Uploaded October 2026 | Updated November 2026, 7 hours ago
AI saying “I tested it” isn't the same as knowing what it tested.
Chris Lucian built an MCP server that makes an agent work through a red–green–refactor state machine. We look at the project, his Gemma 4 tic-tac-toe demo, and the practical question behind it: does controlling the process lead to better software?
We also discuss the neural-scaling paper behind Dan Grib's video, why its cost extrapolations shouldn't be treated as a universal ceiling, and Alberta Tech's story about AI inventing her biography — then that information reaching people preparing to speak with her.
Tripo AI, make 3D models from text or images: tripo.pxf.io/c/7912812/3975844/55781
Affiliate link: I may earn a commission at no extra cost to you.
CHAPTERS
00:00 Can AI actually follow test-driven development?
00:53 Chris’s state-machine experiment
01:18 Inside Chris’s TDD-MCP project
02:08 The real Gemma 4 tic-tac-toe demo
03:12 The harness, context and tools
04:08 Queued calls and false success
04:56 Does the process create practical value?
06:04 The pull-request review bottleneck
07:16 Will better models make guardrails unnecessary?
08:17 The scaling video Chris references
09:06 Control, context and alignment
10:01 What the primary paper actually studies
10:50 Superposition and ambiguity
11:43 The cost-cube extrapolation
12:30 What changes when hardware improves?
14:05 AI invented Alberta Tech’s biography
15:02 What Chris learned from smaller models
CHRIS'S PROJECT AND DEMO
TDD-MCP: github.com/ChrisLucian-HunterIndustries/TDD-MCP
TDD MCP Server with Gemma4 TicTacToe Kata: youtube.com/watch?v=xoQwpjdE-gc
Chris Lucian: youtube.com/@ChrisLucian
VIDEOS AND RESEARCH DISCUSSED
Dan Grib — MIT Just Found a Hard Limit in LLM Scaling, and Money Won't Fix It:
youtube.com/watch?v=6xQ8LQfkBg4
Superposition Yields Robust Neural Scaling — Yizhou Liu, Ziming Liu and Jeff Gore:
arxiv.org/abs/2505.10465
Paper header cropped and Figure 2 resized for commentary, under CC BY 4.0: creativecommons.org/licenses/by/4.0
Alberta Tech — AI is convinced I'm from Mexico:
youtube.com/shorts/a0k8J0-KluY
youtube.com/@albertatech
Alberta's firsthand post:
substack.com/@albertatech/note/c-331851077
Shacknews Evening Reading, September 10, 2026 — Asif Khan's roundup embedded her post:
shacknews.com/article/150691/evening-reading-september-10-2026
Chris's demo is a small experiment, not a production benchmark. The MCP checks control actions routed through those tools; passing tests still depend on good requirements and useful assertions.
CURRENT ASSET STORE OFFER
Unity's Autumn Sale has selected top sellers at 50% off, plus rotating flash deals up to 85% off:
assetstore.unity.com/?on_sale=true&orderBy=1&aid=1011lkXUB
Flash deals:
assetstore.unity.com/?flashdeals=true&aid=1011lkXUB
Current Publisher of the Week — Matei Ioan Alexandru:
assetstore.unity.com/publisher-sale?aid=1011lkXUB
Check Unity's page for the current selection and offer terms.
Asset Store links are affiliate links. I may earn a commission at no extra cost to you.
AI saying “I tested it” isn't the same as knowing what it tested.
Chris Lucian built an MCP server that makes an agent work through a red–green–refactor state machine. We look at the project, his Gemma 4 tic-tac-toe demo, and the practical question behind it: does controlling the process lead to better software?
We also discuss the neural-scaling paper behind Dan Grib's video, why its cost extrapolations shouldn't be treated as a universal ceiling, and Alberta Tech's story about AI inventing her biography — then that information reaching people preparing to speak with her.
Tripo AI, make 3D models from text or images: tripo.pxf.io/c/7912812/3975844/55781
Affiliate link: I may earn a commission at no extra cost to you.
CHAPTERS
00:00 Can AI actually follow test-driven development?
00:53 Chris’s state-machine experiment
01:18 Inside Chris’s TDD-MCP project
02:08 The real Gemma 4 tic-tac-toe demo
03:12 The harness, context and tools
04:08 Queued calls and false success
04:56 Does the process create practical value?
06:04 The pull-request review bottleneck
07:16 Will better models make guardrails unnecessary?
08:17 The scaling video Chris references
09:06 Control, context and alignment
10:01 What the primary paper actually studies
10:50 Superposition and ambiguity
11:43 The cost-cube extrapolation
12:30 What changes when hardware improves?
14:05 AI invented Alberta Tech’s biography
15:02 What Chris learned from smaller models
CHRIS'S PROJECT AND DEMO
TDD-MCP: github.com/ChrisLucian-HunterIndustries/TDD-MCP
TDD MCP Server with Gemma4 TicTacToe Kata: youtube.com/watch?v=xoQwpjdE-gc
Chris Lucian: youtube.com/@ChrisLucian
VIDEOS AND RESEARCH DISCUSSED
Dan Grib — MIT Just Found a Hard Limit in LLM Scaling, and Money Won't Fix It:
youtube.com/watch?v=6xQ8LQfkBg4
Superposition Yields Robust Neural Scaling — Yizhou Liu, Ziming Liu and Jeff Gore:
arxiv.org/abs/2505.10465
Paper header cropped and Figure 2 resized for commentary, under CC BY 4.0: creativecommons.org/licenses/by/4.0
Alberta Tech — AI is convinced I'm from Mexico:
youtube.com/shorts/a0k8J0-KluY
youtube.com/@albertatech
Alberta's firsthand post:
substack.com/@albertatech/note/c-331851077
Shacknews Evening Reading, September 10, 2026 — Asif Khan's roundup embedded her post:
shacknews.com/article/150691/evening-reading-september-10-2026
Chris's demo is a small experiment, not a production benchmark. The MCP checks control actions routed through those tools; passing tests still depend on good requirements and useful assertions.
CURRENT ASSET STORE OFFER
Unity's Autumn Sale has selected top sellers at 50% off, plus rotating flash deals up to 85% off:
assetstore.unity.com/?on_sale=true&orderBy=1&aid=1011lkXUB
Flash deals:
assetstore.unity.com/?flashdeals=true&aid=1011lkXUB
Current Publisher of the Week — Matei Ioan Alexandru:
assetstore.unity.com/publisher-sale?aid=1011lkXUB
Check Unity's page for the current selection and offer terms.
Asset Store links are affiliate links. I may earn a commission at no extra cost to you.










