Frontier models can recover up to 65% of facts they can't directly recall — just by thinking longer
THE SO WHAT
If frontier models can recover ~65% of “forgotten” facts by simply extending reasoning time, your first lever is inference-time strategy, not another retrain or heavier RAG stack. Teams should be benchmarking depth-of-thought settings and cost/latency tradeoffs before committing to more complex retrieval architectures.
READ THE SOURCE
MORE FROM THE WIRE
Applied AIAnthropic launches Fable 5.1 as AI security worries mount
Model vendors are starting to sell safety deltas — Anthropic is explicitly tying Fable 5.1 and Mythos 5.1 to agent security incidents and scrutiny. If you're deploying agents, assume buyers will soon ask for concrete evidence of security posture, not just capability benchmarks.
Applied AIDell’s AI servers drive a stellar earnings performance, and a raised outlook
A $95 billion backlog tied to AI servers means infrastructure constraints — power, space, delivery times — will shape what AI projects are actually feasible in 2026–2027. CIOs should align AI roadmaps with realistic server delivery and colocation timelines, not just model availability.
Applied AIOpenAI says Astra is its first model to reach its "Critical" cyber threshold and warns safeguards may mistakenly flag legitimate activity as cyber misuse
Once a model crosses a “Critical” cyber threshold, false positives become a governance problem, not just a UX annoyance — Astra may block legitimate admin and security work. Enterprises integrating these models need explicit escalation paths and override policies for cyber-related queries.
Applied AI'The disguise becomes part of the answer key': Researchers find that AI is redefining what human writing means by self-correcting itself
If 18,989 abstracts show AI-style vocabulary rising and self-correcting, then “AI detection” is on a path to futility in text-heavy domains. Organizations that still rely on stylistic signals to police AI use in writing need to shift toward process controls and provenance, not content forensics.