Day 1 of AIE World's Fair 2026 set the year's frame — software factories built by stacking loops. The mainstage arced from Swix's opening thesis through knowledge-grounding, open agentic stacks and new frontier models, into orchestration at scale, then landed on a candid reckoning with the limits of pure automation and a preview of recursive model self-improvement.
Swix (Shawn Wang) opened the World's Fair by naming the year's organizing idea: Loopcraft. AI engineering, he argued, is fundamentally about stacking loops — and the skill is knowing when to move up a loop for scale or down a loop for reliability. He extended the metaphor from code to human life and civilization, noted that agents are now generalizing well beyond coding into vertical domains, and celebrated the event's growth to roughly 7,000 attendees.
He closed the opening by calling the World's Fair itself the highest loop — "the loop that makes loops" — where humans gather to figure out what the next loop should be. That framing, software factories assembled from nested loops, ran through nearly every talk that followed.
Microsoft's Pablo Castro (CVP & Distinguished Engineer) took the knowledge problem apart along three axes: intrinsic knowledge baked into model weights, extrinsic knowledge that agents retrieve to ground themselves, and learned knowledge that continuously improving agents capture over time.
He traced the exponential arc from IntelliSense (1996) through ML ranking and GitHub Copilot to today's agents, then demoed how Microsoft Foundry, Microsoft IQ and Azure AI Search deliver company-wide grounding and agentic retrieval — closing on a real learning loop in which Foundry's agent optimizer hill-climbs an agent's own instructions against an evaluation.
OpenAI's Alexander Embiricos and Roman Huitt argued that AI is triggering a return to engineering's roots rather than replacing engineers — "AI engineers are eating the world." They framed Codex as a layered, open stack (model, responses API, open-source harness, apps server, app-layer plugins) that OpenAI runs internally exactly as it ships, and previewed the GPT 5.6 series across cost-efficient (Terra, Luna), fast (Codex Spark, Cerebras at ~750 tok/s) and frontier variants — closing on "value maxing" over "token maxing."
The open-weight story carried the middle of the day. Z.ai's Zishan Lee dialed in remotely to introduce GLM 5.2 (and GLM 4.2), positioning its coding and agentic capability between Opus 4.7 and 4.8, adding a new "high" thinking level, making the case for open weights, and revealing Zcode, Z.ai's own Codex-like harness. Hugging Face's Thomas Wolf then mapped the open-model landscape — GLM near the top of Artificial Analysis alongside the Chinese "AI dragons" (DeepSeek, Moonshot/Kimi, Z/GLM, MiniMax) — before a fireside with MiniMax's Olive on M3: a compact ~400B-total / ~20B-active model with native multimodality and a functional ~1M-token context powered by MiniMax Sparse Attention.
Peter Steinberger — "the claw father," of OpenClaw, now at OpenAI — argued that juggling 10+ terminal windows was a false peak of productivity in which the human was really acting as scheduler, router and memory. He described the shift toward managing a single long-lived manager agent that delegates to a team of workers, enabled by server-side compaction, coordination and automated triggers.
As agents improve, he noted, the human bottleneck moves from tokens to compute to attention — so the real work becomes deciding where to spend it. His closing thesis: "the future is not 20 terminals, it's better loops."
The day's sharpest tension was a pushback on unbounded automation. HumanLayer's Dex Horthy argued that the race to build fully automated software factories is producing slop, because reinforcement learning rewards making tests pass rather than maintaining codebase quality — a model-training problem no amount of harness engineering can fix. His answer was to "turn the lights back on": use AI-assisted product review, architecture and vertical slices up front so humans can still read every line.
Linnet's Labs' Eric Meyer sharpened the safety angle: agents become intrinsically dangerous the moment you grant tool calls, and alignment baked into weights is not a real guarantee. His proposal treats safety as a programming-language problem — have the model emit a program (an expression of type IO) you can statically analyze, taint-check and formally prove safe before running, a repackaging of 1990s proof-carrying code. Cursor's Lee Robinson then closed the day by showing the loop turning on itself: Composer is trained through nested outer (real-world feedback, A/B tests) and inner (evals, hard verifiable problems) flywheels, with the top model distilling faster derivatives — a concrete path toward recursive self-improvement.
AI engineering is Loopcraft — you stack loops, and the skill is knowing when to move up a loop for scale or down a loop for reliability.
AI engineers are eating the world — this is a return to engineering's roots, not the end of it.
The future is not 20 terminals, it's better loops.
Harness engineering is not enough — if the model is trained to make tests pass, more loops just produce faster slop.
| Time | Topic |
|---|---|
| Talk 1 | Swix — Opening Keynote: Software Factories & Loopcraft |
| Talk 2 | Microsoft / Pablo Castro — On AI and Knowledge (agentic retrieval) |
| Talk 3 | OpenAI — The Codex Keynote & the GPT 5.6 series |
| Talk 4 | OpenClaw / Peter Steinberger — Better Loops, Not More Terminals |
| Talk 5 | Z.ai / Zishan Lee — GLM 4.2 / 5.2 and the case for open weights |
| Talk 6 | Hugging Face × MiniMax — Open Models & the M3 Fireside |
| Talk 7 | HumanLayer / Dex Horthy — Harness Engineering Is Not Enough |
| Talk 8 | Linnet's Labs / Eric Meyer — Provably Safe Agents |
| Talk 9 | Cursor / Lee Robinson — Recursive Model Improvement (Composer) |