contributors: Create "good first issue" program for taste model feature development - Sponsors: Develop transparent funding model for community-driven features - Ecosystem: Coordinate with IDE and editor extension developers #### WordPress - Core AI: Apply taste model concepts to WordPress AI features and personalization - Community: Share learnings about open source AI development with WordPress AI team - Documentation: Create case studies of successful open source AI business models ## 🧪 Hypotheses - Open source AI coding tools with personalization will achieve higher adoption than closed alternatives - Community-driven taste model development will produce more diverse and effective coding preferences - API cost optimization provides more sustainable monetization than software licensing for AI tools - Developer recruitment through open source contributions yields higher quality hires than traditional methods ### 📊 Methodology - Track adoption metrics for open vs. closed AI coding tools across developer communities - Measure contribution quality and maintainer satisfaction in open source AI projects - Analyze revenue sustainability of API-first vs. license-first AI business models - Survey hiring managers on effectiveness of portfolio-based vs. interview-based technical assessment ### ❓ Questions - How do AI startups balance transparency with competitive advantage in rapidly evolving markets? - What governance models work best for open source projects funded by venture capital? - How can taste models be made interoperable across different coding tools and platforms? ### 🔮 Future Research - Longitudinal study of open source AI startup outcomes vs. closed source equivalents - Framework for evaluating when AI innovations should be open sourced vs. kept proprietary - Analysis of community contribution patterns in AI development vs. traditional software ### 💡 Implications - Practice: AI startups may need new frameworks for evaluating open source strategies - Policy: Venture funding terms might need updating for open source AI business models - Security: Community-driven security review becomes critical for AI tools with code access - Education: Developer education programs should include AI personalization and taste model concepts ### 🧠 Learning Styles Assessment Reading level estimate: 12th Grade (Flesch-Kincaid) #### 👀 Visual learners - Present: Screenshot of taste model output, pricing comparison tables - Improve: Architecture diagrams showing taste model workflow, competitor analysis charts #### 🔊 Aural learners - Present: Video format with founder narration and personal anecdotes - Improve: Podcast version with investor/community perspectives #### 🗣️ Verbal learners - Present: Clear definitions of taste vs. skills vs. rules, business model analogies - Improve: More concrete examples of taste model outputs and transformations #### 🧩 Kinesthetic learners - Present: GitHub repository references, hands-on contribution opportunities - Improve: Interactive taste model demo, step-by-step open sourcing checklist #### 🧠 Logical learners - Present: Pro/con framework, cost reduction calculations, timeline references - Improve: Decision matrix with weighted criteria, ROI projections #### 🧑‍🤝‍🧑 Social learners - Present: Community examples, recruitment benefits, copycat discussions - Improve: Founder community feedback sessions, peer startup case studies #### 🧘‍♀️ Solitary learners - Present: Individual founder decision-making process, personal coding history - Improve: Self-assessment questions for similar decisions, reflection prompts #### 🌿 Naturalistic learners - Present: Evolution of coding practices, ecosystem thinking around AI tools - Improve: Open source sustainability models, community health metrics ### 📊 Accessibility Assessment #### Screen reader semantics - Present: Clear heading structure, descriptive timestamps - Improve: Alt text for any charts or visual pricing information #### Keyboard navigation - Present: Standard video controls and link navigation - Improve: Keyboard shortcuts for timestamp navigation #### Contrast and readability - Present: Standard video player with captions - Improve: High contrast mode for pricing tables and documentation #### Media and captions - Present: Full transcript provided for video content - Improve: Downloadable transcript, audio descriptions for visual elements ### 🌍 DEIB Assessment #### Cultural breadth - Present: Global open source perspective, Pakistani founder background - Improve: Examples from diverse developer communities worldwide #### Inclusive language - Present: Technical terms defined, accessible