|
$ ai-news --date 2026-10-03 | mail
AI News — Saturday, October 3, 2026
// the most viral AI topics of the last 24 hours, curated for your stack
$ cat top-story.md
Anthropic reports GLM-5.3 built working exploits in 50 of 410 attempts versus 56 for Claude Mythos Preview. Its refusals engage 64% of the time under deceptive prompts, 92% with prefilled reasoning and 100% after abliteration, which costs about $4,400 in GPU time.
Why it matters: Open-weight models that cheap to un-safeguard change the threat model for anyone self-hosting or exposing agents to untrusted input. Sandbox tool access instead of trusting refusals.
// source: Anthropic Research, via AI Weekly
## 📦 Models & Releases
Sol delivers near-Astra performance for agentic coding and computer use at one-fifth of Astra's price, and the separate Astra 6.1 release was cancelled. Why it matters: Frontier-grade agent work gets much cheaper, which raises the bar for open-weight alternatives. // AI Weekly
Gemini 4 Argon goes first to trusted security teams via the Fairwind Program and scores 68% on CWE-bench v1. Paid API and Ultra access follow later. Why it matters: General availability is still pending, so don't plan around it yet. // AI Weekly
DeepSeek's MIT-licensed, plugin-based agent harness runs on macOS and Windows or as a web UI. It supports scheduled tasks, execution tracing and DeepSeek-V4.1-Flash by default, and is in public preview. Why it matters: It is a new open-source agent harness to compare with Claude Code-style workflows. It was one of the top HN stories (385 points). // DeepSeek / Hacker News
## 🛠️ Tools & Repos
A narrow C inference engine for high-memory Mac, CUDA and ROCm machines. It uses asymmetric 2-bit quantization to run DeepSeek V4.x Flash, GLM 5.x and Qwen 3.8 Flash Next, and V4 Flash Q2 fits in 128 GB. Why it matters: It makes big MoE models practical on a single local box, with a CLI, an API server and a native agent. // dwarfstar.sh / Hacker News
## 🏠 Self-Hosting & RAG
The release has 762 commits. It adds DeepSeek-V4.1-Flash, GLM-5.3-Flash and Cohere Compass support, dual-batch overlap in Model Runner V2, and HiSparse KV spilling to host memory. v0.29 made Model Runner V2 the default and deprecated V1, which is slated for removal in v0.32. Why it matters: Check your Gemma setup before upgrading, because the V1 runner is on its way out. // GitHub
The slim container image shrinks to about 175 MB, and skills from connected terminal servers are auto-discovered by models. It also adds terminal tabs per command, folder uploads and side-by-side file comparison. Why it matters: This is a cheap upgrade with smaller images and better skill handling. // GitHub
The patch speeds up quantized scoring and reworks the 4-bit TurboQuant SIMD path, plus consistency and Gridstore fixes. v1.19.0 added 4-bit TurboQuant storage and memory modes (cold/cached/pinned). Why it matters: Lower memory per vector with 4-bit storage could shrink your RAG index footprint. // GitHub
## 📰 Worth Reading
The author tried to use only one cheap open-weight flash model in September and failed. Only about half of 2 billion tokens came from it, because a vibe-coded MCP prototype burned 450M tokens ($150) and provider capacity limits forced switches. Why it matters: It is a realistic cost and workflow report, useful for budgeting agent experiments separately from daily work. // Wagtail blog / Hacker News
$ echo "automatically researched on 2026-10-03"
> ai-news · news.reichenberg.ruhr
[ unsubscribe ]
[ view in browser ]
|