The Archive

Plants from OpenAI, Meta, xAI, and Microsoft could emit more than 129M tons annually.

Proposes GROUNDING.md field-scoped epistemic grounding document to improve agentic AI coding reliability in scientific domains.

Mobile image enhancement model with gated encoding and multi-scale refinement bridging training-deployment gap for quantized inference.

Unsubstantiated post with no context or verifiable claim.

Mixed-methods study on university students' psychological willingness to disclose AI use via Cognition-Affect-Conation framework.

Ramen framework enables test-time adaptation of vision-language models under mixed-domain distribution shifts via active sample selection.

Agent Evolving Learning framework enables LLM agents to accumulate and leverage experience across open-ended episodes via two-timescale Thompson sampling.

X-GRAM proposes frequency-aware dynamic token injection and hybrid hashing to improve embedding parameter efficiency for large lookup tables.

Code LLMs generate ML pipelines with 87.7% inclusion of sensitive attributes in feature selection, revealing underestimated bias beyond simple conditional statements.

Framework for fairness in sequential decision-making under uncertainty via counterfactual inference, addressing online ML applications with cascading decisions.

Tencent releases Hy3 preview: 295B MoE model with 21B active parameters on Hugging Face.

Cross-lingual dysarthria severity assessment across 3,374 speakers from 12 languages shows aetiology-specific phonological degradation patterns in SSL speech models.

Technical walkthrough: Qwen 3.6 27B achieves 85 TPS, 125K context on single RTX 3090 using llama.cpp.

User reports success running Qwen 3.6 35B with local coding agent using structured planning skill file on production code.

BadStyle backdoor attack framework uses natural language style triggers against LLMs with reliable payload injection in long-form generation.

Topological data analysis via persistent homology applied to eye-tracking fixation sequences for dyslexia detection using hybrid statistical-topological features.

TEmBed benchmark evaluates tabular foundation models across four representation levels for table retrieval, semantic search, and prediction tasks.

Usually the Claude session window always starts and ends on the full hour but today I just noticed it's not matching up with the full hour. Another small change is they removed the refresh button from their UI so you can't see your real-time usage without refreshing the whole page. Is this happening for everyone else? Are they changing some things about their usage limits?

Logic Gate Networks replace floating-point feature extractors with binary representations for compact, efficient video copy detection at scale.

Today on Decoder, I want to lay out an idea that’s been banging around my head for weeks now as we’ve been reporting on AI and having conversations here on this show. I’ve been calling it software brain, and it’s a particular way of seeing the world that fits everything into algorithms, databases and loops — software. Software brain is powerful stuff. It’s a way of thinking that basically created our modern world. Marc Andreessen, the literal embodiment of software brain, called it in 2011 when he wrote the piece “Why software is eating the world” as an op-ed in The Wall Street Journal. But s...

StrictlyVC San Francisco is in just a week. Now’s the time to grab yourself a ticket. Join VCs and founders at Sentro Filipino Cultural Center on April 30.

TechCrunch has confirmed that Delve was the compliance company that performed the security certifications for Context AI, the AI agent training startup that last week disclosed a security incident.

Reddit anecdote claims Chinese workers at OpenAI contractor training AI models that may displace them; unverified sourcing and secondhand reporting.

Opinion piece identifying five research strands—solvable settings, tractable limits, feature analysis, loss landscapes, implicit regularization—converging toward unified deep learning theory.

AttnLRP explanation method applied to DNABERT-2 genome language model reveals whether Transformer attention captures relevant genomic patterns versus CNNs.

A-IC3 augments IC3 hardware model checking with learning-guided inductive generalization to accelerate counterexample generalization and clause synthesis.

He said it on the [Dwarkesh Podcast ](https://mrkt30.com/anthropic-mythos-triggers-chinas-ai-arms-frenzy/)this week and I have not been able to stop thinking about it. His argument was not that China is not a threat. It was that cutting them off and treating them as an enemy is probably not the smartest long term play. His actual words were that victimising them and turning them into an enemy likely is not the best answer. The context here is Huawei targeting 750,000 AI chip shipments this year. It is nowhere near Nvidia's compute but the direction of travel is clear. And if DeepSeek ends u...

Earlier this month, millions of OpenClaw users woke up to a sweeping mandate: The viral AI agent tool, which this year took the worldwide tech industry by storm, had been severely restricted by Anthropic. Anthropic, like other leading AI labs, was under immense pressure to lessen the strain on its systems and start turning a profit. So if the users wanted its Claude AI to power their popular agents, they'd have to start paying handsomely for the privilege. "Our subscriptions weren't built for the usage patterns of these third-party tools," wrote Boris Cherny, head of Claude Code, on X. "We wa...

GEM: smooth rational activation functions matching ReLU performance with C^2N differentiability for deep networks.

Maggie Appleton on social signaling benefits of public learning via blogging and podcasting.