Jonathan Gray

@ujjwalscript Completely disagree. We were not bottlenecked by these important decisions and technical thoughtfulness. It took too long to put it into action. Not all vibe code is created equal. There have always been bad developers and good developers. Same shit, different day.

Unpopular Opinion: We aren't building the future 10x faster with AI. We are just generating legacy code 10x faster. Everyone is currently bragging about developer velocity. "I built this entire backend in a weekend!" "AI wrote 80% of my codebase!" But here is the reality check we are ignoring: Code is a liability, not an asset. If an AI tool spits out 1,000 lines of functional boilerplate in five seconds, that is still 1,000 lines that a human being has to read, review, secure, and maintain when the dependencies inevitably break next year. We are treating code generation like a pure productivity win, but we are optimizing for the wrong metric. The bottleneck in software engineering was never how fast we could type. The bottleneck has always been comprehension, architecture, and maintenance. If we don't shift our focus from "generation speed" to "architectural sanity," the tech debt of the next five years is going to be an absolute, unmaintainable nightmare.

@Scobleizer @dreamina_ai Try our free little demo at aixmascards.com

I need to make some Christmas cards, and images for sending to family and friends, and Felix caught my eye with how he used @dreamina_ai. So I tried it. Took literally a few seconds to take my step mom's photo and put her into this Christmas scene. I just uploaded a photo to it (actually just copy and pasted out of Apple Photos into Dreamina) and told it what I wanted "put the woman and the dog into a Christmas scene." Took seconds.

Any response from @cursor_ai regarding sudden crashes with no logs or trace on Mac? Happened for the 4th time in 6 weeks today. There are multiple threads out there on it across github and reddit. Any idea at all what's happening or am I being forced to switch?

we're living in the future of dev except everything is expensive and breaks. in the past 8 hours I had a completely silent crash of @cursor_ai (where i pay $200/mo) then a few hours later I got blocked by @AnthropicAI claude code (also $200/mo) and now @github is broken

@Scobleizer I use AI image and video models all day every day for work. I do not agree with or even understand this statement. Yes, you can make art with AI, but it’s better than all of the world’s pre-AI art? 🤔

I've been to a lot of modern art museums around the world and what I'm seeing from the AI artists lately is better than anything I saw in any of the museums. (I have a list of 3,500 here on X).

@Vtrivedy10 @jackmuva @lmstudio Agreed… Frontend work seems to be the only thing cursor+claude still seriously struggles with. I have the best luck with vanilla html/css now 😂

@jackmuva @lmstudio thanks!! also neovim and no plugins…you madman 😂 but I hear you, you try Claude Code? the terminal interface grew on me a bunch, no fluff, that and cursor have been my mains i’m convinced perfect centering divs + perfectly installing PyTorch is true AGI

What I learned coding without any internet + a local LLM on a 7 hour flight ✈️ Setup: I tried rawdogging with no local LLM for ~2 hours, it was pretty brutal lol. Then I used Qwen-2.5 7B with @lmstudio on an M2Pro-32GB. Probably coded 5 hours total. 1. I’m way worse at rawdogging typing code. Having no Tab Autocomplete was the real blocker because I forget a lot of syntax and just stare at the screen. I think AI users’ collective memory for syntax has fallen off a small cliff. 2. I’m way better at thinking in systems/modules compared to pre-LLMs! Managing LLMs as a sort of AI PM over the last 3 years forces you to delegate in modular pieces because LLMs excel when they have that structure. Making a design doc was actually totally fine alone! I usually use an LLM partner but alone I probably thought even deeper. I’m in practice of immediately dumping my ideas into Apple Notes, this probably helps train that muscle. 3. If you only use the best models every day (Claude 4, o3, Gemini-2.5), the downgrade to a local 7B level of intelligence is VERY noticeable, way more logical mistakes even if syntax is right. I was working on a local multimodal search + scene verification system for images/videos. Really this boils down to stitching a set of modules (ex: detect objects, match against prompts, function to track across frames). If I wasn’t deeply familiar with this domain, using a local model would’ve been impossible, ex: logical errors when stitching modules that closed models wouldn’t make. 4. It’s beneficial long term to not “Vibe Code” and blindly Agent Mode accept everything if you want to learn how to build (and most importantly maintain) your system. Even though LLMs are great teachers, I bet that very few Vibe Coders actually dive into the code to learn as they Accept everything. Although the copy-paste method is more friction, it definitely helped me internalize the code more deeply. I will probably go back to it more with a WebUI, the friction is good for learning. 5. The “Blank Canvas problem” is solved by LLMs (you basically always have a starting point). Going back and doing that from scratch felt completely alien, will never do that again and ended up scrapping everything after I switched to local Qwen-2.5 which scaffolded everything for me. I don’t think it makes sense to start from a blank screen for 99% of applications anymore. 6. I use Web Search a lot and I didn’t realize how much of a pain coding is without it. I very quickly became aware how much I use Web Search to augment context both to look up syntax and to find examples for the LLM. Tool Use is only getting more integrated, not having it will be even more burdensome (like having your fave MCP unavailable). Overall I don’t think LLMs have brain rotted my abilities as a programmer (yet) and in many ways I feel way better as an engineer and researcher. Using offline local LLMs for primary development is kind of an eye opening experience if you use the best models and Cursor/Windsurf for everything I’m eager to hear other plane coding stories and if my experience lines up! vacation = over, back to shipping 🫡

