The Archive

Google launches The Small Brief, pairing ad industry figures with local businesses to create AI-assisted marketing campaigns using Gemma.

Three months ago, Elon Musk wrote on X that Anthropic was “evil,” “misanthropic,” and that the AI lab hated Western civilization. On Wednesday, he leased Anthropic one of his most valuable assets: the world’s biggest supercomputer. But Anthropic-lovers shouldn’t bask too long in Musk’s newfound praise (even if he did decide that “nobody set off my evil detector” ). The deal has little to do with them as a company, analysts told Fortune, and everything to do with an upcoming prospectus. SpaceX is expected to begin its public roadshow next month, with a confidential S-1 filed April 1 targeti...

Community critique: LLM benchmarks should include realistic context sizes, multimodal feature usage, and agentic/RAG workloads rather than speed-only metrics.

Reddit user reports file upload failures on Claude across Windows and Android platforms.

Reddit user shares subjective experience of using Claude in Microsoft Office suite; anecdotal product feedback without technical depth or novel findings.

Z-lab releases Gemma-4-26B with DFlash inference optimization, claiming improved performance over MTP via stateful parallel block diffusion.

Gemma 4 26B achieves 578 tok/s on RTX 5090 using DFlash speculative decoding in vLLM, 2.5× faster than baseline.

This is an automatic post triggered within 2 minutes of an official Claude system status update. Incident: Elevated Errors on File Operations Check on progress and whether or not the incident has been resolved yet here : https://status.claude.com/incidents/vtt35dc73941 Also check the Performance Megathread to see what others are reporting : https://www.reddit.com/r/ClaudeAI/comments/1s7f72l/claude_performance_and_bugs_megathread_ongoing/

Reddit discussion of image upload failures and infinite loading loops on Claude web and mobile app.

Astronomers used AI to identify 10,000+ exoplanet candidates from telescope data; application domain outside core AI research.

Sometimes, companies pick CEOs based on carefully laid succession plans designed to maximize investor confidence and future performance. Other times, apparently, companies pick CEOs based on a bunch of video calls while the current CEO is texting the former CEO about who the new CEO even is. Such was the story of The Blip, the days in 2024 when Sam Altman was ousted from OpenAI. We knew that situation was chaotic; the ongoing Musk v. Altman trial is showing just how chaotic it really was. Verge subscribers, don't forget you get exclusive access to ad-free Vergecast wherever you get your podca...

Reddit user reports Claude Code agent limiting itself with overly cautious time estimates, underutilizing its own capabilities.

Community analysis of etymology and symbolism behind Claude model names (Haiku, Sonnet, Opus, Mythos) and their alignment with model capabilities.

SimCT method recovers lost supervision signals in cross-tokenizer on-policy distillation by matching representations across heterogeneous vocabularies.

HTN planning uses LLM-generated heuristics to improve hierarchical task decomposition search efficiency beyond classical planning baselines.

Bayesian LoRA in projected subspaces enables uncertainty quantification in parameter-efficient fine-tuning without inflating trainable parameters.

Novel logical characterization of encoder-decoder transformers via temporal logic and distributed automata provides theoretical foundation for cross-attention architectures.

Finite-time convergence analysis of MCTS in continuous POMDPs with probabilistic bounds addresses theoretical gaps in existing solvers like POMCP.

Dynamic guidance learning via RL replaces fixed classifier-free guidance scale in diffusion language models, optimizing control across tasks and generation stages.

DRIP-R benchmark evaluates LLM agents on real-world retail policy ambiguities with multiple valid interpretations, addressing evaluation gaps in agent robustness.

Reddit post title with no substantive content; lacks model name, capabilities, or release details.

Proves grammar-constrained speculative decoding cannot recover grammar-conditional distribution under local masking; proposes Φ-estimation to quantify distribution gap.

Decomposes room impulse response to isolate early vs. late reverberation effects on speaker distance estimation; off-topic to frontier AI systems.

Improves geometric representations for molecule generative models by decoupling representation learning from 3D structure generation.

GASim uses graph-optimized memory to accelerate LLM-agent social simulations by replacing expensive retrieval with efficient graph indexing.

DTW-certified robust anomaly detection for time-series using certified defenses against adversarial perturbations on temporal structure.

Unitree launches SDK store for G1 robot, enabling downloads of motor skills, dances, and martial arts tasks.

GRPO fails under binary rewards due to gradient starvation when all group responses are correct/wrong; group-mean centering fix demonstrated on Qwen3.5.

Chain-of-thought reasoning exhibits coupling tax: long reasoning traces compete with answers in fixed token budgets, reducing accuracy on GSM8K/MATH.

TRACE benchmark evaluates tourism recommender systems on multi-turn recommendations with verifiable review-span evidence and rejection recovery.