2026

Transformers at the wheel: how Yandex uses AI to control self-driving vehicles

Google DeepMind's Gemma 4 arrives on Amazon Bedrock: three models with MoE and multimodality

Anthropic Releases Claude Fable 5 and Mythos 5: One Model, Two Access Tiers

LLM·AI News

AI Red Teaming — How Companies Test AI Systems for Vulnerabilities Before Release

AWS Strands Evals automatically identifies root causes of AI agent failures and suggests fixes

Csquare Files for IPO in US: Data Center Backed by Brookfield Goes Public

LLM·NVIDIA Developer Blog

NVIDIA BioNeMo lets you fine-tune biological AI models with LoRA in hours

NVIDIA has released BioNeMo Recipes — ready-made pipelines for fine-tuning the ESM2 and Evo 2 language models using the LoRA method. Special

01Amazon Bedrock Guardrails Gets a New API for Flexible Protection of Agentic AI ApplicationsJun 30, 2026·2 min 02LangSmith Added End-to-End OpenTelemetry Support for LangChain and LangGraphJun 30, 2026·2 min 03LangChain released Promptim, a library for automatic prompt optimizationJun 30, 2026·2 min 04Sarah Guo of Conviction: why LLM won't replace work that can't be measuredJun 30, 2026·2 min 05New Computer Raised Agent Memory Search Accuracy by 50% with LangSmithJun 30, 2026·2 min

Top themes

LLM·3DNews AI

Qualcomm Develops Over 40 AI Devices and Predicts the End of the App Era

Qualcomm CEO Cristiano Amon believes AI agents will render traditional applications unnecessary — the company is already developing around 4

01Why AI pilots don’t make it to production: what must be in place before scalingJun 30, 2026·2 min 02LangChain Launches Headless Tools: Agents Work Directly in the BrowserJun 30, 2026·2 min 03Vulnerability in Microsoft Copilot Allowed Stealing Two-Factor Authentication CodesJun 30, 2026·2 min

LLM·TNW

YouTube deleted 16 channels with 35 million subscribers in AI-slop purge — and caught honest creators

In January 2026, YouTube simultaneously blocked 16 channels with 35 million subscribers and 4.7 billion views — among those deleted were rea

LLM·Guardian

Anthropic Fable classified as a munition: the US blocked the AI model worldwide

The US classified Anthropic's Fable as a military munition and barred foreign nationals from accessing it — the company was forced to disabl

LLM·Machine Learning Mastery

Scikit-LLM: multi-label text classification without a training dataset using LLMs

The Scikit-LLM library makes it possible to assign multiple tags to each text without a training dataset, using GPT-4 and compatible models through a familiar scikit-learn interface.

LLM·Habr AI

AI assesses Russians' job suitability from their faces — developers move into the public sector

LLM·Meta Corporate News

Meta explains the difference between GPU, CPU, and its own MTIA chips in AI infrastructure

How to speed up transformer training with NVIDIA Apex: FusedAdam, FusedLayerNorm, and torch.amp

LLM·NVIDIA Developer Blog

NVIDIA speeds up LLM pretraining: NVFP4 on Blackwell paired with JAX and MaxText

LLM·AWS Machine Learning Blog

JetBrains released Mellum2 — an open 12B MoE model for fast tasks in AI pipelines

LLM·NVIDIA Developer Blog

NVIDIA: how to design battery energy storage systems for AI factories

Claude Fable 5 arrives on GitLab Duo — multi-day agent tasks without manual stops

GitLab launched a managed DevSecOps platform on Google Cloud with Gemini 3.5 Flash

Alibaba launched Qwen3.7-Plus: multimodal agent with video analysis and autonomous coding

Alibaba released Qwen3.7-Plus, a multimodal agent, on the Bailian platform — the model can process images and video, invoke tools, and itera

01LangGraph gets fault tolerance: retries, timeouts, and error handlers for production agentsJun 30, 2026·2 min 02OpenAI launches Codex: the AI agent that writes code instead of engineersJun 30, 2026·1 min 03Google released Gemma 4 12B: 150 million downloads, runs on a laptop with 16 GB of VRAMJun 30, 2026·3 min 04LangChain: why model neutrality protects businesses from vendor lock-inJun 30, 2026·2 min 05Together AI beat TensorRT-LLM by 31% in benchmarks for code agentsJun 30, 2026·2 min

Top themes

LLM·@Replit

Replit: describe an idea in words — get a complete app with a database and deployment

Replit combines code generation, UI, authentication, databases, and deployment in one place — plus parallel AI agents and real-time team col

01Anthropic taught Claude chemistry: Opus 4.7 outperformed specialized NMR softwareJun 30, 2026·3 min 02Together AI at NVIDIA GTC 2026: Dynamo, multi-agent models, and voice AIJun 30, 2026·2 min 03Together AI launches MiniMax M3 with 1 million-token context and multimodal supportJun 30, 2026·2 min

LLM·DeepMind Blog

Google DeepMind: Gemini AI tutor gave Sierra Leone students 1.5 years of progress in 8 weeks

Google DeepMind published the results of a randomized study: students who studied math with Gemini advanced by 1.5 school years in 8 weeks.

LLM·@Replit

Replit launched parallel agents: an app, a video, and a pitch deck at the same time

A developer created a mobile app, an ad video, and a pitch deck for a travel startup in a single session — using Replit’s new parallel agent

LLM·xAI News

Gopuff launched Grok-based AI agent Go — it handles voice, photos, and text

Fast-delivery service Gopuff, together with xAI, integrated the Grok-based AI assistant Go into its app — it understands customers' text, voice, and visual requests.

Together AI received ISO 27001:2022 certification for enterprise AI workloads

LLM·Together AI Blog

eToro integrated xAI models into Tori agent for real-time market sentiment analysis

PyGraphistry: Graph analytics for threat detection in enterprise data

MIT gathered experts to discuss ethics in AI and computing technologies

Google DeepMind released DiffusionGemma — an open 26B MoE model with 4x faster generation

Microsoft SkillOpt: automatic prompt optimization instead of manual trial and error

HSBC deploys Google Cloud AI targeting $100+ million from each project

Cohere's North Mini Code: open-weight 30B MoE model for agentic coding

Cohere has released its first open-weight model for developers — North Mini Code, built on MoE architecture with 30B parameters, a 256K context window, and the ability to run on a single H100 GPU.