Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
A developer compiled a production-readiness checklist for AI agents based on an 'Ask HN' thread, focusing on specific behaviors rather than infrastructure. The checklist includes observable logging, graceful fallbacks ac…
Google launched information agents in AI Mode for AI Ultra subscribers, with expansion to AI Pro planned for summer 2026. The agents monitor blogs, news, social media, shopping, finance, and sports feeds, delivering synt…
The Atlantic published an essay on June 15, 2026 arguing that surface 'tells' used to identify machine writing are unreliable, drawing an analogy to 19th-century criminologist Cesare Lombroso. The essay recommends lookin…
WALT Labs joined Anthropic's Claude Partner Network as a Registered Partner after completing the learning path and earning the Claude Certified Associate certification, enabling the company to help organizations deploy C…
Redgate's tutorial outlines five PII anonymization techniques for LLM pipelines, including format-preserving masking, pseudonymization, and synthetic-data substitution, to reduce exposure at entry points such as source d…
Apple has struck a $1 billion-a-year deal to run a 1.2-trillion-parameter Gemini model in Google's Cloud for Siri, highlighting Apple's failure to capitalize on its 2010 AI headstart and its reliance on external AI infra…
A developer ran GLM-4 Plus and DeepSeek V4 for 30 days and reported findings on their performance. The comparison highlights differences in capabilities and practical use cases.
A developer traced a hybrid Mamba-Transformer MoE inference run and found that MoE all-to-all collective stalls dominate the tail latency, with a 69x tail ratio, despite dashboards showing 96% GPU utilization. The per-la…
The KV cache, a memory store for attention keys and values, grows linearly with context length and can exceed model weights in VRAM usage, causing out-of-memory errors for local LLM users. At 32k context, a 7B model's ca…
A coalition of 42 state attorneys general, led by New York's Letitia James, subpoenaed OpenAI for records on advertising, user retention, personal and health data, and how ChatGPT treats minors and seniors, signaling tig…
HBM costs have grown from 52% of AI chip component costs in Q1 2024 to 63% in Q4 2025, with total AI chip component spending rising from $22B to $52B. Reasonix is an open-source terminal coding agent optimized for DeepSe…
Anthropic researchers introduced Natural Language Autoencoders (NLAs), an unsupervised method that generates natural language explanations of LLM activations by training two modules to reconstruct activations through a t…
Salesforce acquires Fin, the customer-service AI company formerly known as Intercom, for $3.6bn to integrate its autonomous support agent technology into Salesforce's Agentforce platform. The deal, expected to close in Q…
A developer compared three AI certification programs—Simplilearn, Great Learning, and IIT Kharagpur's Applied AI & ML EPGC—and chose IIT Kharagpur's 8-month live program for its comprehensive curriculum covering ML, Deep…
OpenAI's Operator moves beyond standard LLM workflows by enabling multi-step, iterative browser interactions. The model holds state across events, parsing DOM elements and handling dynamic content to perform complex task…
A developer built SetTern, an AI-powered dependency resolver for moving abroad, after realizing that international relocation is a graph traversal problem with 27 tasks, unknown dependencies, and deadline constraints. Th…
Anthropic pulled its Fable 5 and Mythos 5 AI models worldwide after the US government issued an export directive citing national security concerns over a potential jailbreak method. The company disputed the claim, statin…
Agentic AI workloads are shifting the compute ratio, making CPUs the critical coordination substrate rather than a support component for GPUs. This inversion is exposed by current Xeon supply tightness, which reveals tha…
A developer outlines a practical approach to building a multi-model LLM fallback layer that allows applications to use multiple providers without spreading provider-specific logic throughout the codebase. The system defi…
Z.ai shipped GLM-5.2 on June 13, 2026, with a 1-million-token context window and zero published benchmark numbers. The model, based on a 744-billion-parameter MoE architecture, introduces a dual thinking-effort system (H…