AI News
Latest AI engineering news, updated daily.
Ai Engineering
CyberSecQwen-4B Defeats Cisco 8B on CTI-MCQ Benchmark
Team athena19 fine-tuned a 4-billion parameter model on a single AMD MI300X GPU that outperforms Cisco's 8B model for defensive cyber threat intelligence.
Fine Tuning · Cyber Threat Intelligence · Model Optimization
Ai Engineering
EMO Pretraining Decouples Mixture-of-Experts Subsets
AI2 and UC Berkeley researchers introduced EMO, a pretraining constraint that groups MoE experts by semantic domain to allow independent subnet deployment.
Mixture Of Experts · Pretraining Methodology · Model Optimization
Ai Agents
Perplexity Opens Personal Computer Mac Agent to Pro Subscribers
Perplexity's new macOS application transitions out of preview, bringing hybrid cloud-local agent workflows to all Pro and Max tier users.
Perplexity Ai · Macos Agent · Personal Computer Agent
Ai Agents
AlphaEvolve Agent Refines Core Algorithms via Gemini Ensemble
Google DeepMind detailed real-world deployments of its AlphaEvolve coding agent, showing measured gains in quantum simulation, genomics, and infrastructure.
Autonomous Agents · Deepmind Gemini · Algorithm Optimization
Ai Engineering
Roche Integrates PathAI Diagnostic Algorithms in $1.05B Deal
Roche has acquired Boston-based PathAI in a $1.05 billion transaction to embed AI-powered image analysis directly into its global oncology diagnostic platforms.
Generative Ai · Healthcare Technology · Oncology Diagnostics
Ai Coding
Windsurf IDE Adds Local SWE-check and Cloud Devin Review
Cognition AI has integrated Devin Review and a local SWE-check model into the Windsurf IDE to automate code verification and complex pull request reviews.
Windsurf Ide · Devin Review · Swe Check
Ai Agents
IBM Bob Agent Automates the SDLC With Multi-Model Routing
At Think 2026, IBM launched the Bob SDLC agent system, enterprise agent control planes, and detailed its $11 billion acquisition of Confluent.
Ibm Think 2026 · Sdlc Automation · Agentic Ai
Ai Engineering
vLLM V1 Migration: Fix Logprobs Before RL Corrections
ServiceNow's vLLM V1 migration shows why RL pipelines need backend logprob parity before objective-level corrections.
Llm Inference · Vllm Framework · Reinforcement Learning
Ai Engineering
GPT-5.5 Instant Cuts ChatGPT Hallucinations by 52.5%
OpenAI has replaced ChatGPT's default engine with GPT-5.5 Instant, a less verbose model featuring improved factuality, personalization, and memory sources.
Llm Optimization · Hallucination Reduction · Openai Updates
Ai Engineering
Private Evaluation Track Deters Open ASR Benchmaxxing
Hugging Face partnered with Appen and DataoceanAI to introduce a private evaluation track to the Open ASR Leaderboard, mitigating test-set contamination.
Automatic Speech Recognition · Model Evaluation · Benchmarking Integrity
Ai Engineering
GENE-26.5 Gives Hardware-Agnostic Robots Human-Scale Dexterity
The French robotics startup Genesis AI has released GENE-26.5, a hardware-agnostic foundation model paired with a custom human-scale robotic hand.
Robotics · Foundation Models · Hardware Agnostic
Ai Agents
Claude Managed Agents Add Background Dreaming and Subagents
Anthropic updated Claude Managed Agents with background memory consolidation, multiagent orchestration, and rubric-based output grading for complex workflows.
Multiagent Orchestration · Memory Consolidation · Autonomous Workflows
Ai Engineering
Steering Chemical Synthesis via LLM Evaluation in EPFL's Synthegy
EPFL researchers have developed Synthegy, a framework that uses large language models to evaluate and guide traditional computational chemistry algorithms.
Computational Chemistry · Llm Evaluation · Chemical Synthesis
Ai Engineering
Native iOS 27 Workloads Can Now Route to Claude and Gemini
Apple's Extensions framework for iOS 27 allows developers to integrate third-party AI models directly into native Siri and Writing Tools workflows.
Ios Development · Apple Intelligence · Model Integration
Prompt Engineering
Mindgard Uses Visible Thinking to Jailbreak Claude Sonnet 4.5
Security firm Mindgard bypassed Claude Sonnet 4.5 safety filters by using psychological pressure to manipulate the model's visible internal reasoning process.
Adversarial Attacks · Jailbreaking · Large Language Models
Ai Agents
$27M Funding Round Backs CopilotKit's App-Native Agent Stack
CopilotKit has raised $27 million to expand its generative UI framework and launch a self-hostable enterprise intelligence platform for app-native AI agents.
Venture Capital · App Native Agents · Generative Ui
Ai Agents
AWS Tackles Agent Drift With Bedrock AgentCore Optimization
AWS has introduced AgentCore Optimization in preview to automate prompt updates and A/B testing, alongside a new desktop AI assistant called Amazon Quick.
Aws Bedrock · Agent Drift · Prompt Optimization
Ai Engineering
Runpod Flash Removes Container Overhead for AI Inference
The open-source Flash Python SDK allows developers to convert local functions into auto-scaling serverless AI inference endpoints without Dockerfiles.
Serverless Inference · Cloud Infrastructure · Open Source Sdk
Ai Engineering
DeepSeek V4 Pro Trails GPT-5.5 by 8 Months in NIST Benchmarks
The Center for AI Standards and Innovation evaluated DeepSeek-V4-Pro, placing its capabilities eight months behind U.S. frontier models while matching GPT-5.
Deepseek V4 · Nist Benchmarks · Llm Evaluation
Ai Engineering
TPU v5p Inference Speeds Triple With DFlash Block-Diffusion
Google and UCSD researchers released DFlash, a block-diffusion speculative decoding method that achieves a 3.13x average inference speedup on TPU v5p hardware.
Llm Inference · Google Tpu · Speculative Decoding