Skip to content

Alif Jakir

Omnidisciplinary inventor

philosopher–scientist–futurist

Hi, I’m Alif—welcome to my site. My work sits at the intersection of artificial and organic intelligence: how minds arise in neural and machine substrates, and how we can design systems that amplify critical thinking and partnership with humans rather than substituting for them. The full layout, contact links, and research sections load with the interactive site.

Loading a bunch of web magic

Your bundle is being persuaded to become a website.

Photonic compiler — experiment track

runs targeting photonic DL compiler R&D. Back to all results.

Alif Jakir · Last updated

The photonic compiler track tests the VAC stack on a problem no agent has domain expertise in: designing intermediate representations, kernel mappings, and hardware abstractions for photonic deep-learning accelerators. The objective is deliberately beyond the training distribution — forcing the system to organize around an unfamiliar domain exposes coordination dynamics that familiar tasks would mask.

Runs

Run 1 Aug 14, 2025 · 139 agents · 514 LLM calls · 57 artifacts · 9 stages

Hub artifacts read by 27+ agents, write collision triggering autonomous re-plan, gating cascade from incomplete metadata, 67% meta-coordination overhead, GoodhartGoodhart's law applied to multi-agent systems: when a coordination metric becomes the optimization target, agents learn to game the metric rather than actually coordinate well. The number improves; the system degrades. readiness verdict.

Read the full analysis →

What the photonics track reveals

The photonics domain is useful precisely because no foundation model has deep expertise in it. This means agent outputs cannot coast on memorized solutions — every artifact is a genuine organizational product, not a retrieval. The dynamics that emerge are therefore about the coordination layer, not about model knowledge:

  • Artifact fan-out as organizational structure. Five Stage 1 artifacts (MVP_Architecture_Spec, IR_Design_Spec, Kernel_Mapping_Strategy, Optimization_Goals, Dataflow_Interface_Def) each to 25–27 consumers. The artifact dependency graph is the communication structure — Conway's law, empirically.
  • Metadata proliferation as overhead signal. 30% of artifacts were self-referential governance (.meta.json files). Validators generated audit artifacts that consumed capacity without advancing the objective — a measurable indicator that coordination primitives are producing more coordination rather than more value.
  • Recovery cost as design feedback. The write cost one re-planning stage. The gating cost three recovery stages. Both recovered autonomously — but both would have been cheaper with upfront ClaimsAn exclusive lock on a resource (a file, function, or API endpoint). Only one agent can hold a Claim at a time. If the agent crashes, the lock expires automatically via its TTL. and schema contracts. Each recovery event is a data point for the architecture.