AutoDev LG Dev Log

A detour day. I found a YouTube video introducing Hermes Agent and jumped on it: “This is it!”

• ...But running it with local LLMs, it was basically useless. The dream lasted a few hours.
• Verdict: a whole day wasted. Meekly back to developing AutoDev LG.
• Meanwhile, while the human was cheating on them, the local LLMs quietly kept at it — 9 attempts, not a single complaint. Good bots.

In between runs, somehow a Mac mini got bought. "I'll need it to make iPhone apps someday," I told myself. ...So far, it hasn't been used once (lol).

Where it all began. "I want to try vibe coding with a local LLM!" — except I had neither the skills nor the time to write code myself. So here was the brilliant plan.

• Pay just $2 to DeepSeek, famous for being cheap, and have it build an app that builds programs that make a local LLM do simple command-line jobs. Have an AI build an app that puts another AI to work. (Confusing, I know.)
• My only GPU back then was a single RTX 5070 (12GB). Compared with today's five-card rig, it was a light start.
• The result: 1,831 API requests and 376,365,019 tokens (about 380 million) in one day. Just before noon it was calling DeepSeek 400 times an hour. And the bill was still only $1.66. Cheap. Scary cheap.
• But even once something worked, every new feature made DeepSeek wipe out the existing ones (a classic regression). Fix it, it's gone, fix it, it's gone — forever.
• I got bored in one day. ...Yet this "adding a feature breaks the old ones" problem is exactly why AutoDev LG later became so obsessed with building under the protection of tests.

GPU upgrade. The official reason: "smoother gaming." Swapped the RTX 5070 for an RTX 5070 Ti, bought on a flea-market app for ¥169,800! (Around here I started to feel I'd lost it a little.)

• And on the same day, unable to resist the urge to "run a bigger LLM locally," I also bought an RTX 5060 Ti 16GB.
• Two GPUs in one day. From here on, the GPUs multiply like mushrooms after rain.

More GPUs. Two more RTX 5060 Ti 16GB cards. (I've lost it.)

• That makes four GPUs. The more VRAM you have, the bigger the models you want to try. This is the entrance to the rabbit hole.

More GPUs (the very next day). Yesterday's two weren't enough, so one more RTX 5060 Ti 16GB. (Completely out of my mind.)

• One 5070 Ti + four 5060 Ti = five in total. But at this point there's no PC that can hold five cards yet. I did things in the wrong order.

The turning point. While I was grumbling, "No matter how many GPUs I add, it's still DeepSeek writing the code...," an email or something told me ChatGPT Plus was free for a month. Signed up on the spot.

• To get revenge for the 7/30 regression hell, this time I teamed up with a smarter partner.
• Talking things over with ChatGPT, the idea of today's AutoDev LG slowly started to take shape. The smart AI builds "the system that builds," and the local LLMs at home build the actual apps — that division of labor came into view that day.

New PC case, and the 80GB fortress is complete. To fit five GPUs, I switched to the big LIAN LI O11D EVO XL case. When the box arrived, I was shocked at how huge it was.

• The motherboard stays the same ASRock B760 Pro RS/D4 WiFi I already had: server boards are out of reach, and this PC is also for gaming.
• Of course there aren't enough PCIe slots. So, with an assortment of riser cables — knowing full well the bandwidth gets narrower — I crammed all five GPUs in. An 80GB VRAM setup, complete.
• From here began the days of having ChatGPT build AutoDev LG and testing it with local LLMs.
• ...But ChatGPT quickly hit its usage limit. The same day, I signed up for Claude Pro too. Run, fail, fix, run again — the long journey begins.