The number that dominates today: Anthropic closes a Series H at $965 billion — an order of magnitude above anything previously seen in the sector — and simultaneously ships Opus 4.8 with Dynamic Workflows and ultracode. The timing is deliberate: raising at that valuation requires a credible product roadmap on agents and code generation, two markets where Anthropic competes directly with OpenAI o3 and Gemini 2.5 Pro. Dynamic Workflows suggests a native orchestration architecture rather than a simple API wrapper, positioning Anthropic on the agent infra layer, not just the model layer.
Two papers published today converge on the same underlying problem: LLMs know more than they say. The first (Reddit r/ML, code at github.com/synthiumjp/metacog-engineering) shows via LoRA + causal activation patching (ρ=0.976) that 7B–70B models correctly detect their own errors internally (AUROC 0.76–0.88) but consistently output 99% verbal confidence. Probe-targeted fine-tuning closes that gap. The second, MechELK (arXiv:2605.28825v1), attacks the same problem through mechanistic interpretability: SAE localization + causal probing + representation engineering → 84.7% on TruthfulQA, +6.2% over Contrastive Consistency Search, and 78.3% recovery of hidden knowledge when model output is wrong. The two approaches are complementary: one fixes the behavior, the other explains it.
On operational safety, the benchmark of 14 open-source guard models (79,331 samples, 8 NIST categories) produces a result worth keeping in mind for any architecture decision: Qwen Guard 4B hits 83.97% recall, ahead of Llama Guard 12B and GPT-OSS Safeguard 20B. Model size does not correlate with detection performance. For teams sizing their moderation stack, the signal is direct: optimize against targeted benchmarks (HarmBench, StrongREJECT, BeaverTails, RealToxicityPrompts) rather than parameter count.
Research on probe-targeted fine-tuning (LoRA) for verbal confidence calibration in LLMs. Models internally detect correct answers (0.76–0.88 AUROC) but output 99% confidence uniformly. Fine-tuning across 8 models (7B–70B) with causal activation patching (ρ=0.976). Code and pre-registration available.
Comprehensive evaluation of 14 open-source safety guard models on 79,331 samples across 8 NIST AI Risk Framework categories. Qwen Guard (4B) achieves highest recall (83.97%), outperforming Llama Guard (12B) and GPT-OSS Safeguard (20B). Model size does not correlate with safety detection performance.
MechELK is a mechanistic interpretability framework for extracting latent knowledge from LLMs. Through three stages (localization via SAE, verification by causal probing, elicitation via representation engineering), it achieves 84.7% accuracy on TruthfulQA, outperforming CCS by 6.2% and identifies 78.3% of hidden knowledge when model output is incorrect.
Mathematical framework for surrogate modeling of oil reservoirs (Norne, 46×112×22 grid) using Fourier Neural Operators (FNO) and physics-informed variant (PINO). Empirical validation: R²>0.99 (oil), R²>0.90 (gas), R²≈0.80 (pressure) over 3298 days. 10⁴× speedup vs OPM simulator, 1000-member ensemble in <1 min on B200 GPU.