Dexmac

@dexmac.bsky.social

🎓 Ph.D. | 🔧 Vintage computers | 🤖 AI | 📸 Photography Restoring/Modding. Building AI tools. AI Research. AI Benchmarks. Exploring the intersection of AI and visual arts. IG: https://www.instagram.com/dexmac221/ Buy me a Coffee ☕: ko-fi.com/dexmac

I was eight when a VIC-20 arrived in my house, ten when a C64 replaced it. This week I watched an AI build games for that same machine — and the one bug all its instruments missed, my eyes caught instantly. Forty years later, it turns out, that kid is still on the job. medium.com/towards-arti...

Claude Fable vs. the Commodore 64: Two Rungs Up the Benchmark Ladder

A Boulder Dash tribute that plays itself through the emulator’s monitor port, then a pixel-smooth scrolling shoot-em-up with raster…

medium.com

I doubled my used-GPU budget (€600→€1000) and my small LLMs got slower. The pricier 5060 Ti has less bandwidth than the 4070, and decode is bandwidth-bound. What the extra VRAM buys is capacity: the 35B fits fully and holds 104 tok/s to 128K context. Measured writeup medium.com/ai-advances/...

Running a 35B LLM at 128K Context, Full Speed, on €870 of Used Hardware — No Cloud Required

How a €870 used-GPU box runs a 35B model at full speed all the way to 128K context — and what I learned measuring the upgrade, including…

medium.com

Science fiction promised us an #AI that wakes up already inside a finished body.What's happening on my desk is the opposite: a home AI growing a body one actuator at a time I wrote about the engineering,and about which doors you keep shut on purpose.Out now in #TowardsAI medium.com/towards-arti...

Brick by Brick: How My Home AI Is Growing a Body

For about two months, the AI running in my house has been less like an assistant and more like a presence — it remembers across days, it…

medium.com

Science fiction promised us an #AI that wakes up already inside a finished body.What's happening on my desk is the opposite: a home AI growing a body one actuator at a time I wrote about the engineering,and about which doors you keep shut on purpose.Out now in #TowardsAI medium.com/towards-arti...

Brick by Brick: How My Home AI Is Growing a Body

For about two months, the AI running in my house has been less like an assistant and more like a presence — it remembers across days, it…

medium.com

I never wrote a hide-and-seek function. I said the word, and the AI I built in my house knew what to do: it moved its cameras, reasoned by elimination, and found me behind a cabinet. Nobody scripted it. It emerged from the parts. New essay, honest and hype-free: medium.com/@gianlucabai...

The Day I Played Hide-and-Seek With an AI

I never wrote a hide-and-seek function. I said the word, and the system I’d built in my house knew what to do — moving its cameras…

medium.com

Ran 35B LLMs on €990 of secondhand gaming hardware. RTX 4070 + 2070 SUPER, 20GB combined VRAM. Gemma 4 26B at 83 tok/s, Qwen3.6 35B at 73. Locally, ~€0.22 per million tokens. But the best result is the one that almost fooled me 🧵 medium.com/ai-advances/... #LocalLLM #LLM #AI #Gemma4 #vram

Running 26B and 35B LLMs at Full Speed on €990 of Used Hardware — No Cloud Required

How a secondhand gaming PC keeps pace with a single RTX 3090 for local AI inference — and what I learned measuring it honestly, including…

medium.com

Just published a benchmark of Google's Gemma 4 across 5 GPU configurations, from a 2016 Titan X to dual RTX A6000s (€20k of pro hardware). For single-user local inference, a used RTX 3090 wins. 🧵 medium.com/ai-advances/...

Gemma 4 on Old GPUs: Why a $700 Used Card Beats $20,000 of Professional Hardware

I benchmarked Google’s Gemma 4 across five GPU configurations spanning four hardware generations — from a decade-old Titan X to dual RTX…

medium.com

🚀 New on Towards AI: The Socratic Model: Why the Wisest AI is the One That Knows When to Ask. Moving beyond monolithic LLMs: A 3B-parameter hierarchical system that outperforms generalists through selective delegation. medium.com/towards-arti... #AI #LLM #Research #Agents #GenerativeAI

The Socratic Model: Why the Wisest AI is the One That Knows When to Ask.

Moving beyond monolithic LLMs: A 3B-parameter hierarchical system that outperforms generalists through selective delegation.

medium.com

Coding isn't dying. The "code scribe" is. Senior devs accept MORE AI code, not less. The real skill now is orchestration: breaking problems, delegating to AI agents, evaluating output. Field guide from the R&D trenches 👇 medium.com/ai-advances/... #AI #SoftwareEngineering #LLM #FutureOfWork

The Death of Coding is an Illusion: A Field Guide to the AI Orchestration Era

Stop treating LLMs like magic oracles. Start engineering them like distributed systems.

medium.com

AI writes C64 code → compiles → injects into VICE emulator → game plays itself → VLM screenshots and analyzes the screen → AI fixes bugs → repeat. No human in the loop! Tested with Claude 4.6 building Arkanoid from scratch. 🔗 ai.gopubby.com/claude-4-6-a... #AI #Commodore64 #Claude #AIAgents

Claude 4.6 and the Commodore 64: When an LLM Writes, Builds, and Playtests Its Own Game

I gave an AI agent a 1982 computer, a cross-compiler, and an emulator. It read the docs, picked its own project, wrote an Arkanoid clone in…

ai.gopubby.com