Quick recap of the I/O:\n- 13 assignable keys with individual LEDs\n- rotary encoder\n- joystick\n- touch disc\n- optional edge-lit LED band
The cool part for us was just HOW our agent was able to do it when every other agentic skills, tools, and even commercial platforms failed. We put together a more detailed writeup in the linked blog post (plus pretty pictures!), but long story short, we asked our agentic scientist, Marvin, to treat the electrical engineering as a research problem.
Marvin was able to synthesize vast technical documents and best practice guidelines (e.g. scientific literature), test specific placement and routing strategies (e.g. experiments), and selectively backtrack and combine learnings (e.g. interpreting data analysis). The end result was a fully routed board (made especially challenging by some hard constraints to keep cost down) plus a reusable runbook worfklow that demonstrated Marvin's ability to enter an unfamiliar domain, learn and improve through experimentation, and unlock novel capabilities.
Some future work we've got marinating in our mind grapes:\n- an app for cross coding agent compatibility (shoutout to lop1381997 for contributing a codex-native firmware)\n- optional OLED screen, something like an SSD1306 that could fit right into the space between the knob and joystick\n- optional base with bluetooth and battery for full wireless functionality
Let us know what you think, have any ideas or feature requests, and share pictures if you build it!
Repo: https://github.com/yuz207/agentpad13\nConfigurator: https://yuz207.github.io/agentpad13/configurator/site/"},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["codex"],"value":"Show HN: AI scientist builds an open-source Codex Micro from scratch for $40"},"url":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["codex"],"value":"https://iluvatarlabs.com/blog/2026/08/how-we-cloned-openai-codex-micro/"}},"_tags":["story","author_ben_iluvatar","story_49453466","show_hn"],"author":"ben_iluvatar","created_at":"2026-08-26T18:16:57Z","created_at_i":1787768217,"num_comments":0,"objectID":"49453466","points":4,"story_id":49453466,"story_text":"A few weeks ago, we open-sourced our design for a cost-effective but full featured alternative to OpenAI x Work Louder's sold out Codex Micro macropad. And now, we're pleased to share that we've actually built them and they work!
Quick recap of the I/O:\n- 13 assignable keys with individual LEDs\n- rotary encoder\n- joystick\n- touch disc\n- optional edge-lit LED band
The cool part for us was just HOW our agent was able to do it when every other agentic skills, tools, and even commercial platforms failed. We put together a more detailed writeup in the linked blog post (plus pretty pictures!), but long story short, we asked our agentic scientist, Marvin, to treat the electrical engineering as a research problem.
Marvin was able to synthesize vast technical documents and best practice guidelines (e.g. scientific literature), test specific placement and routing strategies (e.g. experiments), and selectively backtrack and combine learnings (e.g. interpreting data analysis). The end result was a fully routed board (made especially challenging by some hard constraints to keep cost down) plus a reusable runbook worfklow that demonstrated Marvin's ability to enter an unfamiliar domain, learn and improve through experimentation, and unlock novel capabilities.
Some future work we've got marinating in our mind grapes:\n- an app for cross coding agent compatibility (shoutout to lop1381997 for contributing a codex-native firmware)\n- optional OLED screen, something like an SSD1306 that could fit right into the space between the knob and joystick\n- optional base with bluetooth and battery for full wireless functionality
Let us know what you think, have any ideas or feature requests, and share pictures if you build it!
