§03//BLOG
The log.
Half work-journal, half scratchpad. New models, agent experiments, and small things I learned the hard way.
§ log.016 — the memory eaters
AI data centers now consume 70% of global memory output. Console prices have doubled, studios are closing, and games that will never be announced are being killed at the pitch stage. All because of RAM. Meanwhile China, locked out of the memory boom by US export controls, is building cheaper models and a domestic chip supply chain. The United States is cannibalizing its own consumer economy to feed an arms race it might not win.
§ log.015 — the missing question→
Open source just caught the frontier and KV cache compression keeps breaking the memory wall. The next milestone isn't a smarter model — it's a curious one. Recursive self-improvement won't come from better answers. It'll come from a machine learning to ask questions. That's also what makes it dangerous.
§ log.014 — canirunai got adsense, referral links, and a fresh coat of paint→
Been poking at CanIRunAI again. Adsense is live, referral links to RunPod and Vast.ai are wired in, and the design finally feels finished.
§ log.013 — i'm letting the internet design my album cover→
Just finished a song I've been sitting on for months. Now I'm building a voting chain where strangers decide what the cover art becomes.
§ log.012 — i designed a crewneck for a product nobody uses→
I designed a crewneck for a product with single-digit users and put it on my portfolio. It'll probably never get printed. That's the most honest thing I've built all year.
§ log.011 — debugging yaml at midnight is the real ai experience→
Everyone posts the cool demos. Nobody posts the part where you're squinting at a config file at 11pm trying to figure out why your agent is throwing phantom warnings. That part is the actual job. Also: I'm going to start making videos.
§ log.010 — i finally found my use case→
For a year I chased benchmarks and the 'best' model, like AI was a thing happening to me. Then I started using it as scaffolding for the way my brain actually works — and the use case finally clicked. The politics keep shifting; the use case doesn't.
§ log.009 — Gemma 4 just got 3x faster for free→
Google dropped Multi-Token Prediction drafters for all Gemma 4 models under Apache 2.0. Up to 3x speedup, zero quality loss. And these run on consumer hardware.
§ log.008 — i scored 6 product ideas with a spreadsheet and still haven't found the right one→
When EAC threatened WrapsRL's future, I built a weighted decision matrix across 6 niches, picked one, and shipped within 24 hours. Then I realized the spreadsheet can't tell you if an idea actually hits.
§ log.007 — anthropic's ai courses→
I'm working through Anthropic's free AI engineering courses. Two down, a few to go. Most of the content I already knew, but the structure is clean.
§ log.006 — iMessage agent setup→
Bridged Hermes Agent to iMessage using Sendblue. One afternoon, a few API calls, and now I can text my agent like a friend.
§ log.005 — DeepSeek cut V4-Pro to $0.036/MTok. That's 75% off.→
DeepSeek slashed V4-Pro input pricing by 75% through May 5 and permanently dropped cache-hit pricing to one-tenth across all their APIs. V4-Pro at $0.036/MTok is cheaper per token than V4-Flash.
§ log.004 — DeepSeek V4 dropped today. Benchmarks are wild, tool calling is broken.→
DeepSeek released V4-Pro (1.6T params) and V4-Flash (284B params). The benchmarks look incredible for the price — but I ran Flash through Hermes Agent and it completely falls apart on tool calls.
§ log.003 — OpenAI dropped GPT-5.5. It's fast, agentic, and probably overkill for me.→
GPT-5.5 just dropped. Six weeks after 5.4. Better at coding, autonomous task planning, and using fewer tokens. I don't have a use case for it yet — but the shift is real.
§ log.002 — 27B just ate 397B for breakfast.→
Qwen3.6-27B dropped yesterday. Dense model, open weights, flagship coding. I hooked it up to Hermes Agent.
§ log.001 — starting the log.→
First entry. Why I'm building this site and what comes next.