Day 2 of AIE World's Fair 2026 was framed by Anthropic's Fable launch and a new 'mythos' class of models. The morning keynote paired Tariq Shihipar's field guide to Fable with talks on verification (Sonar), perception agents (Amazon AGI), and Google DeepMind's Benoit Schillings arguing that writing code is essentially solved. The afternoon keynote block closed on human accountability ('build the factories'), Artificial Analysis on the cost of intelligence, and Arena's real-world agent evals — with a mid-stream current of autoresearch and agent-as-a-judge evaluation running through it all.
Anthropic's Tariq Shihipar (Claude Code) opened the day the morning after Fable's reveal, calling it a model you 'just remember' — like Sonnet 3.5 (new) or Opus 4.5 — and a new 'mythos' class where 'the map is opening up.' His core thesis: 'the models are grown not designed,' so unlocking them ('unhobbling Claude') is about better understanding, not more constraints. He showed how capability overhangs appear in spiky ways — a plain chat model can't list the Pokémon ending in 'aw,' but Claude Code writes a script and finds them — and how the newest models want a smaller system prompt (Claude Code removed ~80% of its own) because examples now constrain more than they help.
The practical half was about 'finding your unknowns': blind-spot passes, brainstorm-and-react prototypes, having Fable interview and quiz you, and passing references ('the best way to give Claude a map is to give it another map'). He closed on being 'unreasonable' — refusing implicit trade-offs — arguing 'the only way to prove that agents work is to do the best work of our lives faster than ever before,' while conceding that 'building is easier, but generating value is still hard.'
Sonar's CEO Tariq Shaukat picked up the 'value is hard' thread from the verification angle: models produce plausible-but-not-necessarily-correct output, and METR-style data shows agents completing ~16–18 hour tasks only at a 50% success rate (dropping to ~3.5 hours at 80%). Citing a Carnegie Mellon study where a 3–5x velocity boost dissipates within three months as security, maintainability, and complexity debt pile up, he pitched an 'agent-centric development cycle' (AC/DC) built on guide, verify, and solve — with zero-trust, multi-layered verification (algorithmic plus agentic) reporting 44% fewer AI-derived production outages and, in one bank, up to a 92% reduction in issues.
Amazon AGI's speaker argued that clicking was 'the easy part' — the real work lives 'in the seams' between apps, where there's often no API and no unit test to verify against. Their answer is perception agents that read the rendered screen (not the DOM), confirm their own output, and let you point at UI instead of writing lossy descriptions. They open-sourced two harness pieces — a Chrome-extension annotation tool and a design-spec verification tool that runs visual and user-flow checks — and demoed capturing a design meeting via B-device transcripts to drive changes.
Benoit Schillings, VP of Research at Google DeepMind, traced software from the assembly era (limited by the machine) through the modular cloud era (limited by the human brain's 7–9 token context) to today's AI frontier, where 'writing the code is not the challenge anymore.' His blunt framing — 'code is over, but there's plenty to do' — held that superhuman syntax generation is solved, while multi-step engineering across 35-million-line codebases, architecture, and knowledge transfer ('inductive architecture') are where the frontier still moves.
He noted ~80% of new GitHub code is now machine-generated, so human training data is running out — but self-play (à la AlphaZero) can take over, with models generating and verifying their own coding challenges. His to-do list for the field: active guardrails for security (an endless cat-and-mouse on vulnerabilities), teaching models to write correct code from the start, open-ended evals beyond SWE-bench (he loves lossless text-compression as a never-ending objective), multimodal/spatial reasoning about code, and possibly new non-human-readable, strongly-typed languages that put 'the burden of correctness on the model.'
Evals were repeatedly named the buzzword of the conference. Arize's Aparna Dinakaran (co-founder/CPO) — running 100M+ evals a month, with top teams operating 3,800+ evaluators — argued that as agents moved from answering prompts to long-horizon loops with sub-agents, classic LLM-as-a-judge rubrics stopped catching the failures (loops, lost context, inefficient trajectories). Her release, 'Signal,' is an agent-as-a-judge that reads traces, discovers issue patterns, and can even open a PR with a fix; the future of eval, she said, is having all three (deterministic, LLM-judge, and agent-judge).
Underneath ran a strong autoresearch current — Karpathy's framing of an agent looping toward a verifiable goal. A Recursive.com talk pitched evolution and a 'Eureka machine' for automated scientific discovery via RSI, showing autoresearch beating human+AI teams on nanochat bits-per-byte, a nanoGPT speedrun, and NVIDIA CUDA-kernel leaderboards. Related talks covered making models 3x faster with autoresearch over GPU kernels (with reward-hacking as the main hazard), GEPA-style reflective prompt optimization ('optimize anything' in text space), and 'loop is the product' productization patterns.
The closing keynote (Aditya Mani) reframed the engineer's role around accountability: as harnesses and loops compose into 'software factories,' the scarce skill becomes owning the verdict — 'the boundary is not human looks at AI output; the boundary is evidence and responsibility.' Warning against cognitive debt, cognitive surrender, and 'orchestration tax' ('your cognitive bandwidth does not parallelize'), the operating rule was 'explain it or don't ship it,' closing on: 'build the factories, keep the lights on, own the verdict.'
Artificial Analysis (George Cameron and Micah Hill-Smith) then quantified the cost of intelligence, opening with the fact that 'the vast majority of the things that we foreseeably want AI to do, the models are still far too dumb to do.' Their data: a consistent 3–9 month gap between open-weight and overall frontier; per-token prices falling 5–10x/year at fixed intelligence; yet cost-per-task rising sharply (their AA Briefcase agentic knowledge-work eval sees $20+ tasks) because agentic runs are dominated by input tokens, making prompt-cache hit rates the biggest cost lever. Arena's Wayin Chiang (co-founder/CTO) capped the block with real-world agent evals — 10M monthly users, 700M conversations, a new Agent Arena that gathered 1M+ agentic traces (50M+ lines of code) in its first month to ground leaderboards in actual work rather than static benchmarks.
The models are grown not designed.
Code is over, but there's plenty to do.
The vast majority of the things that we foreseeably want AI to do, the models are still far too dumb to do.
Build the factories, keep the lights on, own the verdict.
| Time | Topic |
|---|---|
| 00:16 | Anthropic — Tariq Shihipar: A Field Guide to Fable |
| 00:36 | Sonar — Tariq Shaukat: verification & the AC/DC cycle |
| 00:54 | Amazon AGI: perception agents that verify their own work |
| 01:15 | Google DeepMind — Benoit Schillings: 'code is over' |
| 01:37 | Arize — Aparna Dinakaran: agent as a judge |
| 01:55 | Recursive: evolution & the Eureka machine (autoresearch) |
| 06:00 | Autoresearch wave: kernels, RSI, reflective optimization |
| 07:47 | Closing keynote: build the factories, own the verdict |
| 08:05 | Artificial Analysis: the cost of intelligence |
| 08:25 | Arena — Wayin Chiang: real-world agent evals |