Repo: https://github.com/yuz207/agentpad13\nConfigurator: https://yuz207.github.io/agentpad13/configurator/site/","title":"Show HN: AI scientist builds an open-source Codex Micro from scratch for $40","updated_at":"2026-08-27T01:10:24Z","url":"https://iluvatarlabs.com/blog/2026/08/how-we-cloned-openai-codex-micro/"},{"_highlightResult":{"author":{"fullyHighlighted":true,"matchLevel":"partial","matchedWords":["codex"],"value":"codex"},"comment_text":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["ssd"],"value":"Surprise feature: 512GB of DRAM associated with the SSD. Is this a write cache? If so, is it battery backed?"},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Mac Pro Late 2013 Teardown"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"http://www.ifixit.com/Teardown/Mac+Pro+Late+2013+Teardown/20778"}},"_tags":["comment","author_codex","story_6990518"],"author":"codex","children":[6992230],"comment_text":"Surprise feature: 512GB of DRAM associated with the SSD. Is this a write cache? If so, is it battery backed?","created_at":"2013-12-31T18:43:25Z","created_at_i":1388515405,"objectID":"6992073","parent_id":6990518,"story_id":6990518,"story_title":"Mac Pro Late 2013 Teardown","story_url":"http://www.ifixit.com/Teardown/Mac+Pro+Late+2013+Teardown/20778","updated_at":"2024-09-19T20:28:55Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"cahaya"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["codex","ssd"],"value":"Until recently my M1 Max 32GB 2021 was great until memory became the bottleneck.\n My M1 Pro 8GB lasted me less than a year bcs with Docker dev stuff made it sluggish. \n32GB was great for years until AI assisted productivity got so high with Codex.\nRunning 3 projects, 3 different worktrees, each either Docker compose or running/building Electron apps. So in this high AI era productivity my Mac is using insane amount of RAM and SSD space (bcs of all the worktrees)"},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Apple introduces M6 and M5 Ultra"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://www.apple.com/newsroom/2026/08/apple-introduces-m6-and-m5-ultra-for-a-big-leap-in-performance-and-ai-compute/"}},"_tags":["comment","author_cahaya","story_49433292"],"author":"cahaya","comment_text":"Until recently my M1 Max 32GB 2021 was great until memory became the bottleneck.\n My M1 Pro 8GB lasted me less than a year bcs with Docker dev stuff made it sluggish. \n32GB was great for years until AI assisted productivity got so high with Codex.\nRunning 3 projects, 3 different worktrees, each either Docker compose or running/building Electron apps. So in this high AI era productivity my Mac is using insane amount of RAM and SSD space (bcs of all the worktrees)","created_at":"2026-08-26T10:12:36Z","created_at_i":1787739156,"objectID":"49446476","parent_id":49440111,"story_id":49433292,"story_title":"Apple introduces M6 and M5 Ultra","story_url":"https://www.apple.com/newsroom/2026/08/apple-introduces-m6-and-m5-ultra-for-a-big-leap-in-performance-and-ai-compute/","updated_at":"2026-08-26T14:53:53Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"pixelbro"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["codex","ssd"],"value":"I literally have had a SOTA Codex agent delete a bunch of files just last month, but it took very specific circumstances. It was working within my game repo, in a branch, and it ran out of SSD. It had to free up some space to work, so it looked around and found 25GB of untracked files in the project, in a folder called Recordings. Might as well clean that up, it thinks. There goes all the raw Unity Recorder footage I'd ever recorded of my game, over many years of development.
So yeah, it won't go on a spree outside of its lane even with full access, but if you give it a box and tell it to go ham, it's on you to make sure you didn't leave precious unrecoverable assets in that box."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"GPT-5.6"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://openai.com/index/gpt-5-6/"}},"_tags":["comment","author_pixelbro","story_48849066"],"author":"pixelbro","comment_text":"I literally have had a SOTA Codex agent delete a bunch of files just last month, but it took very specific circumstances. It was working within my game repo, in a branch, and it ran out of SSD. It had to free up some space to work, so it looked around and found 25GB of untracked files in the project, in a folder called Recordings. Might as well clean that up, it thinks. There goes all the raw Unity Recorder footage I'd ever recorded of my game, over many years of development.
So yeah, it won't go on a spree outside of its lane even with full access, but if you give it a box and tell it to go ham, it's on you to make sure you didn't leave precious unrecoverable assets in that box.","created_at":"2026-07-10T10:15:17Z","created_at_i":1783678517,"objectID":"48857991","parent_id":48855452,"story_id":48849066,"story_title":"GPT-5.6","story_url":"https://openai.com/index/gpt-5-6/","updated_at":"2026-07-11T18:54:55Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"kev009"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["codex","ssd"],"value":"Grok build already punched above its weight and is the nicest TUI, claude and codex are clearly vibecoded by web developers that don't understand systems (eating SSDs, spaghetti logic, extinguishing kernel watch limits, etc). I think Anthropic and OpenAI are both engaged in their own theatre, trying to define and redefine what game they are even playing, trying to shift to immeasurables like safety or security or exclusivity. There's definitely room at the top."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Grok 4.5"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://x.ai/news/grok-4-5"}},"_tags":["comment","author_kev009","story_48835111"],"author":"kev009","comment_text":"Grok build already punched above its weight and is the nicest TUI, claude and codex are clearly vibecoded by web developers that don't understand systems (eating SSDs, spaghetti logic, extinguishing kernel watch limits, etc). I think Anthropic and OpenAI are both engaged in their own theatre, trying to define and redefine what game they are even playing, trying to shift to immeasurables like safety or security or exclusivity. There's definitely room at the top.","created_at":"2026-07-08T19:30:24Z","created_at_i":1783539024,"objectID":"48836377","parent_id":48836079,"story_id":48835111,"story_title":"Grok 4.5","story_url":"https://x.ai/news/grok-4-5","updated_at":"2026-07-11T16:27:52Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"verditelabs"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["codex","ssd"],"value":"It has gotten harder, unfortunately. One of the barriers to entry is simply the massive amounts of data; not everyone can set aside $100s worth of HDD or SSD space to play around. That said I have done a lot of work to dramatically reduce the amount of storage and bandwidth needed.
