About this episodeThe video analyzes Steve Yaggi's $1.5 million annual AI software development experiment, concluding that his c…AI summary
The video analyzes Steve Yaggi's $1.5 million annual AI software development experiment, concluding that his custom harness approach resulted in a low-quality game with significant lag and no tangible value. The host argues that bespoke AI harnesses are only viable if they solve expensive problems cheaply, warning against 'AI psychosis' where agents build irrelevant systems like legal constitutions instead of functional software.
Key takeaways 5
Steve Yaggi's custom harness involves 600,000 lines of bash code running in Emacs, utilizing 21 Claude Pro Max subscriptions that grow by two per week, costing approximately $122,000 per month.
The resulting game, 'Wyvern', exhibits approximately 231 milliseconds of input lag and is described as 'slop' with poor tooltips and chunky movement, failing to justify the massive API spend.
Yaggi's agents built an entire legal system (constitution, courts, case law) instead of functional game engineering systems, which the host identifies as a sign of 'AI psychosis' and lack of direction.
The host distinguishes between two failure modes: being too closed off to new tech versus spending heavily on AI without producing valuable outcomes (spending $100k to save $100).
Treating AI agents like people (e.g., telling them to 'take a deep breath') improves performance empirically due to training data correlations, but anthropomorphizing them leads to dangerous ethical distortions and poor engineering practices.
Notable quotes 4AI-generated: wording and quote attribution may be wrong. Use the play link to verify.
“Building large software remains hard. And it always will be because our ambition will forever outstrip the metal.”
▶ 4:41Steve Yaggi's reflection on why AI cannot simply replace the hard work of software engineering.
“Instead, what they have built was an entire legal system complete with a constitution, jurist, prudence, courts, offices, jurisdictions, case law, rulings, registries, ledgers, rosters, and a full-fledged apparatus for running something resembling a manorial state.”
▶ 17:04Description of the useless output generated by Yaggi's AI agents when left without specific direction.
“You can't write articles telling people what the future is going to be when this is what you've produced.”
▶ 20:43Host's critique of Yaggi's claims about the future of AI development given the poor quality of his actual product.
“If you have been building a harness... but you're spending a ton of money and not really saving much time out the other end, you've done the wrong thing.”
▶ 7:06Host's principle for evaluating the success of custom AI harnesses.
Chapters & Sections (9)▼
0:02AI Software Development Costs and Custom Harnesseschapter4
2:14Wheelhouse Harness and AI Development Costs
4:13AI Coding Ambition and Custom Harnesses
6:32AI Efficiency and End-to-End Testing Pitfalls
8:41Agent Testing vs Unit Tests
11:34AI Personhood Debate and Game Developmentchapter
10h 35m⏭ 3
9h 51m
2h 35m⏭ 1
1h 28m
2h 12m
47m⏭ 1
22m⏭ 1
4h 25m⏭ 1
1h 35m⏭ 1
2h 42m
1h 3m⏭ 1
2h 36m
1h 28m⏭ 3
35m⏭ 1
2h 24m⏭ 3
1h 13m⏭ 2
1h 54m⏭ 4
1h 59m⏭ 2
1h 26m
1h 0m⏭ 1
33m
46m
3h 54m⏭ 7
54m⏭ 1
3h 59m⏭ 5
1h 13m
54m⏭ 2
58m⏭ 5
1h 26m
1h 16m⏭ 2
45m⏭ 2