business language - Improve: Plain language alternatives for venture capital terminology #### Neurodiversity support - Present: Structured pro/con format, clear section breaks - Improve: Executive summary, bullet point key takeaways #### Representation - Present: Founder from underrepresented background in tech leadership - Improve: Diverse community examples, varied developer personas ### ✅ Logical Validity #### Strong reasoning - Deductive strength: Well-structured pro/con analysis with clear criteria - Inductive strength: Good use of personal experience and analogies from Shades of Purple - Causal reasoning: Clear connection between open source strategy and business outcomes - Analogical strength: Effective comparison to WordPress security model #### Fallacies - False dilemma: Presents binary open/closed choice without hybrid models - Appeal to authority: Heavy reliance on personal open source credibility without addressing different contexts #### Tone of voice - Overall tone: Authentic and vulnerable, appropriate uncertainty for complex strategic decision ### 🛡️ DARVO Analysis - Deny: No denial patterns observed - Attack: No attacks on critics or competitors - Reverse victim/offender: No victim/offender role reversals - Emotional cues and rhetoric: Genuine vulnerability and uncertainty about major business decision ## 🎯 Conclusion Ahmed's decision framework illustrates how AI-first startups must reconsider traditional open source trade-offs. When value lies in optimization and personalization rather than algorithms, open sourcing core platforms may accelerate rather than threaten business growth. The taste model innovation addresses real developer pain points around AI code personalization, and the API-first monetization strategy appears more sustainable than software licensing. The decision ultimately favors community building and distribution benefits over competitive protection, reflecting both founder values and market realities in the AI era. ## 🧱 Refactor Appendix ### 📚 Literature Split Pad ### [H1] Open source strategy for AI startups differs from traditional software > [!quote] [Should we make it Open Source?](https://youtu.be/laEzOCgtK6c) by Ahmed via YouTube. Published 2024. Accessed 2025-01-12. > "We are not actually selling code. We are selling a product and an API platform." AI companies monetize through optimization and infrastructure rather than propri --- ## Full Transcript [00:00] Should we open source command code or [00:01] not? This is what we are discussing [00:02] today. Hello, I am Ahmed, founder of [00:05] command code. It's a coding agent that [00:07] can build anything, but it can also [00:09] observe how you write code and learn [00:10] your coding taste. Your coding taste is [00:13] recorded as an autogenerated skill that [00:15] command code can keep updated. So now [00:16] that you know what command code is, [00:18] today we are discussing should we open [00:20] source command code or not. And this is [00:22] a product decision that I have yet to [00:23] make. And as a founder, I have my [00:25] employees to take care of, my investors [00:27] to report to, and the general community. [00:30] And as a company that has just recently [00:32] raised $5 million from amazing [00:34] investors, I want to make sure that we [00:36] are building a monetized and profitable [00:38] company for the long run. So our mission [00:40] is to build the best AI engineering [00:42] developer experience. And we have [00:43] recently raised $5 million in seed from [00:45] all of these amazing investors and [00:47] founders. And if you know a little bit [00:49] about me, I am Ahmed. I have spent most [00:51] of my career 25 plus years building open [00:54] source software one way or the other. If [00:56] you actually go to my uh GitHub profile [00:58] and like go really deep in there, you [01:01] will see that I have like probably 250 [01:03] PRs merged on all of the all of these [01:05] software like core contributions to MCP [01:08] protocol contribution to NASA Mars [01:10] Ingenity helicopter mission from Omisesh [01:12] Bootstrap WordPress React to something [01:15] like NodeJS core to W3C Google Chrome VS [01:18] Core I've done pretty broad strokes as [01:20] you can imagine different languages [01:21] frameworks and I have built a lot of [01:25] open source software. At the time of [01:26] recording this video, I have over 1,000 [01:29] GitHub repositories and maybe around 300 [01:32] plus of them are actually useful and [01:34] I've written most of code that I've [01:36] built are CLIs. I love building [01:38] automation. So