We unfortunately get a lot of slop submissions, which is unfortunate. I think a _really_ good place to start is simply joining the discord and looking at the data we've published and trying to replicate something or anything really. We understand that not everyone is a researcher that can jump in making awesome immediately applicate submissions.
Granted, that's pretty specifically for people that want to submit for prizes and prize money. Everyone on the team absolutely loves to talk shop and interact with real people with real interest, so if you show it in the discord we are all more than happy to help, engage, fix bugs, gvmive advice, etc.
I would personally love to see more open source and contributed papyrology and translation, musing on difficult readings etc.
For the more technically inclined, testing software, pointing out bugs, and actually running and trying to fix things is a huge positive that we like. We get a lot of slop submissions that are just someone pasting an issue on our GitHub into codex or Claude. We don't want to encourage that. We can do that ourselves."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"An entire Herculaneum scroll has been read for the first time"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://scrollprize.org/firstscroll"}},"_tags":["comment","author_verditelabs","story_48675179"],"author":"verditelabs","comment_text":"It has gotten harder, unfortunately. One of the barriers to entry is simply the massive amounts of data; not everyone can set aside $100s worth of HDD or SSD space to play around. That said I have done a lot of work to dramatically reduce the amount of storage and bandwidth needed.
We unfortunately get a lot of slop submissions, which is unfortunate. I think a _really_ good place to start is simply joining the discord and looking at the data we've published and trying to replicate something or anything really. We understand that not everyone is a researcher that can jump in making awesome immediately applicate submissions.
Granted, that's pretty specifically for people that want to submit for prizes and prize money. Everyone on the team absolutely loves to talk shop and interact with real people with real interest, so if you show it in the discord we are all more than happy to help, engage, fix bugs, gvmive advice, etc.
I would personally love to see more open source and contributed papyrology and translation, musing on difficult readings etc.
For the more technically inclined, testing software, pointing out bugs, and actually running and trying to fix things is a huge positive that we like. We get a lot of slop submissions that are just someone pasting an issue on our GitHub into codex or Claude. We don't want to encourage that. We can do that ourselves.","created_at":"2026-06-26T02:23:40Z","created_at_i":1782440620,"objectID":"48681640","parent_id":48681315,"story_id":48675179,"story_title":"An entire Herculaneum scroll has been read for the first time","story_url":"https://scrollprize.org/firstscroll","updated_at":"2026-06-28T16:02:36Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"jgalt212"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["codex","ssd"],"value":"37 days maybe with Codex?
> "On my machine, after about 21 days of uptime, the main SSD has written about 37 TB," wrote developer Rui Fan, a project management committee member of Apache Flink. "Process/file-level checks show Codex SQLite logs are the main continuous writer
https://www.theregister.com/ai-and-ml/2026/06/23/openai-code..."},"story_title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["ssd"],"value":"F* file system \u2013 file search that reads SSD directly bypassing OS kernel"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/dmtrKovalenko/ffs"}},"_tags":["comment","author_jgalt212","story_48622433"],"author":"jgalt212","comment_text":"37 days maybe with Codex?
> "On my machine, after about 21 days of uptime, the main SSD has written about 37 TB," wrote developer Rui Fan, a project management committee member of Apache Flink. "Process/file-level checks show Codex SQLite logs are the main continuous writer
https://www.theregister.com/ai-and-ml/2026/06/23/openai-code...","created_at":"2026-06-24T21:31:34Z","created_at_i":1782336694,"objectID":"48665867","parent_id":48652972,"story_id":48622433,"story_title":"F* file system \u2013 file search that reads SSD directly bypassing OS kernel","story_url":"https://github.com/dmtrKovalenko/ffs","updated_at":"2026-06-29T20:00:11Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"inventor7777"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["codex","ssd"],"value":"5-10 sessions?! Wow, not sure how you keep track of those.
