>·4 min
Too Much Faith in AI's Ability to Refuse
Refusal is the load-bearing wall of AI safety, but it's probabilistic, can fail catastrophically, and is ripe for censorship.
ai-safetyaillm
>read more_
> booting the archive…
Loading
[ // tag: #llm ] (14)
>·9 min
Claude Mythos: 26K Findings and the Human Triage Bottleneck
Anthropic's Mythos generated 26,153 candidate findings via Project Glasswing, but few reached disclosure. The bottleneck is human triage, not discovery.
aianthropiccybersecurity
>read more_
>decode·11 min
Apriel-H1: Distill Reasoning, Not Pretraining Data
ServiceNow turns a 15B reasoning model into a Mamba hybrid at 2.1x throughput. The interesting part is how they chose the distillation data.
llmresearchopen-source
>read more_
>·10 min
TSX Costs 94% More LLM Tokens Than Pug
Four markups, one tokenizer: TSX uses 94% more tokens than Pug. But the interesting cost is not where most developers expect it to be.
llmwebdeveloper-tools
>read more_
>decode·14 min
Vi: An AI Organism Built Without a Rented LLM
Inside Vi, a sovereign AI built without a rented LLM: own weights, persistent memory, a C++ speech stack, and hippocampal replay during sleep.
aiai-sovereigntyllm
>read more_
>·13 min
GPT-6 Astra vs Agent Harness: The Real Difference
GPT-6 Astra is a model, not an entire agent. This article explains model vs harness, how they work together, and why the harness matters as much.
aiai-agentsopenai
>read more_
>decode·5 min
Claude Fable 5.1: Smarter Coding, 75% Cheaper Cache Reads
Claude Fable 5.1 doubles down on long agentic work — 52.6% on Terminal-Bench-Science — and cache reads drop 75%. Mythos 5.1: same model, looser guardrails.
aianthropicclaude
>read more_
>decode·5 min
llama.cpp Joins Hugging Face: Local AI Gets a Home
The ggml.ai team behind llama.cpp joins Hugging Face. The runtime stays 100% open source, and the transformers-to-GGUF bridge is about to get much shorter.
aiopen-sourcellm
>read more_
>·7 min
GLM-5.3-Flash vs Qwen3.8-Flash-Next: Two Labs, One Recipe
Z.ai and Qwen shipped near-identical hybrid LLM architectures within a day: 3:1 linear attention, a 2048-token sparse budget, four gated residual streams each.
aillmopen-source
>read more_
>decode·5 min
We Still Don't Know How People Use AI
A new independent study of real AI conversations shows that work-focused company reports filter out nearly half of what people actually do with AI.
airesearchllm
>read more_
>·7 min
NVIDIA Nemotron 3 Ultra: 550B Open Model for Agentic AI
NVIDIA releases Nemotron 3 Ultra, a 550B parameter MoE model with 1M context, free on OpenRouter, built for agentic AI systems.
ainvidianemotron
>read more_
>·9 min
GPT-5.6 Sol Ultrafast: 14x Faster at 750 Tokens Per Second
OpenAI's Ultrafast mode runs GPT-5.6 Sol at 750 tokens per second on Cerebras wafer-scale chips, 14x faster with no quality loss.
openaicerebrasinference
>read more_