I have had a pause in the AI tooling topic, just used copilot autocompletion and that was it. And I feel like I was digging a hole with a shovel and then a good neighbor IndyDevDan stopped by and said "hey man I have got an excavator." Sure it costs some gasoline but time, the excitement of being productive has grater value. And honestly I have a peak of productivity which lasts less then 9-5 so I need to use it well. Thank you, good neighbor.
This was really enjoyable to watch. Seeing your workflow process is insightful. Especially that you get specific with Aider and tell it what in the code to focus a change on. It’s absolutely crazy how far AI coding has progressed in a short time
I love these videos, thanks for making them! I appreciate you showing the /token command and showing how much this costs. I had no idea how much this was costing per prompt. It's great to see that even with a couple of files added to the context that it's still just a couple of cents per prompt. I was worried this would end up costing hundreds of dollars in API fees per day.
Dude any coding ai is a huge help I keep running into problems with my setup. I'm trying to make a multi model assistant but I only have like 1 semester of coding knowledge lol
Kudos to you for building something and showing your workflow. It doesn’t seem agentic or something you’d really use though, right? Agentic would be an agent curating information automatically for you, without needing to add specific sites to scrape or push a scrape button. How would an AI system know what’s useful or interesting to you? Probably old school method where you’d “like” or heart updates to teach it. Anyway have a good day and keep building!
Jeremy thanks for the kudos and the super solid comment. I am using this and the definition of 'agentic' is definitely worth a video. I think a good working definition is: if you use probabilistic software (LLM) to make a decision to drive an outcome you can call that agentic which we do in the video via sonnet 3.5. This video is a POC, your version (v5 maybe?) is definitely a great direction as costs for top tier models come down. Balancing real tangible results with autonomous behavior is where all the money is right now. I like to start tangible and add autonomy to find a balancing point. We're on an agentic workflow hot streak right now, I'll demonstrate exactly what I mean in subsequent videos.
Mmmm. Ok so you are very aware of what is coming in the near future. I’ve recently been working on a lot of the systems thinking and product UX around this. Definitely glad to see you building in this direction. This is not just for engineers. Literally for everyone. Let’s face it, there are highly nutritious info calories and highly toxic info calories. Kids especially. Social media is so terrible. RU-vid is too. These are idiotic algos don’t have the end user’s best interests in mind. We are looking at a new internet in coming. Happy to chat further
Nice! Thanks for showing how you work with aider. That’s really cool! Will implement using aider as support soon! By the way with llama 3 70B I get very reliable json output in a production application, just specified in the prompt too.
I liked watching your workflow, just wondering if you've tried using a mechanical keyboard? it may seem trivial, but the tactile feedback of the key clicks does feel like it makes a difference to me. Oh yeah, I'm a huge Aider fan, as much content on that please
> update Workflow: act(): loop through providers, if changes_detected update the config.json file for that provider. ONLY update that one field. reload the config.json file to make sure nothing else changes except that one field for each provider IF 'changes_detected' is true. - seems like you had little faith in your AI companion to understand the instruction and had to reiterate the requirements twice :) IMO this is good demonstration of the key challenge with AI coding. The assistant is supposed to take the hint and make resonable assumptions based on short input, let alone not requiring repeated instructions. Yet that's not the case, often the effort (mental included) to instruct the model creates a barrier to prompt the assistant and you find that you type the code manually.
Tried it already. Decent step up, but still not enough to be used for purposes other than testing. Let's wait for Cursor to implement interpreter mode for Sonnet 3.5 - way better.
Noticed you've returned back to aider in your vids. Did you ever take a pause from it or just not shown in the last few in the series? If you did focus with another assistant, what brought you back (while the assistant is good, and that it now uses claude sonnet even better, for me it's the auto git diffs and updates)?
For a while now, I've been using both Aider & Cursor religiously. Right now, Aider can perform more, higher quality, multi file edits over Cursor. With sonnet 3.5 it has the #1 spot as Best AI Coding Assistant. That being said though, I also use Cursor for quick, inline, single file edits and codebase QA. A lot of engineers fall into the trap of 'one over the other' and 'this over that' - I recommend thinking 'how can I use ANY COMBO of tech to improve my engineering'. If you can duel wield, do it.
Thanks you for the video, if i have project lets say on laravel on my vscode , its can work with it? I mean if i say create a new api ... it can found the correct files to edit? Ty
@@J3R3MI6 ive been watching this guy for over a year. The 'automation' he shows is too error prone and a huge waste of time considering humans do this better and will continue doing it better until a couple of decades when we have true agi
what I know is that the folks who think learning to leverage AI coding assistants isn't valuable only believe this because they haven't taken the time to dig in and do the work. They'll come around eventually because they'll have no choice. Keep this content com coming. It's awesome!