For all my daily use + Codex Desktop and some local AIs in Ollama/LM Studio (which power my Home Assistant Voice Assistants) I have an Mac Studio M4 Max with 48GB of RAM (Which I bought early last year IIRC)
It's a complete beast and runs GPT-OSS 20B at over 100 tok/sec. The GUI never slows down, and I run Sequoia still, so it is always smooth and polished. I've also never heard the fans.
I got mine for $2250 USD because I am eligible for the student discount. I got the base 512GB SSD because I store most stuff on external SSDs/HDDs anyway."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Ask HN: What computer are you using for AI coding tools?"}},"_tags":["comment","author_inventor7777","story_48498766"],"author":"inventor7777","children":[48498837],"comment_text":"5-10 sessions?! Wow, not sure how you keep track of those.
For all my daily use + Codex Desktop and some local AIs in Ollama/LM Studio (which power my Home Assistant Voice Assistants) I have an Mac Studio M4 Max with 48GB of RAM (Which I bought early last year IIRC)
It's a complete beast and runs GPT-OSS 20B at over 100 tok/sec. The GUI never slows down, and I run Sequoia still, so it is always smooth and polished. I've also never heard the fans.
I got mine for $2250 USD because I am eligible for the student discount. I got the base 512GB SSD because I store most stuff on external SSDs/HDDs anyway.","created_at":"2026-06-12T01:44:00Z","created_at_i":1781228640,"objectID":"48498823","parent_id":48498766,"story_id":48498766,"story_title":"Ask HN: What computer are you using for AI coding tools?","updated_at":"2026-06-12T01:47:15Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"zozbot234"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["codex","ssd"],"value":"> ...I'm not sure why people don't try to write more stuff in C to have more control / speed / less dependences.
Codex CLI is written in Rust, which should give comparable raw performance to C/C++. Of course you can care about the "less dependencies" point but this is somewhat less of a concern on a properly maintained project like Codex. That's not so much "wild, out of control" third-party dependencies and closer to the old ideal of proper software componentry.
> Also there is a lot more to imagine, TUI side. The problem is that most projects all copy what they already saw. For instance I just did this in 20 minutes.
This mockup is really nice and the sidebar display gives you a natural way to expose running multiple thinking flows in parallel, at least if you keep them from stepping on each other's toes with code edits (keep them all in read-only "plan" mode or working on completely separate directories/files). That's not so helpful on a 128GB MacBook where a single agentic flow brings you to thermal/power limits already, but it suddenly becomes useful on other hardware (DGX Spark, Strix Halo, lower-RAM machines with SSD offload, multiple nodes with pipeline parallelism) where you have more compute than you could use for single-stream decode."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"A few words on DS4"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://antirez.com/news/165"}},"_tags":["comment","author_zozbot234","story_48142108"],"author":"zozbot234","comment_text":"> ...I'm not sure why people don't try to write more stuff in C to have more control / speed / less dependences.
Codex CLI is written in Rust, which should give comparable raw performance to C/C++. Of course you can care about the "less dependencies" point but this is somewhat less of a concern on a properly maintained project like Codex. That's not so much "wild, out of control" third-party dependencies and closer to the old ideal of proper software componentry.
> Also there is a lot more to imagine, TUI side. The problem is that most projects all copy what they already saw. For instance I just did this in 20 minutes.
