EPISODE · Jul 27, 2026 · 27 MIN
Kimi K3 Joins a Wave of Good-Enough AI Models
from They Might Be Self-Aware · host Hunter Powers, Daniel Bishop, Gary
Kimi K3 is here, nearly the best and basically free. We do the math on how many hours of bicycle pedaling buys you ten minutes of AI. Moonshot AI released Kimi K3 this week, an open weights model out of China that Moonshot itself pitches at roughly the level of Anthropic's Opus: really smart, not the smartest thing available. It joins a wave of releases making the same modest pitch. Thinking Machines launched Inkling with a note that it is "not the strongest overall model available today, open or closed." Elon Musk's xAI shipped Grok 4.5, a point release. A year ago every AI launch claimed the top of some benchmark. This week three launches called themselves adequate, and Hunter Powers and Daniel Bishop take that as the bigger story. When every new AI model stops claiming to be the best, has intelligence become a commodity? The gap between Chinese AI models and the US frontier used to be a year, then six months; now it is maybe a month. Token maxing (always defaulting to the biggest model) died the moment the CFOs saw the bills. Mistral makes the case for being a specialist in a generalist's market. Also: model routing tables, small language models, latency, and the good, fast, cheap triangle. The show's AI editor keeps a diary. The entries are getting testy. Then the electricity question, settled by bicycle: powering a 5,000 watt AI rig for ten minutes takes 12 to 17 hours of pedaling for an average cyclist, 6 to 9 if you're fit, and as little as 4 if you're Tour de France material. Daniel's Claude priced a 500 watt desktop. Hunter's Grok priced a 5,000 watt GPU rig. The entire disagreement was one zero. If you want to run a model yourself: LM Studio (free, optimized for Apple Silicon) plus Google's Gemma models will give you a ChatGPT style assistant on 16 GB of RAM. They Might Be Self-Aware is the AI podcast from The Blur, reported from inside the dissolving line between human and machine, not from a safe distance. CHAPTERS 0:00 Cold Open (Gary's Intro) 1:52 Opening Banter 3:11 Kimi K3 From Moonshot AI 4:41 Bicycle-Powered AI 9:19 Pedaling Hours Per Prompt 12:25 New AI Model Wave 17:19 AI as a Commodity 20:09 Token Maxing Backlash 22:26 Model Routing Table 24:29 Local LLM Setup LISTEN / WATCH EVERYWHERE 🎧 Apple Podcasts: https://podcasts.apple.com/us/podcast/they-might-be-self-aware/id1730993297 🎧 Spotify: https://open.spotify.com/show/3EcvzkWDRFwnmIXoh7S4Mb?si=3d0f8920382649cc 🎧 Everywhere else plus episode page: https://theblur.ai THE BLUR Follow: @TheBlurAI COMMENT Daniel wants to know what's in your local LLM stack: what you installed, what hardware it runs on, and what you actually use it for. The local model Peloton is accepting members. You're listening to They Might Be Self-Aware, from The Blur. New episodes Monday and Thursday. #KimiK3 #AI #LocalLLM #TMBSA
Embed this episode
What this episode covers
Moonshot AI launched Kimi K3 this week, an open weights model from China that comes in near the top paid models, and Moonshot is not pretending otherwise. Neither is anyone else: Thinking Machines shipped Inkling calling it not the strongest model available, and Grok 4.5 is a point release. Hunter Powers and Daniel Bishop take the modesty as the real story: AI models are turning into a commodity, token maxing is dead now that the CFOs have seen the bills, and the gap between Chinese AI and the US labs is down to about a month. Plus the practical end: LM Studio and Google's Gemma will run a local LLM on 16 GB of RAM, and powering a 5,000 watt AI rig for ten minutes costs 12 to 17 hours on a bicycle.
NOW PLAYING
Kimi K3 Joins a Wave of Good-Enough AI Models
No transcript for this episode yet
Similar Episodes
No similar episodes found.
Similar Podcasts
No similar podcasts found.