LiveAI interpretability researchWestern cycleAnchors pending

By end of 2028, mechanistic interpretability research produces a reproducible toolkit that explains at least one major frontier-model capability end-to-end?

As of 6 Oct · 22:19 UTCResolves 31 Dec 20283 movers0 pressure moments
Timeline A · Interpretability LandsReality
35%
Crowd 55%Gap +20

By the end of 2028, Anthropic, DeepMind, or OpenAI releases a detailed playbook showing how a specific frontier model produces one major capability—like reasoning through complex code or planning multi-step tasks—by tracing the exact neural pathways and learned features that fire in sequence. Other labs can follow the same method on their own models and get the same results. The toolkit becomes real enough that safety teams use it to predict and catch dangerous model behaviors before deployment.

Timeline B · Capability Outruns UnderstandingReality
65%
Crowd 45%Gap −20

By the end of 2028, interpretability research remains scattered and incomplete. While researchers publish papers on sparse autoencoders, feature circuits, and causal tracing, no single toolkit ties these pieces together into a reproducible end-to-end explanation of how a frontier model does any major task. Each paper solves a narrow piece; the full chain stays opaque. Safety evaluations continue to rely on behavioral testing and red-teaming rather than mechanistic understanding.

crowd 55
The quiet paththe one nobody’s watching

Uncontrolled AI capability incident before 2028 halts frontier research and resets the timeline.

6%
The read
The underlying structure favors Capability Outruns Understanding (65%) over Interpretability Lands (35%) — the signal is real but noisy.

Two trajectories through ai interpretability research: Interpretability Lands and Capability Outruns Understanding. The Signals below describe what would be visible if each trajectory is landing; the Levers list what could bend it. The third timeline — the one neither Interpretability Lands nor Capability Outruns Understanding captures — is where the honest scope-limit sits.

Where reality is heading

A, 80% bandCrowd today
Projected reality support to 31 Dec 2028A rises from 41% to about 52% with an 80% band of 33 to 71. B falls from 53% to about 44%. Quiet path near 4 to 6%. Crowd today: A 55%.100%7550250Crowd today · A 55%A ~52%B ~44%P ~4%NowJan 2027Jan 202831 Dec 2028
Reality crosses toward A in early 2028. Today’s crowd number sits at the top of A’s band, possible but optimistic. View as table

The movers, ranked by leverage

01
Frontier labs' interpretability team sizes and fundingAnthropic, DeepMind, OpenAI, xAI, Meta FAIR all have interp teams, but headcount and seniority vary wildly
leverage 0
Conditional
02
Open-source interpretability communityDetermines whether methods generalize beyond frontier labs
leverage 0
Conditional
03
Scaling behaviorIf interp methods scale sub-linearly with model size, the field falls further behind each generation
leverage 0
Conditional

What would change our mind

Signals toward A
  • Mechanistic interpretability produces a reproducible toolkit — researchers can take a specific capability of a frontier model (planning, deception detection, in-context learning) and explain the computational mechanism in a way other labs can verify
  • Anthropic, DeepMind, OpenAI interp teams publish converging methods
  • at least one capability has an accepted mechanistic explanation
Signals toward B
  • Capability outruns understanding
  • New model generations outpace the interpretability toolkit — each capability advance resets the frontier, and the gap between what models do and what we can explain widens rather than narrows
  • Interpretability stays a research field, not an engineering discipline

Evidence