This mockup is really nice and the sidebar display gives you a natural way to expose running multiple thinking flows in parallel, at least if you keep them from stepping on each other's toes with code edits (keep them all in read-only "plan" mode or working on completely separate directories/files). That's not so helpful on a 128GB MacBook where a single agentic flow brings you to thermal/power limits already, but it suddenly becomes useful on other hardware (DGX Spark, Strix Halo, lower-RAM machines with SSD offload, multiple nodes with pipeline parallelism) where you have more compute than you could use for single-stream decode.","created_at":"2026-05-15T10:08:23Z","created_at_i":1778839703,"objectID":"48146736","parent_id":48146669,"story_id":48142108,"story_title":"A few words on DS4","story_url":"https://antirez.com/news/165","updated_at":"2026-05-15T14:14:44Z"},{"_highlightResult":{"author":{"fullyHighlighted":true,"matchLevel":"partial","matchedWords":["codex"],"value":"codex"},"comment_text":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["ssd"],"value":"Completely agree. I would also be surprised if this weren't done at the block level, as ZFS does SSD acceleration, rather than at the file level as the article claims. It's just so much more efficient than caching at the file layer, for not that much more RAM cost (for indexing)."},"story_title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["ssd"],"value":"Fusion Drive: Apple jumps on the SSD cache bandwagon"},"story_url":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["ssd"],"value":"http://www.extremetech.com/computing/138535-fusion-drive-apple-jumps-on-the-ridiculously-overpriced-ssd-cache-bandwagon"}},"_tags":["comment","author_codex","story_4693362"],"author":"codex","children":[4694387],"comment_text":"Completely agree. I would also be surprised if this weren't done at the block level, as ZFS does SSD acceleration, rather than at the file level as the article claims. It's just so much more efficient than caching at the file layer, for not that much more RAM cost (for indexing).","created_at":"2012-10-24T17:12:32Z","created_at_i":1351098752,"objectID":"4693893","parent_id":4693747,"points":null,"story_id":4693362,"story_title":"Fusion Drive: Apple jumps on the SSD cache bandwagon","story_url":"http://www.extremetech.com/computing/138535-fusion-drive-apple-jumps-on-the-ridiculously-overpriced-ssd-cache-bandwagon","updated_at":"2023-09-06T21:23:53Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"chrysoprace"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["codex","ssd"],"value":"There's a point at which these things become Good Enough though, and don't bottleneck your capacity to get things done.
To your point, React, while it has new updates, hasn't changed the fundamentals since 16.8.0 (introduction of hooks) and that was 7 years ago. Yes there are new hooks, but they typically build on older concepts. AWS hasn't deprecated any of our existing services at work (besides maybe a MySQL version becoming EOL) in the last 4 years that I've worked at my current company.
While I prefer pnpm (to not take up my MacBook's inadequate SSD space), you can still use npm and get things done.
I don't need to keep obsessing over whether Codex or Claude have a 1 point lead in a gamed benchmark test so long as I'm still able to ship features without a lot of churn."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"My AI Adoption Journey"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://mitchellh.com/writing/my-ai-adoption-journey"}},"_tags":["comment","author_chrysoprace","story_46903558"],"author":"chrysoprace","children":[46947822],"comment_text":"There's a point at which these things become Good Enough though, and don't bottleneck your capacity to get things done.
To your point, React, while it has new updates, hasn't changed the fundamentals since 16.8.0 (introduction of hooks) and that was 7 years ago. Yes there are new hooks, but they typically build on older concepts. AWS hasn't deprecated any of our existing services at work (besides maybe a MySQL version becoming EOL) in the last 4 years that I've worked at my current company.
While I prefer pnpm (to not take up my MacBook's inadequate SSD space), you can still use npm and get things done.
I don't need to keep obsessing over whether Codex or Claude have a 1 point lead in a gamed benchmark test so long as I'm still able to ship features without a lot of churn.","created_at":"2026-02-06T04:26:30Z","created_at_i":1770351990,"objectID":"46909112","parent_id":46909024,"story_id":46903558,"story_title":"My AI Adoption Journey","story_url":"https://mitchellh.com/writing/my-ai-adoption-journey","updated_at":"2026-03-05T23:30:56Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"arajnoha"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["codex","ssd"],"value":"hijacking this, what is the best local model (and tool to use it) for programming, if i only have 256gb ssd on a mac? im very used to codex and while i get that it will never be this smart locally, is there any coding model like it, not too heavy on space?"},"story_title":{"matchLevel":"none","matchedWords":[],"value":"LM Studio 0.4"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://lmstudio.ai/blog/0.4.0"}},"_tags":["comment","author_arajnoha","story_46799477"],"author":"arajnoha","comment_text":"hijacking this, what is the best local model (and tool to use it) for programming, if i only have 256gb ssd on a mac? im very used to codex and while i get that it will never be this smart locally, is there any coding model like it, not too heavy on space?","created_at":"2026-01-29T15:49:20Z","created_at_i":1769701760,"objectID":"46811768","parent_id":46799477,"story_id":46799477,"story_title":"LM Studio 0.4","story_url":"https://lmstudio.ai/blog/0.4.0","updated_at":"2026-03-05T23:26:08Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"dudeinhawaii"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["codex","ssd"],"value":"I think it's hard because it's quite artistic and individualistic, as silly as that may sound.
