The AI Treadmill
Network Chuck’s latest video, "I kind of hate AI," hit a nerve. In it, he admits he nearly quit YouTube. Perhaps tech altogether. It’s a sentiment echoing across the industry: "AI Burnout."
The constant onslaught of news, the fear of displacement, and the sheer speed of change led Chuck to a 3 month sabbatical in Japan just to find clarity. With tens of thousands of IT layoffs directly linked to AI integration, the anxiety is grounded in reality. Yet, despite the "doom and gloom," Chuck’s takeaway isn’t retreat. It’s a shift toward relentless optimism and a return to learning alongside his audience.
Predicting the Future: A Lesson from History
When we feel the ground shifting, it helps to look back. History shows us that transformative tech rarely results in fewer jobs; only different jobs:
The Printing Press: Book copiers feared the end, but the literacy and publishing industries were born.
The Car: Stable boys feared for their livelihood, yet today there are more automotive mechanics than there ever were stable boys.
The Camera: Painters feared obsolescence, but it birthed the entire field of modern visual media.
Even if you were a "diehard stable boy," you could still find work today. Innovation doesn't delete the past; it expands the horizon.
The Software Pivot: Leaving OpenClaw for Hermes
I waited to see if Chuck would follow through on his promise to learn OpenClaw. He did, but my own journey has taken a different turn.
As of today, I no longer run or recommend OpenClaw. The frustration became untenable. Every patch felt like a gamble, frequently breaking in ways that even AI couldn’t diagnose. The GitHub issues are a sea of duplicates with no clear fixes beyond rolling back to previous versions.
The recent "mega patch" was the final straw. It forced me to switch to Hermes. Written in Python and properly architected, Hermes makes patching a dream. It’s done in seconds rather than hours of troubleshooting.
The Evolution of the "Brain"
The choice of model is also shifting. Claude and Gemini have banned third-party access, effectively rendering them worthless for power users who want control. We are seeing a mass exodus of subscribers from these platforms. I couldn’t imagine ever staying with either of them.
Meanwhile, OpenAI, Minimax, and Kimi are actively supporting the new way to use AI. The real excitement is in the generational leaps of how we interact with these models:
1st Gen: The Chat Window (Manual copy-paste).
2nd Gen: Autocomplete (IDE plugins).
3rd Gen: The Agentic Era (AI reads files and edits code directly).
4th Gen: The Orchestrator (You talk to the AI, and it manages the agentic loops for you).
The leap from the 3rd to the 4th generation is massive. We are no longer just using a tool; we are managing a digital workforce.
The 2026 Hardware Landscape
Keeping up is the hardest part. We’ve seen four generations in three years. New techniques like Turboquant (reducing VRAM usage) and Dflash (10x faster speculative decoding) change the math every few months.
The model wars are equally volatile:
Gemma 4 26B: A decent coder, but it struggles to stand out. It’s bizarre because it matches the intelligence of DeepSeek R1, which dominated headlines as one of the bests. Today you have deepseek r1 intelligence available on consumer hardware.
Qwen3.5 27b or Gemma 4 31B: This is the current sweet spot. It’s nearly twice as "smart" as R1, but while DeepSeek requires datacenter hardware, the 31B model can run beautifully on a $2,000 GPU. Though a 5090 is ideal, but $4000.
On the hardware front, the Apple M5 Max offers a staggering 128GB of unified RAM, but the price tag is steep. For those building custom rigs, the Intel B70 with 32GB is an incredible value. If you’re looking at a 4U server with six PCIe x16 slots... well, now we’re talking.
Though I’m waiting. Next year, DDR6 era drops. I’m predicting the AMD Strix Halo successor will have 384bit bus, 192GB of lpddr6 ram. Optionally, even just going a threadripper with 8 channels of DDR6 should be fast as or faster than today’s GPUs. Insanity.
Final Thoughts
Between LM Studio, Ollama, vLLM, and the endless "Claw" variants (Nanoclaw, Nemoclaw), the ecosystem is building itself in real-time. It is exhausting, yes. But if we follow Chuck’s lead of "relentless optimism," we stop seeing the speed as a threat and start seeing it as the ultimate sandbox.
Install Hermes Agent!