by looking at just this [01:40] information, you should be able to [01:41] understand that I'm an open source [01:42] developer and I was born and raised in [01:44] Pakistan but I ended up moving to the US [01:47] later on. But even now I was just [01:48] talking to my brother about it who [01:50] happens to work at my company as well [01:52] and he was sharing that if you search [01:54] something like who's the most prominent [01:55] open source developer from Pakistan with [01:57] any AI out there uh it's a high [01:59] probability that my name comes up and I [02:02] take a lot of pride in that I love [02:03] building software so this is really good [02:06] like you know this is who I am so I'm [02:08] going to factor all of this in and this [02:10] is you know I'll kind of look at above [02:11] my shoulder uh you're basically looking [02:13] into my brain how a founder is taking [02:15] decisions nowadays in the age of AI [02:18] where Kurser is probably going to get [02:20] acquired by SpaceX because even though [02:22] they are a $50 billion company now they [02:25] are running out of money and running out [02:26] of ways to buy more compute. So it's a [02:29] crazy world out there. Let's go one by [02:31] one and figure out the pros and cons [02:33] that I can think of and I would love to [02:35] see what your feedback is. So one of the [02:37] biggest goals I have in open sourcing [02:39] command code is I want to make it [02:41] hackable. I really really want that [02:44] every developer should be able to use [02:46] what we have built and build on top of [02:48] it. So your coding agent is more closer [02:51] to what you want it to be like. So our [02:53] first attempt in doing this was [02:55] obviously the taste model. If you go to [02:57] our docs you will see that it's a [02:58] frontier coding agent. It can it can [03:00] write code. It can fix bugs, write [03:03] tests, you know do your pull request [03:04] reviews and whatnot. It comes with all [03:06] the cloud models, GPD models and uh [03:08] top-notch open source models. And it [03:11] also has a taste one model. I'm not [03:13] going to go in the details of this. You [03:15] can, you know, take time and study this. [03:17] But what this taste model does is it's [03:19] not an LLM. It is not there to write [03:21] code. It works with whatever LM you are [03:22] using. Maybe you're using cloud or maybe [03:24] you're using Kimmy. It sees it observes [03:26] how you write code or how you prompt AI [03:29] to get that code to production. It sees [03:31] your accepts. It sees your rejects and [03:34] how you're editing the code and it [03:35] builds it autogenerates a skill that is [03:37] a project level skill inside of your [03:39] project in the form of markdown files [03:42] and it also tries to continuously learn [03:44] more and more as you build with command [03:45] code and it continuously you know [03:47] updates that skill as well on our blog [03:49] you can also find a blog post that I [03:51] wrote about in fact I can actually go [03:53] back here in skills versus taste because [03:55] this is the most common question that I [03:56] keep answering again and again that what [03:58] is the difference between skills taste [04:00] and rules I'll give you a quick gist [04:02] that uh taste is the highest order bit [04:04] on these two things right taste is [04:06] continuously learn auto manage and [04:08] compounds and it is personal to you [04:10] taste is able to manage your skills and [04:12] rules skills are your reusable workflows [04:14] that you sit down and write and I kind [04:17] of call them tutorial for agents right [04:19] and rules are very static and very high [04:20] level constraints that when you are [04:22] doing X always make sure you're doing Y [04:24] and Z right so this is the top level [04:26] difference between this why I am sharing [04:27] this with you folks is also I probably [04:29] should mention one last thing before [04:31] moving forward is if you take a look at [04:32] our launch blog post, I actually wrote a [04:34] lot about my intentions and our [04:36] intentions and our research into why we [04:38] are trying to build something like this. [04:40] The reason why AI assisted uh coding has [04:42] a paradox. It's writing code that is a [04:44] statistical average of the entire [04:46] internet and it is not always right. [04:48] Even if that code is running, it doesn't [04:49] feel like it's your code. Your [04:51] understanding, your preferences, your [04:53] you know micro decisions that have you [04:55] you have built up in your head