My initial excitement for @OpenAI Codex is now gone. It completely fails to do anything moderately complex, ignoring requirements and taking huge shortcuts. Identical prompt into @cursor_ai Agent w/ @AnthropicAI Claude 4 Opus knocked it out of the park on the first shot.

Fact. It’s taken things to an entirely new level and my entire process, from coding to testing to deploying to debugging, is now changed forever.

Today, we’re announcing our $50 million Series B and that Admiral John M. Richardson (USN, Ret.), former Chief of Naval Operations, is joining our Board of Directors. With over $70 million raised to date, we’re accelerating toward full-scale nuclear battery demonstrations in 2026, enabling deliveries to maritime and space customers in 2027. Read more via @demarest_colin in @axios: axios.com/2025/05/14/zen…

@sama @AravSrinivas Sorry, Perplexity Deep Research is not even close to ChatGPT Deep Research. I guess you get what you pay for. Keep rockin @sama

@AravSrinivas since you nicely apologized to me in person for all the mean tweets last week, im going to let this go :) keep cooking out there! proud of you.

we put out an update to chatgpt (4o). it is pretty good. it is soon going to get much better, team is cooking.

Using o1-preview from @OpenAI reminds of using the web on slow dial-up in the 90s

One pair of founders I talked to today were students at a top US university. I asked what percent of students don't use AI to write their papers for them. They said max 20%. Professors have given up trying to forbid it.

THIS IS #HOMEOFBEISBOL 🩵 See you at 4:10pm: marlins.com/tix

@shl encorestudio.ai

Looking to invest between $25K in 188 startups and $4.7M in one startup in 2024. If you could use some capital to accelerate your growth, please DM me: 1. What you’re working on (URL is great) 2. What you’d spend the money on 3. How much you’re looking to raise, at what post-money valuation 4. Your existing cap table Or you can email me: sahil.lavingia@gmail.com Please, no decks!

It is Saturday before Christmas and there are still AI products shipping.

@nmotgi Transcribed lyrics in this case, text only

@jgrayla Are you provide audio files to Gemini ?

Have been playing with Gemini Pro for our core use cases related to music. Unfortunately if there’s even a single curse word in the lyrics, it denies the request with no detail. Shockingly way more strict than GPT3.5/4 and of course OSS models.

This feels like someone has a time machine and identified the exact point in history to fork the timeline

@KiDCuDi I wish it could be shared with everyone on @ClapForEncore ❤️

This is just for the die hards that love those deep cuts that have never been heard, or maybe u were too young to go to a concert when the albums came out

I have been writing code in shell text editors for almost 30 years, mostly resisting IDEs (hey, java 🤷‍♂️). I'm not sure I will ever code again without both an AI typeahead and AI chatbot assistant. This shit is so wild and constantly blows my mind.