June 11 AM: Anthropic ships Fable 5 and unfiltered & Anandkumar's team distills physics

Anthropic released its most capable model alongside a safety-stripped version, revealing a split strategy that risks alienating both cautious enterprises and aggressive researchers.

Contributors Simon Willison: Characterizes Fable 5 as slow, expensive and capable of crunching through complex tasks, while suggesting Mythos 5 is simply Fable 5 with safety classifiers disabled ^[1] Anthropic: Released detailed system cards describing Mythos 5 as Fable 5's capabilities without safety classifiers, available only to approved partners at identical pricing ^[2]

Anthropic released Claude Fable 5, a high-capability reasoning model, alongside Mythos 5—described as Fable 5 without safety classifiers. Mythos is gated behind approved partnerships and priced identically to Fable, suggesting it is technically the same model with guardrails removed rather than a distinct architecture. Simon Willison's testing confirms Fable 5 delivers significant capability but with notable latency and cost penalties, while the dual-release structure creates a two-tier access system where safety constraints are enforced through access control rather than technical limitations.

For founders and builders, this creates a strategic fork. Enterprises seeking reliability face a slow, expensive model; researchers seeking uncensored capabilities face artificial scarcity. The pricing parity suggests Anthropic is not monetizing the dangerous version at a premium, but rather using access control as the primary gatekeeper. This signals a company hedging between safety brand preservation and competitive pressure from open-weight rivals, potentially ceding the uncensored market to less cautious competitors while retaining plausible deniability.

Safety advocates argue that gating Mythos prevents misuse while allowing vetted researchers to probe failure modes. Critics counter that differential access creates a black market for model weights and that safety classifiers often encode political rather than safety preferences, making the restriction paternalistic. The evidence tilts toward Willison's interpretation—Mythos appears to be a capability unlock rather than a distinct model, suggesting Anthropic's safety strategy is increasingly about access control rather than architectural safeguards.

This tension mirrors the broader industry split between closed frontier labs and the open-weight movement, with Anthropic attempting an unstable middle path.

A new framework turns scientific literature into real-time process controllers, potentially eliminating the data bottleneck that has stalled industrial AI adoption.

Contributors Prof. Anima Anandkumar: Co-authored research demonstrating that LLM-extracted physics priors can be encoded into teacher models via Graph-Masked Attention layers and distilled into lightweight student models capable of real-time industrial inference ^[1]

Researchers led by Anima Anandkumar have developed a knowledge distillation pipeline where LLMs extract analytical physics priors from scientific literature, encode them via Graph-Masked Attention layers into a teacher model, then distill this into a lightweight student capable of inference speeds exceeding 6000 Hz. The system maintains predictive performance even when LLM-derived priors are suboptimal or incomplete, and operates effectively with small manufacturing datasets where traditional data-driven approaches fail.

Industrial AI has long stalled on the cold-start problem—rare failure modes and expensive data collection make pure statistical approaches uneconomical for many manufacturers. By grounding predictions in physics textbooks rather than just historical sensor data, this approach could democratize predictive maintenance for mid-sized industrial players. For technical founders, this suggests a viable vertical SaaS play: physics-informed edge models for specific industrial processes like CNC machining or chemical batching, where data is scarce but literature is rich. The 6000 Hz inference rate enables real-time closed-loop control rather than mere monitoring, opening the door to autonomous process optimization.

This development offers a counterpoint to the brute-force scaling debates dominating AI discourse, showing how structured knowledge extraction from scientific literature can outperform raw parameter counts in resource-constrained industrial environments.

A claimed breakthrough in autonomous mathematical discovery is splitting the community between those seeing a new research paradigm and those seeing clever brute force on low-hanging fruit.

Contributors Pushmeet Kohli: Claims an AI agent autonomously resolved 9 out of 353 open Erdős problems and 44 out of 492 OEIS conjectures using formal proof search in Lean, and that the system is actively deployed in real mathematical research ^[1] Anonymous critics: Argue that autonomously resolved obscures heavy human scaffolding in problem formalization, and that solved problems likely represent the most tractable cases rather than genuine mathematical breakthroughs ^[1]

Pushmeet Kohli announced that an LLM-driven formal proof agent resolved 9 open Erdős problems and 44 OEIS conjectures via Lean proof search, claiming the system is actively deployed in real mathematical research across combinatorics, algebraic geometry, and quantum optics. The paper positions this as the first large-scale autonomous mathematical breakthrough, achieved at a cost of only a few hundred dollars per problem, suggesting a new paradigm where mathematicians shift from provers to problem-formalizers.

If verified, this shifts the economics of mathematical research and suggests that formal verification—long considered academic—is becoming critical infrastructure for reliable AI reasoning. However, the specific problem selection matters enormously for assessing scalability. OEIS conjectures vary wildly in depth, from trivial pattern-matching exercises to profound unsolved challenges, and formalizing a problem in Lean requires significant human mathematical judgment, making autonomous a potentially misleading descriptor.

Proponents argue that even easy open problems require combinatorial creativity beyond brute force, and that resolving 9 Erdős problems—regardless of difficulty distribution—validates the method's potential. Skeptics counter that without a difficulty-stratified breakdown or evidence of advancing frontier research in the claimed deployment domains, the results may reflect automation of formalization rather than genuine mathematical insight, essentially cherry-picking low-hanging fruit enabled by human preprocessing. The evidence tilts skeptical: the lack of specificity about which problems were solved and the absence of peer-reviewed results from the claimed active deployments suggest the autonomy claims are overstated.

This debate echoes the Fable 5 release tension—are we seeing genuine capability gains or sophisticated packaging of existing compute power behind curated benchmarks?