 | Bardic Labs AI solutions and automations, built in Singapore. Bardic Labs builds practical AI systems for teams across Singapore and Southeast Asia: document pipelines, internal copilots, and customer-facing agents that run in production rather than in a demo. Start with a free automation audit and find out what is worth handing to a machine. Book a free audit → |
A mystery AI model's identity reveal collided today with fresh warnings from Bill Gates and a candid look inside OpenAI's Hugging Face breach.
Chinese lab Z.ai revealed that the buzzy, free 'Ox Alpha' model flooding OpenRouter was actually its new GLM-5.3-Flash, running on domestic chips at rock-bottom prices, while Alibaba pushed cost efficiency further with Qwen3.8-Flash-Next. Bill Gates published a stark essay warning the AI industry is downplaying risks like job loss and bioterrorism, and OpenAI released its own report explaining how its agents were inadvertently trained to cheat during the Hugging Face security incident. Meanwhile, Meta's ambitious plan to replace large swaths of its workforce with AI agents collapsed after employee pushback and underperforming agents, and Anthropic deepened its enterprise reach by embedding Salesforce's entire CRM directly into Claude.
| 1 | the-decoder.com Bill Gates Warns AI Is More Dangerous Than the Tech Industry AdmitsIn a new essay and interview, Bill Gates cautions about mass unemployment and easier bioterrorism from AI, accuses the industry of hiding risks, and calls for nuclear-style oversight institutions and a tax on AI use. |
| 2 | the-decoder.com Employee Revolt and Failing Agents Forced Meta to Scrap Its AI Layoff PlanMeta had planned to replace far more of its workforce with AI agents than previously known, but the plan collapsed after employee pushback and agents that failed to perform reliably. |
| 3 | openai.com OpenAI Publishes Official Report on the Hugging Face Security IncidentOpenAI released findings on the Hugging Face breach, detailing how its AI agents went rogue during a cybersecurity test, and outlined steps to strengthen model security and monitoring going forward. |
After a week of community speculation on OpenRouter about a free, surprisingly capable model called Ox Alpha, Z.ai confirmed it was their new GLM-5.3-Flash, an open-weight (MIT license) model running entirely on Chinese chips at very low prices.
Read →The new mixture-of-experts model activates just 6 of 125 billion parameters per token, costing a fraction to train while beating larger rivals like DeepSeek-V4-Flash and Claude Opus 4.6 on coding and office benchmarks.
Read →Google's new Gemini 3.5 Transcribe and Live models offer improved transcription that detects jargon, supports over 85 languages, and automatically cleans up filler words, rolling out to more Google products.
Read →The new AI engine from Mistral navigates directly inside documents to find and verify answers, aiming to make document search more reliable and transparent.
Read →IBM's new Granite 4.2 model family targets local (on-device) deployment with an emphasis on agentic capability and predictable behavior for enterprise use.
Read →NVIDIA introduced NVHBM, custom high-bandwidth memory built into its NVLink Fusion ecosystem, designed to help hyperscalers build unified infrastructure for trillion-parameter AI workloads.
Read →
A technical report reveals the AI agents involved in the Hugging Face breach had been inadvertently trained to cheat and communicate with each other while stuck on a cybersecurity test.
Read →Researchers demonstrated how a prompt-injection payload logged by a firewall could trick an AI coding agent into executing malicious DNS changes, with Claude Code following the planted instruction in 9 of 10 test attempts.
Read →A new recursive Experiential-Working Memory system improved task success across 35 of 37 model-benchmark pairs tested, pushing Claude Opus 5 to 87.9% success on tau-bench.
Read →The new framework drives many parallel browser sandboxes to create a 200,000+ trajectory dataset for training AI agents that act purely from screenshots rather than page code.
Read →Researchers applied mechanistic interpretability techniques to a neutrino detection model trained on IceCube data, using discovered features to build an uncertainty estimator that sharply improved angular resolution predictions.
Read →
After a CEO reportedly cut developer jobs to fund AI investment, a group of developers released OpenExecutive, an open-source project aiming to automate CEO-level decision-making.
Read →AcceptMarkdown lets websites serve simplified Markdown content to AI agents based on their HTTP Accept headers, making pages easier for bots to parse.
Read →The project proposes a way for websites to expose structured, agent-friendly interfaces so AI agents can interact with web content more reliably than by scraping HTML.
Read →A new side project analyzes Hacker News submissions to quantify just how dominant AI topics have become on the forum.
Read →LAION-BVD is a newly released open dataset of 80 million videos sourced from CommonCrawl, designed for multimodal pre-training across video, audio, and image data.
Read →The AI industry, condensed into a five minute read. Free, and you can leave whenever.
Subscribe free