The Apache 2.0 code runs V4.1 Flash on Nvidia SM100 GPUs and Huawei Ascend 950 chips, with about 8,400 GitHub stars. It covers offline inference only, so you still bring your own cache and load balancing.
Its new Auto router matched Opus 5.5 and Astra on 20+ common tasks at about 90% lower average cost, Notion says. If you pay for Notion AI, test Auto before your next credit bill.
In Cloudflare's test a pipeline using it took 19.3 seconds against 61.7 for an all-LLM version, at about half the cost. It is in preview on Workers AI, with open weights promised.
The text-to-speech model claims 70 ms latency, 25+ languages and a $1 self-hosted rate before hardware. If you run voice agents, price it against your current voice bill.
Backers IBM, Red Hat and Google report 3,000 coding agents running at once on self-hosted gear, 5 to 10 times cheaper than a hosted service. Worth a look if your agent API bill keeps climbing.
They test decision models as a cheap stand-in for LLM judges, with OpenAI Decisions at F1 0.87 for 3.72 cents per 100 runs. Tune the threshold on labeled traces instead of keeping 0.50.
Import AI 477 carries his IMF lecture, citing FrontierMath scores rising from 2% to 94%, and asks governments to build power, compute and oversight now. A big-picture read from Anthropic's co-founder. His GPT-4.1 site drew about 72,000 ChatGPT fetches in 30 days, and its fake numbers turned up in an academic paper. Check the source behind any statistic an AI hands you.
He codes for fun on weekends, from a buggy Breakout clone to type tricks in Java, and argues joy needs no business case. A calm read if agents have you doubting the craft.
Beam is a new Mac app that takes a window running on one Mac and puts it on another, where The post Beam pulls single app windows off the Mac Mini… · The New Stack
How a live GreyNoise product demo turned into an incident response. Inside a proxyjacking campaign that hit 143 devices across 34 countries. · GreyNoise
Claroc found shared chats holding loan applications and revenue reviews, because the chat title sits in the web address. If your staff use Muse, tell them share links are public and delete old ones.
Redact found 102 hidden AI-only notes on 80 help pages, and WhatsApp told WIRED they were for its support bot and showed up by a bug, now fixed. Hidden page text can steer the answers an AI gives.
Terra Security says 2.3% of GitHub security fixes this year needed a follow-up advisory within 90 days, up from about 0.2%. Treat a patch for an exploited bug as a map attackers will read.
Documents obtained by The Intercept describe ECHOTRIBBLE, agents that copy themselves across devices, even air-gapped networks, then self-destruct. The program is early, and experts warn of collateral damage.
An Aachen court barred three false AI answers about a doctor, with fines up to 250,000 euros per violation, after a Munich ruling. There, AI search answers now count as Google's own words.
About 200 agents now pay through Stripe's Link, most buys run about $100, and the growth starts from a small base. Each agent gets its own card number and spending limit, a pattern worth copying.
Its agents, called townies, give each employee an assistant that can share approved info with other agents, and the first ones shipped in June. Forerunner and Menlo led the round.
The money comes from General Catalyst's Customer Value Fund, and Otter passed $100 million in yearly recurring revenue last year. It says the cash goes to enterprise growth and going global.
The BriefThe US Department of Homeland Security published an order on October 20 telling federal civilian agencies to stop buying DeepSeek products and to remove them from their systems, including third-party products and services built using DeepSeek. Removal is immediate unless an agency asks for an exception in writing within 60 days, and the order says data sent to DeepSeek could reach Chinese authorities under Chinese law. If you sell software or IT services to US agencies, check today whether any tool, AI gateway or vendor feature you supply calls a DeepSeek model.
Level UpFind every AI model your tools call. Search your code, configs and gateway logs for model IDs that start with deepseek. If you route through OpenRouter, its account-wide ignore list keeps requests off DeepSeek's own servers, but the same models hosted by other providers still need the model ID check. →OpenRouter provider routing docs (ignore and only lists)