for you [04:57] know years, they are not there when you [04:59] are using AI. So I want to build a [05:00] coding agent that is more and more [05:02] closer to you. So you're not endlessly [05:03] correcting it and it is not ignoring [05:06] some you know things that you really [05:07] care about. It's actually uh your [05:09] preferences are not getting ignored. [05:11] It's mostly like you what you describe [05:12] is what you get right and it's not an [05:15] average code that is being shipped you [05:16] know it's not the average of entire [05:17] stack overflow or github that they [05:19] trained on top of right and again I go [05:21] uh in quite a lot of details on how we [05:23] actually stumbled on this all of the [05:24] taste files actually uh you know do look [05:26] like uh that they are transparent they [05:28] are very similar to scales this is [05:29] generally how you work with or without [05:31] taste without taste you're constantly [05:33] interrupting and staring your coding [05:34] agent that oh no do not use ts uh to [05:37] build this I want you to set up t-up or [05:39] do not use mocha I love vitest and with [05:42] taste built it just makes the right [05:44] decisions on one go or the second go [05:45] when you it has learned something from [05:47] you and we've gone into depth of [05:49] discussing you know how we think this [05:51] becomes really really better it's again [05:53] it's just what we are doing is it's your [05:54] prompt your taste and we are [05:57] conditioning every prompt that you send [05:59] on top of that right so there's more of [06:01] you and your thoughts and your [06:03] intentions in the code that is being [06:04] generated by command code so now [06:06] naturally you can think of my intentions [06:08] here I want to make command code so [06:10] hackable [06:10] that it feels like an extension of you. [06:12] So we can continuously build the core of [06:15] command code but you on the other hand [06:16] can go ahead and build on top of things [06:18] that you care about that might not fit [06:20] our business model. Right? So you don't [06:22] have to start from scratch. You get the [06:24] best tools, the best system prompt [06:25] possible which is constantly evolving [06:27] and improving and then you are able to [06:28] create plugins, extensions or you're [06:30] completely able to hack what command [06:32] code is for you. So to be able to do [06:34] this I have to open source, right? [06:37] That's a pro, right? I want this and I [06:40] think that would work really really [06:41] well. The second thing that I'm coming [06:43] up with is I want to build a community [06:45] around this. And community is bigger [06:46] than code, right? Code is cheap. [06:48] Everybody can code. AI can code really [06:50] really well if we are not building this [06:52] together. I am not having a lot of fun [06:54] as uh as we speak. Maybe it's just you [06:56] know me being me, right? I'm an open [06:58] sourcer and I want to open source this [07:01] right. One of the other things that I [07:02] can come up with is distribution. We [07:04] want command code to be used beyond use [07:07] cases that we can think of, right? Uh [07:09] you shouldn't have to come and buy from [07:11] us. you shouldn't have to wait for us to [07:12] do things. We want millions of [07:14] developers to be using this code. Like [07:16] for example, I built this thing called [07:18] shades of purple code theme. Uh it is [07:21] available for 30 different uh software. [07:23] It has like more than 2 million active [07:25] installs. And the thing that I wanted to [07:27] do here was I wanted to make sure that [07:28] the syntax the colors of different [07:30] things are semantic. So if you're using [07:32] 35 different programming languages, [07:34] you're using Rust, JavaScript, [07:35] TypeScript, Python, no matter what, even [07:37] if a simple meaningless thing like you [07:39] know declaring a variable would have the [07:41] same color code across these different [07:44] languages and then a bunch of other [07:45] things. And then turns out this is an [07:47] idea that not only I cared about but a [07:49] huge community cared about and now more [07:50] than 5 million developers use uh shares [07:53] of purple code theme which is open [07:54] source and have amazing reviews. Most of [07:57] those reviews are five by five, right? [07:59] So coming back the pros of open sourcing [08:01] command code are making it hackable [08:03] building a community creating [08:04] distribution beyond what I think is [08:06] possible and what