Is Coding Solved? Opus 4.5, "Infinite" Memory, & The End of Scaffolding
In this episode of Next Token, we sit down with McKay Wrigley (Founder of Takeoff AI) to discuss the massive shift in AI coding agents. We break down the "Waymo-like" experience of using Claude Opus 4.5, why complex prompting scaffolding might soon be obsolete, and how Google DeepMind’s new "Titans" architecture could solve the context window problem forever.
We also debate the future of junior developers, the rise of "asynchronous" agents from AWS, and why software engineering is evolving into an orchestration role.
Visualizing The Tech
During the "Nerd Alert" segment, we discuss the bottlenecks of current LLMs. To better understand how models handle memory during generation, this diagram breaks down the specific caching mechanism discussed:
Later, we dive into Google's new paper on "Titans." This architecture proposes a "neural memory" that differs significantly from standard context windows.
Timestamps
00:00 - Intro: The end of "Scaffolding" and the Gen Harness
00:30 - Welcome & Guest Intro: McKay Wrigley (Takeoff AI)
02:50 - Tech Buzz: Google DeepMind’s "Titans" Architecture & Neural Memory
06:20 - Why "Infinite Context" changes everything for coding agents
08:15 - The problem with current AI memory (Is it annoying or helpful?)
12:10 - Is Go or TypeScript the best language for AI Agents?
18:30 - AWS Frontier Agents & The "Autonomous Worker"
19:30 - The Debate: What happens to Junior Engineers?
30:35 - Nerd Alert: What is KV Cache?
32:20 - Deep Dive: The Claude Opus 4.5 Experience
33:00 - The "Waymo" Analogy: From driving to being a passenger
40:40 - Predictions: Where will AI coding be in 6 months?
46:50 - Adapting codebases for Agents (Agents.md vs. Raw Intelligence)
53:10 - Game Time: True or False (Google Code Stats, Amazon Q Savings, & Code Churn)
Key Topics Discussed
Opus 4.5 vs. Sonnet: Why Opus feels like reviewing a Senior Engineer's code rather than fixing a Junior's mistakes.
Google Titans: Moving beyond finite context windows to persistent neural memory.
The "Orchestrator" Shift: How developers are moving from writing syntax to managing swarms of agents.
Code Churn: Why "disposable code" is increasing and why that might actually be a good thing.
Guest Info McKay Wrigley is a builder, teacher, and the founder of Takeoff AI. He is well known for his practical AI coding tutorials and experiments.
Follow McKay on X: https://x.com/McKayWrigley
Hosts
Ryan Carson: https://ryancarson.com
Thorsten: https://thorstenball.com/
Follow Amp: https://x.com/ampcode
#AI #SoftwareEngineering #ClaudeOpus #DeepMind #CodingAgents #NextToken