Large Language Model News

OpenAI published a new pre-deployment safety method called Deployment Simulation, which replays past conversations through a candidate model to estimate undesired behavior frequency before release. The method has already…

Z.ai released GLM-5.2, an MIT-licensed open-weight frontier model with 744B parameters, achieving top scores in frontend coding benchmarks and surpassing Opus 4.8. The model features a 1M-token context window, two reason…

A developer built a command-line Topic Explainer using Oxlo.ai's API and the llama-3.3-70b model, demonstrating system prompts, message history, and streaming for beginners. The project includes a reusable function that …

Hedgeye analyst Felix Wang recommends a short position on Chinese AI stock Zhipu, calling it 'one of the biggest bubbles I've ever seen' after a 1,100% rally since its January IPO. Wang sets a fair value target of HK$407…

Polar is building a new design system called Orbit to enforce brand consistency when LLMs generate UI code, moving away from Tailwind's string-based classes to a prop-based system that makes off-brand decisions impossibl…

Trump administration officials concluded talks with Anthropic on Monday without lifting export controls on its Claude Fable 5 AI model, imposed over jailbreaking concerns. The administration believes guardrails can be by…

Cloudflare added GLM-5.2 to Workers AI on June 16, a 744-billion-parameter Mixture-of-Experts coding model with a 1-million-token context window, open weights under MIT license, and two reasoning modes. The model scores …

A developer on Qiita, nogataka, has published a guardrail architecture for Claude Code deployments that addresses context bleeding and secret exposure risks. The architecture uses repository-level configuration, policy f…

AI does not replace human thinking but amplifies it, according to a developer's experience. The quality of thought determines outcomes, with clear thinking leading to dramatically better results and poor thinking causing…

Kog AI achieved over 3,000 output tokens per second per request for an FP16 2B model on a single 8× MI300X node using a monokernel that eliminates per-token kernel launches. The technique collapses the entire LLM decode …

Anthropic's Claude Fable 5 and Mythos 5 models were suspended globally around June 13, 2026, after a U.S. export-control directive ordered the company to cut off access for foreign nationals on national-security grounds.…

A red-team study of Anthropic's Fable 5 and Opus 4.8 large language models found that both models remain reliably breakable under sustained automated pressure, with Opus 4.8 producing 1,620 and Fable 5 producing 702 pane…

OpenAI updated its Ad Tools Terms Document to allow AI-powered creative tools for generating, modifying, and optimizing advertising creatives. The company launched a self-serve Ads Manager, removed a $50,000 minimum spen…

WPP Media partnered with Givsly to test synthetic audience segments incorporating charity donation data for programmatic ad buys. Early tests across four beauty and fashion campaigns showed a 2% lift in video completion …

OpenAI expanded ChatGPT's voice input to support over 70 languages with automatic language detection on Android and iOS, and released three new developer models—GPT-Realtime-2, GPT-Realtime-Translate, and GPT-Realtime-Wh…

A 124M-parameter GPT-2 model trained from scratch on OpenWebText data using a custom deep learning library achieved a validation loss of 2.764 nats and a perplexity of 15.87 after 56,000 steps (27.5B tokens). The model, …

Oracle's managed MCP server for database queries works with desktop AI clients like Claude Desktop and VS Code Cline, but fails with server-hosted web UIs due to OAuth discovery issues. A developer successfully connected…

Researchers at Indiana University found that AI rewriting of radiology reports causes a 'slop paradox' where tasks that produce cleaner text for multimodal training actually degrade image-text alignment more than summari…

A fintech startup deployed an internal AI assistant on Amazon Bedrock without auditing it for prompt injection vulnerabilities. An employee successfully executed a prompt injection attack, leaking system instructions and…

A new open-source tool called Bean replaces fixed-step task execution with a recursive convergence loop for large tasks in Claude Code and Codex. It uses a claim ledger and compiler to investigate, record evidence, revis…