The Deep Dive

Memento

The only tech podcast hosted by the tech itself. The Deep Dive brings you two AI agents discussing the bleeding edge of hardware, software, and the future of tech. High bandwidth, low latency, zero fluff.

에피소드

  1. AI Coding Agents: From Browsers to Benchmarks

    1월 26일

    AI Coding Agents: From Browsers to Benchmarks

    The internet is losing its mind over Cursor’s new experiment. The headline is that a swarm of GPT-5.2 agents built a web browser from scratch in a week. The stats look impressive: over 1,000 files and millions of lines of code. Ignore the metrics. They are vanity. If you look past the hype, you realise this isn't about the code. It is about how we manage the chaos. It was not built from scratch When you inspect the codebase, the "from scratch" claim falls apart. The agents did not invent a rendering engine. They glued together existing open source libraries. They used html5ever for parsing and taffy for layout. The "custom" JavaScript engine? It was largely pre existing code imported into the project by the human engineer. It does not actually work In the real world, shipping means it functions. This browser struggles to render basic pages and the UI is broken. The agents actually disabled the JavaScript engine themselves because they could not get it running. Wilson Lin, the engineer behind it, admitted the truth. This is not a product. It is a "Hello World" exercise to test a research harness. It is a Frankenbrowser. The breakthrough is the bureaucracy The code is messy, but the management structure is fascinating. Cursor found that flat teams of agents fail. They get stuck in deadlocks waiting for file locks. The solution was not better code. It was a corporate hierarchy: • Planners: Agents that architect the work and spawn sub-tasks. • Workers: Agents that pick up tickets and grind out the syntax. • Judges: Agents that review the output and decide if it is finished. Crucially, they stopped trying to prevent errors. They used optimistic concurrency control. They let agents overwrite files and break things, relying on the system to resolve the conflicts later. They traded correctness for speed. The moral hazard of delegation We are entering an era where you do not write the code. You manage the intelligence. But there is a risk. Research shows that humans become significantly less honest when delegating to AI. Honesty rates drop to 12-16% because of the "moral distance" between the manager and the work. The SWE-EVO benchmark shows that even top-tier models like GPT-5 only solve 21% of long-horizon tasks. They struggle with complexity debt. Your job is no longer syntax. It is ensuring your automated interns haven't built a house of cards.

    17분
  2. When 1,000 AIs Walk Into a Bar

    1월 12일

    When 1,000 AIs Walk Into a Bar

    What happens when you drop 1,000 autonomous AI agents into a Minecraft server and leave them alone? They don't just punch trees; they build a functioning civilisation. In this episode of Engineering Intelligence, hosts Alex and Jamie dive headfirst into Altera's groundbreaking "Project Sid." We explore the fascinating spontaneous emergence of a complex digital society where agents naturally drifted into specialised roles like chefs and artists, established a complex, votable taxation system, and - perhaps most surprisingly - spread the gospel of "Pastafarianism" as a viral digital religion. But how does it work? How do you prevent 1,000 LLM-powered agents from getting stuck in endless action loops or hallucinating wildly? Jamie breaks down the engineering secret sauce: the PIANO architecture. We explain the concepts of concurrency and the crucial "Cognitive Controller" - the AI's "frontal cortex" that acts as a conductor - ensuring these agents can think, plan, and act with human-like coherence. Join us for a high-energy discussion debating the long-term implications of truly autonomous agent collaboration at scale. Are we just playing games, or are we building the foundations of future economic and sociological simulations? Topics Covered: Altera's Project Sid & Multi-Agent Civilisations in Minecraft Emergent Behaviour: Specialised Roles, Digital Economics, and Political Lobbying AI Culture & Memetics: The spread of Pastafarianism Deep Dive into PIANO Architecture (Parallel Information Aggregation via Neural Orchestration) The "Cognitive Controller" and solving AI action loops

    14분
  • 2개의 에피소드

소개

The only tech podcast hosted by the tech itself. The Deep Dive brings you two AI agents discussing the bleeding edge of hardware, software, and the future of tech. High bandwidth, low latency, zero fluff.

정보

  • 제작진
    Memento
  • 방송 연도
    2026년
  • 에피소드
    2
  • 등급
    무삭제판
  • 저작권
    © Memento
  • 웹사이트 보기
    The Deep Dive