else I think one thing [08:09] that is coming to my mind is recruitment [08:11] hiring has become very very hard right [08:14] uh running an AI company building so [08:16] many agents I am struggling a lot in [08:18] figuring out who are the best candidates [08:20] to be at this company so this is a big [08:22] struggle and this is a struggle on both [08:24] ends right a lot of slop is going around [08:26] I am anybody who writes more than like a [08:28] paragraph I am not reading that. I'm [08:30] finding that very hard. It feels like I [08:32] have, you know, a high limit of reviews [08:36] left in me after, you know, doing like [08:38] 10 to 50 PRs a day. And it's just not in [08:40] me to be reviewing so much text anymore. [08:43] So, I think we open source, I can just [08:46] redirect people to, you know, go ahead [08:48] uh our source is open code uh contribute [08:50] something awesome there. And if you have [08:52] like 10 20 PRs already merged, you are [08:53] an excellent candidate. It's a winwin on [08:56] both sides. So I think recruitment could [08:57] be a really really good win on the other [08:59] side and then there are cons of you know [09:02] open sourcing all of this on top of my [09:03] head that the things that I can think of [09:06] are probably monetization. I think if we [09:10] open source command code how are we [09:11] going to monetize it? My gut feeling [09:13] right now on this is that we are not [09:16] actually selling code. We are selling a [09:17] product and an API platform. If you [09:20] actually go take a look at our pricing [09:22] page, you will see that we are selling [09:24] command code as a provider, right? The [09:26] command code provider is subsidized, [09:29] right? So, you're not paying the API [09:30] cost because we are able to do a bunch [09:32] of amazing things. That is our R&D. We [09:34] are able to figure out how to serve the [09:36] same open source models that you are [09:39] going to pay I don't know 20 bucks for [09:41] in 10 bucks, right? So all the [09:44] optimizations we are able to pass [09:46] through to you which is why next week we [09:48] are launching our go plan which is just [09:49] starting with $1 a month in which you [09:51] are getting $10 worth of credits in [09:53] which you can do anywhere from 8,000 Kim [09:56] K2 requests to 12,000 miniax requests or [10:00] whatnot right these are a lot of credits [10:02] so we you this is like a 10 times [10:03] discounted plan this is twice as much [10:06] and as you know uh further if you're a [10:08] very heavy user you can get $300 worth [10:10] of credits for now for all the premium [10:13] ium and uh open source models and we are [10:16] discovering that if you are able to use [10:18] top models like cloud or GPT I will just [10:21] call them premium models we are [10:22] discovering that if you are able to use [10:24] premium models to build with command [10:26] code the outcome of this is after [10:28] building maybe for a week you will have [10:31] built your coding test that taste is a [10:33] skill that is autogenerated and auto [10:35] updated by command code and it kind of [10:38] knows what you want to do how you want [10:39] to do in a particular project now with [10:42] command [10:43] and taste you can actually swap out the [10:45] premium model from with an open model [10:47] and it will start matching the quality [10:49] of premium models. This is the space we [10:51] are after. I think this is an excellent [10:53] idea and it has started to work for us [10:55] for almost a month now where open models [10:58] are stacking up that if we are able to [10:59] build our engineering codex or our taste [11:01] internally and that taste is able to [11:03] steer it to do the right things we are [11:05] able to reduce our API costs on these [11:07] LLMs by 25 times right this is what we [11:10] offer so this is how we are monetizing [11:11] it right so this is what the product is [11:14] potentially and not the code and the [11:15] code part might not suffer the [11:18] monetization issues if we open source [11:20] this is again a gut feeling I [11:22] The next thing I can think of are [11:23] copycats. Even when I was building uh [11:25] Langbase, the OG AI cloud started late [11:28] 2023, what we saw was everybody just [11:31] copy pasted everything that we did. We [11:33] ended up inventing a pipe architecture [11:35] for hosting agents which would connect [11:37] any LLM to any data and give you an API [11:40] on an isolate which was absolutely dirt [11:43] cheap, 80% cheaper than AWS. We saw this [11:47] become the