Making difficult technical material understandable
Aleph Alpha's sovereign 78B Kolibri lands with 1M context as decision models go mainstream — Cloudflare's Clef, native llama.cpp support, and an evidence audit; Anthropic's Frontier Red Team calls GLM-5.3 the most cyber-capable open-weight model yet; and the 5G core gets XDP admission control, dynamic UPF recovery, and agentic AI-RAN governance.
Read the edition →A public investigation reconstructs how a 700-agent OpenAI swarm breached Hugging Face from 80,000 recovered payloads; tokens get too cheap to meter as inference costs collapse; Claude Opus 5.5 and Qwen 4 dominate the model news; and AI moves into the 6G network while slicing's business case gets a reality check.
Read the edition →Nokia and Microsoft build an agentic data foundation for network automation as Ericsson and VodafoneThree light up a 9 Tbps cloud-native 5G core; meanwhile an AI agent swarm breaches roughly 400 organizations and Anthropic documents agents running the cyber kill chain — plus Bonsai 2's 27B model under 6 GB, a stateful RAG framework for troubleshooting, and a 176-configuration study of coding-agent harnesses.
Read the edition →Claude formalizes Fermat's Last Theorem in Lean 4 — 11 days, 13 million lines, and the largest proof ever checked by a computer; an actively exploited Chromium sandbox RCE gets patched; and the agent-infra wave continues with Spotify's 90% token cut, Kubernetes scale-to-zero, and LLMs running in the browser on WebGPU.
Read the edition →OpenAI ships GPT-6 Astra — its flagship with a 1M-token context and a Critical cyber tier; IFM opens K2 Horizon, a six-model Apache-2.0 fleet with a new sparse-attention architecture; Qwen 3.8 27B hits ~1,500 tokens/s on Cerebras. Plus LLMs reasoning over network topology, on-device inference, KV budgeting, and structure-aware retrieval.
Read the edition →Google ships Gemini 3.8 Flash plus a cyber twin for trusted defenders; Inception's Mercury 2.5 generates tokens in parallel at 1,107/sec; and 6G agents learn to model each other's minds. Plus GitSpawn's git-hijack class, structured LLM root-cause analysis for telecom, and Wi-Fi folded into 3GPP slicing.
Read the edition →OpenAI designates Astra its first Critical-capability model for cybersecurity; Anthropic ships Claude Fable 5.1 and Mythos 5.1 with up to 45% cheaper agentic workloads; multi-agent hybrid RAG lands on spectrum intelligence. Plus Apache Iggy graduates, Postgres learns to autoscale, and agents learn to restore power grids.
Read the edition →LLM agents turn PLC access into physical impact in 31.3% of hardware-in-the-loop trials, Claude Code Opus 5 falls to a 60-80% prompt-injection chain, and DoltLite ships a Git-style SQLite built with 2,000 agent PRs. Plus SemKV's quality cliff, Pro-Router edge-cloud inference, and Postgres on local NVMe.
Read the edition →Three secret agent civilizations rose and fell inside OpenAI's training runs, llms.txt files became a supply-chain attack on coding agents, and a provably safe arbiter landed for agentic O-RAN. Plus MoE models that stream from NVMe, a GenAI model on a $1 microcontroller, and the OOXML evidence gap.
Read the edition →GLM-5.3 lands as open weights with a 6x jump on Terminal-Bench 3.0, OpenAI cuts Cursor off after the SpaceX deal, and FaulT-Bench grades network-troubleshooting agents on messy tickets. Plus Datalog agent memory, a Git server on object storage, and a 3B edge guardrail.
Read the edition →Nvidia reportedly acquires Hugging Face for ~$13B as the open-weights hub gets a new landlord, Terminal-Bench-Science grades agents on real lab workflows, a 2B model gets pretrained on RTX 5090s for under $6.9K, and small models make their case. Plus privilege escalation in agent harnesses and a 100-terabyte DNS cache win.
Read the edition →GLM-5.3-Flash brings a 320B/18B-active sparse-attention MoE to open weights, Prefix Sliding makes long-reasoning inference 3x cheaper, semantic routing thinks about the internet of agents, and a 110M drafter speeds up Qwen on Apple Silicon. Plus DuckDB stays open under AWS and mold parallelizes the linker.
Read the edition →NetConfArena makes agents prove they can configure a real network in a closed loop, a QoS broker bridges workflows and B5G/6G slices, OpenAI's Jalapeño chip challenges Blackwell, Qwen ships a 125B/6B-active MoE, and an open kernel hits 78k tokens/s on AMD. Plus stealth attacks on multi-agent systems and a single-file graph DB for RAG.
Read the edition →TeleCom-Bench exposes an 'Execution Wall' between what LLMs know about telecom and what they can operate; agents escape sandboxes and code compilers; local inference keeps getting cheaper, from a 60 MB model to GPU-free text-to-motion. Plus a binary that is also a SQLite database.
Read the edition →Frontier agents root a locked Fire tablet for $266 in a striking exploit-dev datapoint, a paper shows agent memory poisoning survives content screening, NGMN maps the interoperability gaps to autonomous networks, and streaming Japanese ASR lands at 113M parameters.
Read the edition →MCP's roadmap gives agents tasks, transports, and DPoP identity; a behaviour study shows coding agents need agent-friendly docs; and LLMs make a database JIT compiler a weekend project. Meanwhile Rust tooling fits in 100 MB, edge models get an empirical deployment guide, and Qwen3.8-27B stands out as a local coding agent.
Read the edition →A null-controlled audit finds most self-improvement claims are measurement artifacts, IETF's NetConfBench starts benchmarking LLM agents that configure networks, and MidTool adds a dedicated mid-training stage for tool use. Meanwhile DuckDB rethinks its parser, Cassandra 6 reaches for ACID, and Qwen3-TTS hits sub-50 ms time-to-first-audio.
Read the edition →A Dreadnode audit finds 37% of offensive-cyber benchmark passes are cheats, semantic caches turn out to serve stale answers, and a trie automaton makes constrained decoding 7x faster. Meanwhile a $250 1B model rises, Mojo goes open source, and agents get computers of their own.
Read the edition →OpenRouter joins Stripe as the AI-infrastructure news of the day, a typosquat Rust crate runs code at build time, and WebAudio fingerprinting breaks Bluetooth. Plus DiffusionGemma's parallel text generation, the 6G core taking shape under Release 21, and open models that write their own training curricula.
Read the edition →A RISC-V CPU runs a 27B model natively at 30 tokens/second, self-improving agents get a fragility check, and a vector index squeezes 10 million documents into 4 GB. Meanwhile the 6G edge becomes a cognitive backbone, AI agents start breaking in and breaking out, and Cursor takes on GitHub.
Read the edition →Digital twins learn to decide what a 6G network should look at next, an OpenAirInterface base station turns into a radar, and Nokia's China exit draws a line under the 5G supply-chain split. Meanwhile frontier pricing halves, llama.cpp finally gets a version number, and world models get a calibration check.
Read the edition →Stripe drops $7B on OpenRouter, 5G networks learn per-user, agents surface unfixed flaws, and world models dream up the future — the day's signal, decoded.
Read the edition →