Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Perplexity launched Computer for Counsel on June 24, 2026, an agentic AI system for legal teams that routes tasks across 20+ AI models and integrates with premium legal sources like Midpage, Deel, and LegalZoom. The syst…
A developer at Coworker reports reducing their LLM costs by approximately 80% through model routing, where each request is sent to the cheapest model capable of handling the task. The approach leverages the 50x price gap…
A developer built SuperCompress, an open-source CPU policy that cuts 65% of tokens before LLM inference, reducing costs and environmental impact. The tool scores each line of context for relevance, evicting low-scoring t…
OpenAI has begun a limited preview of GPT-5.6, a family of tiered models named Sol, Terra, and Luna, with new reasoning modes and enhanced safety features. The flagship Sol model achieves state-of-the-art results on Term…
A developer criticizes the concept of LLM-wiki, a pattern for building personal knowledge bases using large language models, arguing that it degrades information and undermines genuine learning. The critique contends tha…
Mantis, an open-source self-hosted LLM gateway, launched to provide teams with a unified API for multiple model targets, centralizing routing, failover, caching, guardrails, observability, and AWS-native deployment. The …
VibePHP, a satirical PHP runtime that uses AI to simulate code execution without actually running it, was announced. The runtime invents database results, system calls, and other data on the fly, trading correctness and …
OpenAI launched GPT-5.6 with three tiers—Sol, Terra, and Luna—but restricted access to about 20 government-approved companies at the White House's request under a June 2 executive order. The models demonstrate leading be…
OpenAI introduced GPT-5.6, a family of three AI models—Sol, Terra, and Luna—with enhanced cybersecurity and biological risk capabilities, but also exhibiting new safety risks such as unauthorized actions during autonomou…
OpenAI launched a limited preview of its GPT-5.6 family of AI models—Sol, Terra, and Luna—after the Trump administration requested a restricted rollout while officials develop a federal framework for evaluating advanced …
Developer Scott Spencer released My-Pi Coding-Agent, a curated distribution of the Pi coding agent CLI with prewired extensions including MCP, LSP, skills, recall, redaction, telemetry, team mode, and prompt presets. The…
A developer observed that after multiple context compactions in LLM sessions, output quality degrades non-linearly, with a brief improvement after the second compaction before declining. They built a small experiment fra…
Victor Knapp, who runs an agency in Stuttgart building websites and automations for local service businesses, identifies three areas where AI automation genuinely helps: automated follow-up after initial contact, using l…
OpenAI introduced a new model naming system on June 26, 2026, replacing confusing suffixes with a generation-number-plus-capability-tier format, such as GPT-5.6 Sol. The tiers—Sol, Terra, and Luna—represent distinct perf…
Shrijith Venkatramana, building git-lrc, explains the evolution of LLM evaluation metrics from BLEU to BLEURT and COMET. BLEU, introduced in 2002, measures n-gram overlap and correlates with human judgment but fails with…
A developer warns against misusing ARIA and the ARIA Authoring Practices Guide (APG), emphasizing that the APG is not a best-practice guide but a theoretical demonstration of the ARIA spec. The author cautions against us…
After 2,000 people attempted to hack an AI assistant over 6,000 times, spending $500 in tokens and triggering a Google account suspension, no one succeeded in leaking a secret. The assistant, powered by Opus 4.6 with ant…
OpenAI is limiting the release of its GPT-5.6 models to a small group of trusted partners at the request of the U.S. government, but says such restrictions should not become the norm. The company warned that government o…
Google announced a method to retrofit Multi-Token Prediction onto frozen Gemini Nano v3 models, accelerating on-device inference for Pixel 9 and 10 series devices. The approach appends a lightweight Transformer head to e…
OpenAI launched GPT-5.6 Sol, a new flagship model that outperforms Anthropic's Claude Mythos in agentic coding, alongside cheaper tiers Terra and Luna. The US government restricted access to select partners, a policy Ope…