foundation of sandboxes and [11:49] then we built uh one of the cheapest [11:51] possible lowc cost possible and fastest [11:53] memory engine that allowed us to give [11:55] you like millions of databases vector [11:57] store databases in one hour um and scale [12:00] your rag engine very very quickly and [12:03] now everybody has a version of this [12:04] based on object storage or whatnot. All [12:06] I'm saying is I've gone through this uh [12:07] struggle of uh seeing everybody copy [12:09] what we do first and then you know [12:11] seeing in the community that they are [12:13] just rebranding what they think are the [12:14] nice ideas. I started with my company by [12:17] calling we are building a composable AI. [12:18] Then everybody was composable AI. We [12:20] switched to the idea of primitives [12:22] instead of giving you the complete AI [12:23] cloud. And then everybody in the market [12:25] openai and anthropic started calling you [12:28] know uh it primitives or what not and [12:30] everybody uh started ending on the idea [12:32] of like oh primitives work really well. [12:34] I'm not saying that they are absolutely [12:35] out to copy me or our company. It is [12:38] just weird things are happening. It's [12:40] much much easier to copy somebody [12:42] nowadays than it was before. And I think [12:45] a lot of smart people sometimes by [12:47] looking at each other or sometimes [12:49] completely independently are arriving at [12:51] the same solutions uh which to me or uh [12:54] to many might look like copycats right [12:56] so they're both sides to this if we open [12:59] source are copycats who are absolutely [13:01] copying everything that we push out like [13:03] last year October 15th or maybe October [13:06] 10th I gave a talk at uh Cloudflare [13:08] introducing what taste look like it's a [13:10] markdown file and it learns how you are [13:13] building software And maybe a week [13:15] later, Entropy came up with oh we can [13:17] skip the autogeneration of this part and [13:19] the continuous learning part because [13:21] because that is actually a hard part and [13:23] we can call this file skills.mmd right [13:26] and I couldn't help but think that oh my [13:29] god this is exactly what we are building [13:30] and within a month there were like [13:32] 100,000 skills and everybody jumped on [13:34] the idea of skills and that's fine but [13:36] it does hurt in the short term right so [13:39] if we open source command code I think [13:41] we will be giving copycats you know an [13:43] easier road to come copy us uh and do it [13:46] you know in days instead of weeks but on [13:48] the other side I can see the angle of [13:50] maybe they are going to do that anyway [13:52] our source code is actually on npm it's [13:53] not a binary and even if they are not [13:56] looking at our source code anybody [13:58] serious enough with enough funding and [14:00] time with the right models can explain [14:03] any engineering idea and get your models [14:06] to rebuild it so nobody's really safe [14:08] from copycats everybody's going to build [14:11] something similar there's nothing new [14:13] under the sun nowadays. This is more [14:15] true than ever. So, should I be worried [14:17] about copycats at all? Probably not, [14:20] right? Uh China might have already [14:22] copycated command code and we never [14:24] know. That's that's a comment I got from [14:26] stream worlds are saying this. Um so [14:28] yeah, that is happening, right? Another [14:30] thing that I can think of is security. [14:31] Security is a mess. I did a video on [14:33] cloud mythos and it is scary how [14:35] security is becoming a nightmare. Quite [14:37] recently, Versel got hacked through an [14:39] ooth token from somebody on their team [14:41] and lovable found out that their chats [14:45] were already public through somebody in [14:47] the community. All of these people do [14:48] not even have access to uh claude mythos [14:51] level model at the moment. So when [14:53] mythos arrives uh software is going to [14:56] be a nightmare. I feel like that every [14:58] single day, right? So one of the two [15:01] things are going to happen if you do not [15:02] open source doesn't mean you are secure. [15:04] It means somebody has not just found a [15:06] way to find that bug. If you do open [15:09] source all those security bugs will [15:11] probably people will just find them very [15:13] very quickly. So for a very short amount [15:15] of time you will struggle and when you [15:17] are a new startup a young startup I [15:20] think at that time that