I've built "large projects" with AI, which is 10k-30k lines of algorithmic code and 50k-100k+ lines of UI/Interface.
I've found a few things to be true (that aren't true for everyone).
1. The choice of model (strengths and weaknesses) and OS, dramatically affect how you must approach problems.
2. Being a skilled programmer/engineer yourself will allow you to slice things along areas of responsibility, domains, or other directions that make sense (for code size, context preservation, and being able to wrap your head around it).
3. For anything where you have a doubt, ask 3 or more models -- have them write their findings down in a file each -- and then have 3 models review the findings with respect to the code. More often than not, you march towards consensus and a good solution.
4. GPT-5-Codex via OpenAI Codex CLi on Linux/WSL was, for me, the most capable model for coding while Claude is the most capable for quick fixes and UI.
5. Tooling and ways to measure "success" are imperative. If you can't define the task in a way that success is easy to define -- neither a human nor AI would complete it satisfactorily. You'll find that most engineer tasks are laid out in very "hand-wavy" way -- particularly UI tasks. Either lay it out cleanly or expect to iterate.
6. AI does not understand the physical/visual world. It will fail hard on things which have an implied understanding. For instance, it will not automatically intuit the implication of 50 parallel threads trying to read from an SSD -- unless you guide it. Ditto for many other optimizations and usage patterns where code meets real-world. These will often be unique and interesting bugs or performance areas that a good engineer would know straight out.
7. It's useful to have non-agentic tools that can perform massive codebase analysis for tough problems. Even at 400k tokens context, a large codebase can quickly become unwieldy. I have built custom python tools (pretty easy) to do things like "get all files of a type recursively and generate a context document that will submit with my query". You then query GPT-5-high, Claude Opus, Gemini 2.5 Pro and cross-check.
8. Make judicious use of GIT. The pattern doesn't matter, just have one. My pattern is commit after every working agentic run (let's say feature). If it's a fail and taking more than a few turns to get working -- I scrap the whole thing and re-assess my query or how I might approach or break down the task.
9. It's up to you to guide the agent on the most thoughtful approaches -- this is the human aspect. If you're using Cloud Provider X and they provide cheap queues then it's on you to guide your agent to use queues for the solution rather than let's say a SQL db -- and it's on you to understand the tradeoffs. AI will perhaps help explain them but it will never truly understand your business case and requirements for reliability, redundancy, etc. Perhaps you can craft queries for this but this is an area where AI meets real world and those tend to fail.
One more thing I'd add is that you should make an attempt to fix bugs in your 'new' codebase on occasion. You'll get an understanding for how things work and also how maintainable it truly is. You'll also keep your own troubleshooting skills from atrophying."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Vibe Code Warning \u2013 A personal casestudy"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/jackdoe/pico2-swd-riscv"}},"_tags":["comment","author_dudeinhawaii","story_45874987"],"author":"dudeinhawaii","comment_text":"I think it's hard because it's quite artistic and individualistic, as silly as that may sound.
I've built "large projects" with AI, which is 10k-30k lines of algorithmic code and 50k-100k+ lines of UI/Interface.
I've found a few things to be true (that aren't true for everyone).
1. The choice of model (strengths and weaknesses) and OS, dramatically affect how you must approach problems.
2. Being a skilled programmer/engineer yourself will allow you to slice things along areas of responsibility, domains, or other directions that make sense (for code size, context preservation, and being able to wrap your head around it).
3. For anything where you have a doubt, ask 3 or more models -- have them write their findings down in a file each -- and then have 3 models review the findings with respect to the code. More often than not, you march towards consensus and a good solution.
4. GPT-5-Codex via OpenAI Codex CLi on Linux/WSL was, for me, the most capable model for coding while Claude is the most capable for quick fixes and UI.
5. Tooling and ways to measure "success" are imperative. If you can't define the task in a way that success is easy to define -- neither a human nor AI would complete it satisfactorily. You'll find that most engineer tasks are laid out in very "hand-wavy" way -- particularly UI tasks. Either lay it out cleanly or expect to iterate.
6. AI does not understand the physical/visual world. It will fail hard on things which have an implied understanding. For instance, it will not automatically intuit the implication of 50 parallel threads trying to read from an SSD -- unless you guide it. Ditto for many other optimizations and usage patterns where code meets real-world. These will often be unique and interesting bugs or performance areas that a good engineer would know straight out.
