The Archive

NYT shifts OpenAI/Microsoft copyright claims after SCOTUS ruling against Sony.

AI agents have changed a lot in the last two years. The first could only answer one question at a time. Then came multi-turn chat, where the model could keep... AI agents have changed a lot in the last two years. The first could only answer one question at a time. Then came multi-turn chat, where the model could keep some context across a session. Today, we have long-horizon agents. Systems that plan many steps, split work between sub-agents, keep context across a long task, and run tools in a safe sandbox. The NVIDIA AI-Q Blueprint is an open source… Source

Fernando Irarrázaval's hackmyclaw challenge: 2,000 participants attempted prompt injection attacks on Claude Opus 4.6 instance; zero successful secret leaks across 6,000 attempts.

“We don’t believe this kind of government access process should become the long-term default,” says OpenAI. “It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.”

The hire marks OpenAI's latest push into India, expanding offices, partnerships and hiring.

DexCompose framework reuses pretrained dexterous manipulation policies for multi-task control via finger-level action ownership to resolve conflicting demands.

Fictional incident report imagining competing AI code review agents entering costly disagreement loops, illustrating multi-agent coordination and cost/security risks.

Theoretical characterization of proper learning from positive-only samples settles open question in PAC learning with implications for classification under distribution shift.

Study shows zero-sum game solvers systematically select different Nash equilibria deterministically based on algorithm choice, not randomness, using tabular games.

Theoretical analysis proves Bregman ADMM converges to second-order KKT points under relative smoothness for nonconvex polynomial optimization without Lipschitz gradients.

MDM-VGB applies reward-guided remasking to masked diffusion models for test-time scaling, enabling constraint satisfaction and sample editing via backtracking Markov chains.

Democratic ICAI improves preference-based alignment by eliciting multiple competing rationales through structured debate to extract nuanced steering principles from human preferences.

Nuclear physics paper using interpretable neural networks to model binding energies via SU(3)/SU(4) symmetries; domain-specific, not AI-frontier work.

Lighthearted suggestion that Anthropic should use friendlier model naming conventions.

PAC-Bayesian bounds for quadratic control with finite-sample guarantees; theoretical controls work, limited immediate AI systems relevance.

HORIZON: agent framework for hardware design via repository-level code evolution, evaluated on ChipBench and RTLLM.

Google proposes taxonomy for AI-assisted scientific peer review at four collaboration levels to address peer-review bottleneck.

PEHT transformer for cellular network traffic prediction via urban congestion features; domain-specific infrastructure application.

Mechanistic study of vision-language model conflict resolution: visual grounding dominates by default; knowledge priors require specific attention heads.

Vague Reddit post with no substantive content.

OpenAI launches GPT-5.6 series (Sol, Terra, Luna) with tiered performance/cost; limited preview underway, general availability in weeks.

Agent-Native Immune System (ANIS): runtime defense framework for autonomous agents against memory poisoning, tool-chain manipulation, multi-agent attacks.

TopoTTA: test-time adaptation for anomaly segmentation using topological structure; domain-specific computer vision application.

Stratechery weekly digest covering Apple regulatory challenges in Europe and miscellaneous tech commentary.

Less than 24 hours after news broke that OpenAI would stagger its next model release at the request of the Trump administration, that model, GPT-5.6, is here. On Friday, the company unveiled the limited preview of its new GPT 5.6 model suite: Sol, the flagship; Terra, a medium-tier model for "high-volume work"; and Luna, a "fast and affordable" everyday model. OpenAI says it's especially skilled at coding, cybersecurity, and biology, as well as staying focused during long-horizon agentic AI tasks. Per million tokens, GPT-5.6 Sol is priced at $5 input / $30 output (nearly half the cost of Anth...

Hybrid classical-quantum neural networks for oral cancer detection on edge devices; quantum-computing application, niche deployment focus.

HPRO: preference-based reward optimization for emotional TTS via hierarchical separation of content and emotion features.

Theoretical analysis of generalization scaling laws in quadratic neural networks as functions of model width and sample count.

AI models have progressed to the point where their capabilities have real political consequences. Dealing with those consequences will require collective action.

Study shows autonomous AI coding agents pass individual benchmarks but accumulate integration friction in shared codebases, revealing ecosystem-level risks.