LLMs & Generative AI - Page 5 of 48
Latest breakthroughs in large language models and generative AI shaping the future of artificial intelligence and machine learning.
951 articles
View complete article list
Google I/O 2026 Showcases Gemini‑Powered Infinite Scaler and Code Countdown
Google I/O 2026 wasn’t just a stage for announcing new AI capabilities—it was a proving ground for those very tools.
Gemini App Targets General Users—Students, Writers, Marketers, and More
Google’s Gemini ecosystem now feels split in two. On one side sits the Gemini App – a chat‑style interface that reads like any consumer‑focused AI assistant.
Study fine-tunes honest and deceptive variants of five transformers with LoRA
Why does it matter when a model can “choose” to lie? Researchers are probing that question by training honest and deceptive versions of five popular transformers—Pythia‑1.4B, Gemma‑2‑2B, Gemma‑2‑9B, Qwen2.5‑7B and Llama‑3.1‑8B—using LoRA on the same...
3-large embedding wins 2.1 test; MiniLM wins 2.3; rerankers lag in 2.2
A team building a retrieval‑augmented generation pipeline over a few hundred contracts quickly discovers the same cracks that Article 2 warned about: embeddings stumble on negation, on exact identifiers, and on the distance between a question and...
Proxy-Pointer RAG Bakes Emerson Deltas into Index for AT&T system
Why does this matter? Enterprises are forced to feed every page of a contract—often over 100 pages and more than 500 k characters—into a large language model for named‑entity and relations extraction. The result?
Study finds base AI models predict human behavior better than fine‑tuned chatbots
Why does this matter? Researchers have found that making large language models helpful actually dulls their knack for mimicking human choices.
Chronos-2 uses known covariates such as weather for building demand forecasts
Time‑series data powers a huge swath of industrial workflows—think demand forecasting, anomaly detection, classification of sensor streams.
OpenAI upgrades GPT-5.5 readability, removes Canvas from Instant and Thinking
OpenAI is tweaking the ChatGPT experience again. While the company rolls out a readability upgrade for the newly launched GPT‑5.5 Instant, it’s also pulling the plug on a couple of legacy models.
Deep learning models auto‑detect data features, reducing need for engineer input
Artificial intelligence is reshaping how we work, but it’s also inventing a whole new lexicon. Spend five minutes on any tech site and you’ll hit acronyms like LLMs, RAG and RLHF, enough to make even seasoned engineers pause.
Google's Gemini Spark sees my whole life, then friend‑zones my boyfriend
At Google’s I/O developer conference this spring, the company rolled out Gemini Spark, an “always‑on” AI assistant that plugs directly into your Gmail, Docs and Calendar.
Researchers Find Failure Signatures in LLM Trading Agents' Planning Embeddings
Why does this matter? Because LLM‑driven trading bots are being tested in environments that mimic real‑world markets, and their internal states can betray trouble before a loss materialises.
SSD removes sync bottleneck in speculative decoding on MI300X
Why does this matter? Large language models still churn out tokens one at a time, leaving modern accelerators underused.
Claude Opus 4.8 Trained for Honesty, Flags Uncertainty, Reduces Frustrations
The AI field is no longer just about bigger numbers. A year ago, every release sounded like a brag‑fest of parameters and benchmark scores. Today, developers ask a different question: can the model be trusted in production?
Transformer Architecture Reduces Perplexity by 2.92 vs Fine‑Tuning
The Cognitive Categorical Transformer (CCT) adds a twist to a standard GPT‑2 Small backbone. It’s a 306‑million‑parameter model that stitches together category‑theoretic components and ideas drawn from cognitive science.
Step 3.7 Flash runs on NVIDIA GPUs via SGLang, TensorRT-LLM, vLLM
Step 3.7 Flash is the newest vision‑language model from StepFun, aimed at enterprise‑grade multimodal AI. It packs 198 billion parameters in a Mixture‑of‑Experts architecture, yet only about 11 billion are active during any forward pass.
LLMs Struggle with Causal Discovery While Interventional Agents Succeed
Why do large language models stumble when asked to uncover cause‑and‑effect? Researchers say the answer lies not in a particular architecture or dataset but in the learning paradigm itself.
DynaSchedBench Introduces SESC and SSI to Rank LLM Scheduling Tasks
DynaSchedBench arrives at a moment when research on the Dynamic Flexible Job Shop Scheduling Problem (DFJSP) is split between two opposing practices.
LLM-based Architecture Targets Explicit and Implicit Human Values in Text
Why does this matter? As autonomous systems take on more decisions, the gap between raw optimization and human‑centred judgment widens.
Google AI launches Daily Brief in Gemini app for U.S. users 18+
Google’s I/O 2026 turned the spotlight on a suite of new AI tools that aim to blur the line between input and output. Here’s the thing: Gemini Omni, the headline model, claims it can “create anything from any input,” starting with video.
Google Cloud unveils AI platform with Gemini, Wiz, Codemender to patch gaps fast
Google Cloud has rolled out “AI Threat Defense,” a platform that stitches together four AI‑driven components to hunt for and seal security holes faster than a human team could.
📚 Featured Resources & Reviews
🎓
No Code MBA Course Review
Learn to build AI-powered apps without coding. Our comprehensive review of No Code MBA's course.
🔧
AI Tools & Resources
Curated collection of AI tools, courses, and frameworks to accelerate your AI journey.
📬
Weekly AI Digest
Get the week's most important AI news delivered to your inbox every week.