7. It's useful to have non-agentic tools that can perform massive codebase analysis for tough problems. Even at 400k tokens context, a large codebase can quickly become unwieldy. I have built custom python tools (pretty easy) to do things like "get all files of a type recursively and generate a context document that will submit with my query". You then query GPT-5-high, Claude Opus, Gemini 2.5 Pro and cross-check.
8. Make judicious use of GIT. The pattern doesn't matter, just have one. My pattern is commit after every working agentic run (let's say feature). If it's a fail and taking more than a few turns to get working -- I scrap the whole thing and re-assess my query or how I might approach or break down the task.
9. It's up to you to guide the agent on the most thoughtful approaches -- this is the human aspect. If you're using Cloud Provider X and they provide cheap queues then it's on you to guide your agent to use queues for the solution rather than let's say a SQL db -- and it's on you to understand the tradeoffs. AI will perhaps help explain them but it will never truly understand your business case and requirements for reliability, redundancy, etc. Perhaps you can craft queries for this but this is an area where AI meets real world and those tend to fail.
One more thing I'd add is that you should make an attempt to fix bugs in your 'new' codebase on occasion. You'll get an understanding for how things work and also how maintainable it truly is. You'll also keep your own troubleshooting skills from atrophying.","created_at":"2025-11-11T18:01:33Z","created_at_i":1762884093,"objectID":"45890629","parent_id":45880919,"story_id":45874987,"story_title":"Vibe Code Warning \u2013 A personal casestudy","story_url":"https://github.com/jackdoe/pico2-swd-riscv","updated_at":"2026-03-05T23:01:53Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"diggan"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["codex","ssd"],"value":"> Can you imagine if Excel worked like this?
I mean, why would I imagine that? Who would want that? It's like the argument against legal marijuana, and someone replies "But would you like your pilot to be high when flying?!". Right tool for the right job, clearly when you want 100% certainty then LLMs aren't the tool for that. Just because they're useful for some things don't mean we have to replace everything with them.
> Also, each try costs money!
I guess you're using some paid API? Try a different way then. I mostly use the web UI from OpenAI, or Codex lately, or ran locally with my own agent using local weights, neither is "each try costs money" more than writing data to my SSD is costing me money.
It's not a holy grail some people paint it, and not sure we're across the "productivity threshold" (https://news.ycombinator.com/item?id=44160664) yet, but it's worth trying it out probably before jumping to conclusions. But no one is forcing you either, YMMV and all that."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Cloudlflare builds OAuth with Claude and publishes all the prompts"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/cloudflare/workers-oauth-provider/"}},"_tags":["comment","author_diggan","story_44159166"],"author":"diggan","comment_text":"> Can you imagine if Excel worked like this?
I mean, why would I imagine that? Who would want that? It's like the argument against legal marijuana, and someone replies "But would you like your pilot to be high when flying?!". Right tool for the right job, clearly when you want 100% certainty then LLMs aren't the tool for that. Just because they're useful for some things don't mean we have to replace everything with them.
> Also, each try costs money!
I guess you're using some paid API? Try a different way then. I mostly use the web UI from OpenAI, or Codex lately, or ran locally with my own agent using local weights, neither is "each try costs money" more than writing data to my SSD is costing me money.
It's not a holy grail some people paint it, and not sure we're across the "productivity threshold" (https://news.ycombinator.com/item?id=44160664) yet, but it's worth trying it out probably before jumping to conclusions. But no one is forcing you either, YMMV and all that.","created_at":"2025-06-02T17:36:05Z","created_at_i":1748885765,"objectID":"44161193","parent_id":44160920,"story_id":44159166,"story_title":"Cloudlflare builds OAuth with Claude and publishes all the prompts","story_url":"https://github.com/cloudflare/workers-oauth-provider/","updated_at":"2025-06-03T16:49:00Z"}],"hitsPerPage":20,"nbHits":68808,"nbPages":50,"page":0,"params":"query=codex+ssd&advancedSyntax=true&analyticsTags=backend","processingTimeMS":37,"processingTimingsMS":{"_request":{"roundTrip":20},"afterFetch":{"format":{"highlighting":1,"total":1},"merge":{"mergeLoop":{"prepareNextHit":1,"total":1},"total":1},"total":2},"fetch":{"query":3,"scanning":30,"total":34},"total":37},"query":"codex ssd","serverTimeMS":39}