Qwen3.8-Flash-Next: 125B MoE Model with only 6B Active ParametersGiga-Embeddings: 10B MoE Encoders Boost Text Embedding ThroughputFREF: Function-Level Execution Feedback Boosts Code LLMsLLM Agents Run Controlled Experiments via Simulation ModelsAQLoRA: Zero-Search Recipe for Faster Quantized LoRA Fine-TuningAutonomous Math Discovery in Open-World Multi-Agent EnvironmentESQ-Bench: New Benchmark Tests NL2SQL on Enterprise OracleGoogle Launches Gemini 3.5 Transcribe for Smarter Speech-to-TextGitHub Copilot App Automates Dependabot PR TriageServe Markdown to AI Agents via Accept HeadersWebMCP: Teaching Websites to Talk to AI AgentsLovable CTO: Future SaaS Built for AI Agents, Not Just HumansAnima Anandkumar: We Have Foundation Models for Language, Not PhysicsPaul Dix: AI Wrote 1M LOC, Refined It into Reliable SoftwareOpenAI Report Shows How AI Makes Learning ContinuousGates: AI Risks Are Real but ManageableGates: Turbulent AI Era Has Arrived, Critical Choices AheadAI Suggests, You Finish: Hard to Complete Ideas That Aren't YoursChatGPT for Teachers Expands to 55 US Districts, Reaching 100K+ EducatorsQwen3.8-Flash-Next: 125B MoE Model with only 6B Active Parameters

arXiv released Giga-Embeddings, a text embedding family balancing strong retrieval quality with efficient serving. Its largest model is a sparse 10B-parameter MoE encoder with ~1.8B active per token, improving throughput without sacrificing accuracy. Ideal for large-scale document retrieval and semantic search.

A new paper introduces FREF (Function-Level Execution Feedback) for code preference optimization. Unlike process supervision in mathematical reasoning, code generation lacks standard step definitions. FREF provides more granular guidance via function-level execution feedback, potentially improving generated code correctness.

Recent research demonstrates LLM agents conducting controlled experiments using simulation models. The study notes many scientific and engineering tasks require understanding system responses to interventions, not just generating plausible text. It enables LLMs to autonomously design experiments in simulated environments, expanding their potential in scientific discovery.

Quantized fine-tuning (QLoRA) saves memory but is slow. New research presents AQLoRA (Adaptive-Quantization LoRA), which uses a single CPU pass to reduce training time while retaining quantized memory benefits. This zero-search recipe makes quantized fine-tuning speed closer to fp16 LoRA.

New research explores autonomous mathematical discovery in ‘Station,’ an open-world multi-agent environment. AI agents from different model families share a research goal without a coordinator, choosing their own paths. Experiments show this environment effectively promotes cross-model collaboration and autonomous exploration.

Current NL2SQL benchmarks like Spider report >89% accuracy on simplified academic schemas, but fail to reflect enterprise complexity. New benchmark ESQ-Bench simulates real enterprise Oracle databases, testing models on dialect generalization and silent semantic divergence, providing more reliable evaluation for industrial deployment.

Google DeepMind released Gemini 3.5 Transcribe with improved speech-to-text capabilities. It handles noisy environments, multi-speaker recognition, and domain-specific jargon better, generating accurate transcripts with punctuation and paragraph structure. The model is available via API for meetings, subtitles, and voice assistants.

GitHub published a tutorial on using Copilot app to automate Dependabot pull request triage. The feature identifies dependency update types, assesses impact, and auto-labels or assigns PRs to maintainers, streamlining the tedious update process. Developers can customize rules to match project standards.

A technique to serve AI agents cleaner web content: declare preference for Markdown via HTTP Accept headers, allowing servers to return structured Markdown instead of messy HTML. This reduces parsing cost and boosts extraction efficiency for AI reading tools and crawlers.

A technical blog introduces WebMCP, a concept to help websites communicate with AI agents in a standardized way. Borrowing from Model Context Protocol (MCP), it offers a universal interface for sites to interact with AI. This has practical implications for websites seeking accessibility and interconnectivity in the AI era.

Lovable is pivoting from AI web app creation to MCP-powered capabilities. CTO Fabian Hedin argues future SaaS must be usable by AI agents, requiring standardized, machine-readable interfaces. The company is exploring how agents can invoke Lovable-generated apps via MCP to perform complex tasks.

Caltech professor Anima Anandkumar notes that foundation models for language exist, but not for physics. She is applying AI to weather forecasting and fusion reactor modeling, emphasizing that physical models must obey physical laws, not just be data-driven. She sees this as a key breakthrough frontier.

Paul Dix expressed amazement that AI wrote 1M lines of code, refined it over months, and produced reliable software now running on millions of developer machines. He sees this as a testament to AI coding’s potential, while acknowledging reliance on a strong ‘oracle’.

OpenAI published a new report on how students and educators use ChatGPT to make learning continuous beyond the classroom. It highlights AI use cases in after-school tutoring, personalized practice, and lesson prep, while also discussing potential equity challenges.

Bill Gates published an article on Gates Notes addressing concerns about AI risks. He argues that while the risks of AI are real, they are manageable. Gates emphasizes that through global cooperation and responsible policymaking, humanity can harness AI’s immense benefits while effectively mitigating its potential dangers.

Bill Gates writes again, declaring we are in a turbulent AI era. He stresses AI will reshape jobs, education, and society, but guiding this transformation for the benefit of all is the critical choice ahead. The post sparked extensive discussion on Hacker News.

A blog post explores the experience of using AI for writing and thinking. The author finds completing an idea suggested by AI, rather than originating from oneself, is very difficult due to lack of intrinsic motivation and contextual understanding. The post suggests treating AI as an assistant, not a creative source.

OpenAI is expanding ChatGPT for Teachers to 55 U.S. school systems, delivering secure AI tools, training, and support to over 100,000 educators and staff. The program aims to help teachers leverage AI for lesson planning, student assessment, and other daily tasks.