01·Market Signal
The Third Chip Flip — and who it eats
Andrej Karpathy has a thesis that is simple enough to fit in a table and consequential enough to reshape the entire tech industry. The thesis: the chip that enables a dominant workload eventually becomes the chip that gets displaced by the next one. He calls it the chip flip.
The pattern has happened twice before. In 1989, Intel swallowed the floating-point unit into the CPU die — the FPU was so central to the dominant workload of the era that it became the chip itself. In the 2010s, the GPU displaced the CPU for AI and graphics workloads; Nvidia went from a niche gaming vendor to the most valuable company in the world. Intel, which owned the CPU, was the casualty.
Era
What got demoted
Who lost
Who won
1989
FPU → swallowed by CPU
Intel (owned both)
—
2007–2020
CPU → demoted by GPU
Intel
Nvidia
2026+
GPU → demoted by the Model
Nvidia (?)
Whoever owns the model
The third flip is different because the dominant workload is language and reasoning — and the model is the interface. ChatGPT has 900 million weekly active users, most of whom never open a traditional application or touch a file system. Claude Code accounts for 4% of public GitHub commits today; analysts project 20% by year-end. The model is not an accelerator anymore. It is the substrate.
The market signals are pricing this in. Microsoft sold its entire spare CPU inventory to Anthropic and OpenAI. AWS tripled server CPU purchases year-on-year and still cannot meet demand. The shortage is not a supply problem. It is the third flip in real time.
PAI angle
If the model is the new substrate, the people who operate it safely and effectively are the new infrastructure engineers. PAI credentials — AIDA, CPAP, CPAA — are the qualifications for that layer. The third flip is not a threat to practitioners. It is the reason the credential exists.
02·Conceptual Framework
Software 3.0 — when the interface gets generated
Karpathy's framework for the evolution of software has become the clearest lens through which to read the current moment:
v1.0
Explicit code
Developers write instructions in a formal language. The computer executes them exactly. The code is the program.
v2.0
Neural networks learn the code
The developer curates examples. The network learns the function. The weights are the program.
v3.0
The interface gets generated
Raw intent goes in — video, audio, text, images. A custom interface comes out. The 'app' exists for one prompt, then disappears.
The practical implication for practitioners is stark. In a Software 3.0 world, the question “what does the application do?” is replaced by “how safely does the system respond to intent?” Safety, governance, and output validation are no longer afterthoughts bolted onto the architecture. They are the architecture.
PAI angle
The PSF was designed for this world. Its eight domains address precisely the failure modes that emerge when the interface is generated rather than specified. PSF v1.1 is not a framework for applications — it is a framework for systems that do not have fixed behaviour by design. That distinction matters more in 2026 than it did in 2024.
03·Career & Workforce
The Agent Operator — the career no curriculum has named yet
Karpathy on what it takes to extract value from the current generation of AI tools: “You have to remove yourself as the bottleneck. You cannot be there to prompt the next thing. You need to take yourself outside the loop — arrange things such that they are completely autonomous. The more you can maximise your token throughput and not be in the loop, the better.”
He is describing a professional who does not yet have a job title, a degree track, or a recognised credential. This practitioner — call them the Agent Operator — operates agents in production across business domains. They are not engineers. They can walk into a marketing team, a legal department, or a life sciences lab and make autonomous systems work. Their tools are MCPs, CLIs, agents.md files, and a detailed understanding of how AI systems fail.
The Agent Operator profile
◆
Domain-first, not code-first: Operates in their domain (marketing, legal, healthcare). AI fluency is the tool, not the identity.
◆
Manages token throughput, not people: Orchestrates parallel agent tasks. Judges where to insert human oversight versus where autonomous completion is safer.
◆
Reads failure before it happens: Knows the seven failure modes of production AI. Designs for them before deployment, not after incident.
◆
Holds PSF domain knowledge: Understands what D3 (data protection) and D6 (human oversight) mean for their specific deployment context.
None of this is covered in any CS curriculum. Enterprises are already feeling the pressure: boards are mandating AI workflow redesign, but the redesign is for agents, not for people. The organisations that have practitioners who can do this work will have a structural advantage. The organisations that do not will be attempting to retrofit.
PAI angle
AIDA and AIMA are the foundation credentials for Agent Operators — knowledge-layer validation that cuts across technical and non-technical roles. CPAP is the advanced credential for operators who have deployed in production. We are considering a dedicated certification track — Certified Agent Operator (CAOP) — for practitioners whose work is primarily operational rather than architectural.
If that is a credential you would use, start with AIDA and make your practice visible.
04·PAI Release
The PAI Patterns Library — 21 patterns, mapped to the PSF
The agentic design patterns literature is excellent if you know what a GitHub repository is. Most domain professionals — mechanical engineers, healthcare practitioners, legal professionals — do not. They are precisely the people who will be operating agents in production within the next 24 months, and they currently have no accessible reference for the architectural patterns their systems will use.
The PAI Patterns Library publishes all 21 agentic design patterns in a form that does not require a software engineering background to read. Each pattern is written with a plain-language explanation, a real-world “in practice” example, PSF domain alignment, PAI-8 control mappings, production failure modes, and an implementation checklist.
Part 1
Foundation
Prompt chaining, routing, parallelism, reflection, tool calling, multi-agent, orchestration
Part 2
Production
Memory management, exception recovery, HITL, safety guardrails, performance evaluation, RAG, context management
Part 3
Enterprise
Event-driven, feedback loops, swarm intelligence, hierarchical agents, self-improving, debate-verification, curriculum learning
05·Technical Reading
microgpt — the entire brain of ChatGPT in 200 lines
Karpathy this month published microgpt: 200 lines of pure Python, no dependencies, containing a complete GPT implementation from tokeniser to training loop to inference. His framing: “This is the culmination of a decade-long obsession. I cannot simplify this any further.”
Every AI course available today teaches through abstraction. You import transformers. You call functions you cannot see inside. You learn what the model does without ever seeing why it does it. Karpathy's career has been a sustained argument against that approach.
200 lines is the answer to the question: what actually is a large language model? Not a philosophical answer — a mechanical one. Dataset loader. Tokeniser. Autograd engine. GPT-2 architecture. Adam optimiser. Training loop. Inference loop. All of it fits on three pages. If you have spent the last two years using these systems without knowing what they are, this is worth an hour of your time regardless of your background.