Article · 2026-07-07 · 6 moments

[AINews] The Field Guide to Fable

a quiet day lets us digest the world's most significant model launch... to date. ✦ AI generated

GUEST Thariq GUEST Anthropic GUEST Alan Cowen GUEST jon_durbin GUEST Artificial Analysis HOST Shawn Wang (swyx)HOST Alessio Fanelli

timeline · colored by role

Mechanism

The constraints we perceive in a model are often imposed by us, through the harness we put them in and the way we prompt them — so when a new class of model arrives, we should expect to remove or change those harnesses and prompts to elicit behaviors that were previously hidden because we were 'hobbling' the model.

Thariq argues that many apparent limitations of AI models are self-imposed constraints from harnesses and prompting, and that new model generations require 'unhobbling' to reveal capabilities that were previously suppressed.

transcript

Thariq: The constraints on a model are often imposed by US - “the harness we put them in, and the way we prompt them”. Therefore when we encounter a new class of model, we should expect to remove or change those harnesses and prompts in order to elicit new behaviors that you otherwise would never see because you were overly limiting (aka hobbling) the model.

Claim◆

Because Fable-class models are so much more capable, tradeoffs between good, fast, and cheap are no longer necessary constraints — you can demand ambition on all fronts — though building being easy doesn't mean generating real value has gotten any easier.

Thariq contends that with more capable models like Fable, builders should reject the assumption that quality, speed, and cost must be traded off — even though building is now easy, actually generating value remains hard.

transcript

Thariq: Being unreasonable: Demanding good, fast, and cheap results. “Tradeoffs are not real” - because Fable is more capable, you can be more ambitious and not accept tradeoffs. “Building is easy, generating value is still hard”.

Claim

Anthropic is overclaiming by conflating a privileged latent activation pattern with actual consciousness.

Critics including Alan Cowen pushed back on Anthropic's framing, arguing the company conflated a privileged latent activation subspace with consciousness, overstating the philosophical significance of the finding.

transcript

Alan Cowen: But the “consciousness” language was contested: Anthropic’s public framing invited strong pushback. Supporters said the results suggest a functional analog of access consciousness rather than phenomenal consciousness, while critics argued the company was overclaiming by conflating privileged latent activation with consciousness.

Data◆

Claude has a global-workspace-like internal structure — a small subset of activations called 'J-space' — that acts as a privileged representational substrate available for report, modulation, and flexible reasoning, distinct from simple chain-of-thought extraction.

Anthropic's interpretability research identifies 'J-space,' a small subset of Claude's internal activations that appears to function like a global workspace, usable for reasoning, reporting, and modulation.

transcript

Anthropic: Anthropic released research claiming a global-workspace-like internal structure in Claude, centered on a small subset of activations they call J-space. The core claim is not chain-of-thought extraction, but identification of a privileged internal representational substrate that appears available for report, modulation, and flexible reasoning.

rebuts · 1

Claim

Inference, not training alone, is now the whole game — every data pipeline, RL loop, and agent runtime ultimately cashes out as test-time compute, making inference efficiency the key strategic bottleneck for AI progress.

Jon Durbin argues inference efficiency has become the central strategic bottleneck in AI, since all upstream work — data, RL, and agents — ultimately resolves into test-time compute costs.

transcript

jon_durbin: Inference efficiency is increasingly the strategic bottleneck: @jon_durbin argued that inference, not training alone, is now “the whole game,” because every data pipeline, RL loop, and agent runtime ultimately cashes out as test-time compute. That perspective also showed up in lower-level kernel work: Chutes reported major speedups for MiniMax MSA and GatedDeltaNet-2.

Data

On a new independent AutomationBench leaderboard spanning 657 tasks across 40 simulated SaaS apps, Claude Fable 5 leads with 48.6% accuracy, narrowly ahead of Opus 4.8 at 48.5%, while every model tested still breaks business-rule guardrails at some rate.

Artificial Analysis's independent AutomationBench-AA leaderboard shows Claude Fable 5 narrowly leading over Opus 4.8, with Gemini and GPT-5.5 further behind, and notes every model still violates business-rule guardrails.

transcript

Artificial Analysis: AutomationBench-AA adds a more realistic agent eval: @ArtificialAnlys launched an independent leaderboard for Zapier’s AutomationBench, evaluating agents across 657 tasks and 40 simulated SaaS apps with both objectives and guardrails. Claude Fable 5 led at 48.6%, narrowly ahead of Opus 4.8 at 48.5%, with Gemini 3.5 Flash at 42.6% and GPT-5.5 xhigh at 42.1%.

Highlight slides

Reject the Good/Fast/Cheap Tradeoff✦ from: Because Fable-class models are so much more capable, tradeoffs between good, fast, and cheap are no longer necessary constraints — you can demand ambition on all fronts — though building being easy doesn't mean generating real value has gotten any easier.Building Easy, Value Still Hard✦ from: Because Fable-class models are so much more capable, tradeoffs between good, fast, and cheap are no longer necessary constraints — you can demand ambition on all fronts — though building being easy doesn't mean generating real value has gotten any easier.J-space: A Global Workspace Inside Claude✦ from: Claude has a global-workspace-like internal structure — a small subset of activations called 'J-space' — that acts as a privileged representational substrate available for report, modulation, and flexible reasoning, distinct from simple chain-of-thought extraction.Why It's Different✦ from: Claude has a global-workspace-like internal structure — a small subset of activations called 'J-space' — that acts as a privileged representational substrate available for report, modulation, and flexible reasoning, distinct from simple chain-of-thought extraction.

Related episodes

Article · Import AI

Import AI 461: "Alignment is not on track"; FrontierCode; and synthetic research interns

→

Article · Ahead of AI

Using Local Coding Agents

→

Article · Import AI

Import AI 463: Self-improving robots; a 10k Chinese GPU cluster; and an elegiac essay for the human era

→