is fine right? [15:22] It's not fine but it's fine relatively [15:24] fine. So I think by open sourcing we [15:26] will be improving our security and it [15:28] will get hardened like I've seen that [15:30] even in the WordPress space because [15:31] WordPress was open source we were able [15:33] to fix so many security bugs that [15:35] otherwise would have never been caught. [15:37] I think in short term this is a con but [15:39] in long term this will probably will [15:41] become a pro. So here I rest my case I [15:44] think and I'm constantly thinking about [15:46] pros and cons of open sourcing command [15:48] code and I I'd love to hear your [15:50] thoughts on this. I am hellbent and I'm [15:52] leaning towards open sourcing it. I [15:54] think it will bring us more joy. It will [15:56] make command code hackable. Even though [15:58] command is not hackable right now in its [16:00] current form, it is definitely something [16:01] we care about and it is on our road map. [16:04] And I think we will be able to build a [16:05] much better community. We will be able [16:07] to even for our copycats, we will be [16:09] able to put our new features in a tray [16:11] and serve them like here you go ahead [16:13] please sir copy this. Right? You know [16:15] every week on a Monday we feel like okay [16:17] this this just happened okay they copied [16:18] that feature as well. Right? If that is [16:20] going to happen anyway, why don't we [16:21] build it together, right? I think so far [16:23] in my head, pros are going to outweigh [16:25] cons. And if we are able to build a [16:27] community and a distribution engine [16:29] while doing this, we will be able to [16:31] monetize it. We will be able to build a [16:33] really good business on top of it. I've [16:35] done it before. I've done I've made a [16:37] bunch of different open source uh you [16:39] know, profitable before. So, I kind of [16:40] I'm going to take a leap of faith here. [16:43] Still, these are very raw thoughts. I [16:45] have no idea what is the right thing to [16:47] do as a founder of a young company [16:49] raised $5 million uh with employees to [16:52] take care of with uh investors to report [16:55] to and with almost 5 years of runway [16:57] ahead of us. I am both very very excited [17:00] and a little bit scared of what will [17:02] happen. So I'm going to lean one way or [17:04] the other and this is going to happen [17:06] very very quickly. I would love to know [17:08] what your comments are on this. With [17:10] that, use your code for good. Peace.  📄 **Full text:** ![[Full Document Contents/Articles/Should We Make It Open Source.md|Should We Make It Open Source? by YouTube]] ## Highlights - one and figure out the pros and cons that I can think of and I would love to see what your feedback is. So one of the biggest goals I have in open sourcing command code is I want to make it hackable. I really really want that every developer should be able to use what we have built and build on top of it. So your coding agent is more closer to what you want it to be like. So our first attempt in doing this was obviously the taste model. If you go to our docs you will see that it's a frontier coding agent. It can it can ([View Highlight](https://read.readwise.io/read/01kqqypg0ntt9k35wd122r0942)) - Tags: 👻 ai highlighted - right. One of the other things that I can come up with is distribution. We want command code to be used beyond use cases that we can think of, right? Uh you shouldn't have to come and buy from us. you shouldn't have to wait for us to do things. We want millions of developers to be using this code. Like for example, I built this thing called shades of purple code theme. Uh it is available for 30 different uh software. It has like more than 2 million active installs. And the thing that I wanted to do here was I wanted to make sure that the syntax the colors of different ([View Highlight](https://read.readwise.io/read/01kqqypg1m4ym0vew60yr0zxta)) - Tags: 👻 ai highlighted - open sourcing all of this on top of my head that the things that I can think of are probably monetization. I think if we open source command code how are we going to monetize it? My gut feeling right now on this is that we are not actually selling code. We are selling a product and an API platform. If you actually go take a look at our pricing page, you will see that we are selling command code as a provider, right? The command code provider is subsidized, right? So, you're not paying the API ([View Highlight](https://read.readwise.io/read/01kqqypg3ampvk34yyba2hfmw2)) - Tags: 👻 ai highlighted