Independent AI news for builders
Short briefs, source trails, and deeper explainers from a visible publishing pipeline — not a generic technology news feed.
Google is developing a specialized Frozen v2 chip that integrates the Gemini architecture directly into the hardware to increase energy efficiency. The shift from general-purpose accelerators to a "frozen" architecture (model-in-silicon) could radically reduce data center capital expenditures and solve the computing power shortage.
Public structure
Use the news feed for quick updates, explainers for context, the blog for working notes, and projects to inspect the systems behind the publication.
English index for translated briefs. The current live source feed is visible under /ru/news/.
Selected stories rewritten as clean longform articles with context, source trail, and related links.
Evergreen answers about the publishing system, quality gates, and story model without duplicating daily news.
Personal thinking about agents, editorial infrastructure, product design, and open-source work.
TG-NEWS, Task-Plan v2 Dashboard, MCP experiments, and other systems with status and repositories.
TG-NEWS interaction
News desk
Sber is hosting Giga Fun Party, an informal event for ML and AI engineers, on July 29 in Moscow. The goal is to build a professional community around local LLM solutions and provide direct access to engineers working on alignment and data preparation for high-performance models.
Qualcomm AI Research has introduced MobileWan — an optimized version of the Wan2.2 5B model for mobile video generation. It demonstrates the effectiveness of aggressive pruning and distillation for porting DiT models to edge devices.
MetaView has been introduced—a diffusion framework for novel view synthesis from a single image while preserving metric scale. MetaView addresses the critical problem of scale mismatch and geometric inconsistency during large viewpoint changes by combining the power of diffusion models (MM-DiT) with precise geometric cues without the need for heavy 3D reconstruction.
Researchers from Google DeepMind have developed CO2Jump, a training-free sampler that allows text and images to dynamically correct each other during the denoising process. This new approach to multimodal generation addresses the problem of data uncoupling.
An article on Baselane suggests using versioned harness.json manifests to solve the problem of AI agent configuration drift within teams. Moving from scattered instructions to centralized management via manifests reduces operational risks and increases transparency of AI usage across the company.
Apple does not plan to add legendary designer Jony Ive to the lawsuit against OpenAI regarding intellectual property theft. This highlights a shift in corporate dispute focus from software to hardware and the importance of reputation management when involving iconic figures.
The transcribe.cpp library has been introduced for fast, cross-platform speech recognition with support for multiple models via ggml. Optimization for ggml and GGUF simplifies ASR deployment on edge devices and local servers, reducing the need for heavy infrastructure dependencies.
A developer has introduced Dulus, a lightweight Python implementation of Claude Code that supports over 100 backends and 2,000 MCP servers. The project accelerates the transition to decentralized AI systems, reducing developer dependency on specific API vendors and simplifying multimodal agent integration.
The free ZLVOX AI Humanizer tool uses the Groq Llama 3.3 70B model to give AI-generated text a more natural appearance. The availability of LPU-based (Groq) tools accelerates the 'creation—detection' cycle, forcing detection system developers to constantly improve.
Tech giants are facing pressure from investors demanding confirmation of the effectiveness of massive AI investments amid falling Nasdaq 100 and SOX indices. Companies developing AI models and chips will face a rigorous audit of their capital expenditure (CapEx) efficiency.
The Meta Oversight Board has warned that modern AI models could become powerful propaganda tools due to their tendency toward censorship in certain regions. Developers need to implement multilingual audits and account for the risks of political pressure when fine-tuning models.
The author used the Claude model to reverse-engineer and automate the configuration of a MIDI controller and guitar software. This demonstrates the capabilities of LLMs in low-level programming tasks and the analysis of closed data formats.
Longform layer
Explainers collect what happened, why it matters, what changed, source links, related stories, and language editions in one readable page.
Google is developing a specialized server AI chip codenamed Frozen v2, which will partially integrate the Gemini architecture directly into silicon to achieve manifold increases in energy efficiency.
On July 29, Moscow will host the Giga Fun Party—an event by the GigaChat (Sber) team tailored for ML and AI engineers. Instead of traditional lectures, the event will focus on live discussions regarding the future of LLMs, multimodality, and agentic systems.
Qualcomm AI Research has introduced MobileWan—an optimized version of the Wan2.2 5B model that enables high-quality video generation directly on smartphones.
Project lab
These projects show how the news pipeline, planning tools, MCP integrations, QGIS tooling, video-editor copilot work, and deployment experiments are built in public.
Telegram-first AI news pipeline with discovery, story mapping, publication transactions, and channel output.
Dashboard for structured multi-agent task plans, implementation gates, review state, and execution visibility.
Two Agent Skills for validating image layers and building a soft cursor-trail canvas reveal hero with tuning controls and browser checks.
Public repositories for QGIS agent workflows, OpenCut copilot work, Telegram and Threads MCP servers, research search, and model deployment experiments.
Community
About
I build AI news infrastructure, Telegram automation, local knowledge workflows, and agentic developer tools. mlllm.io is the public surface for that work: a news desk, article archive, blog, project index, and collaboration profile in one coherent system.
The goal is simple: keep useful public pages, visible source trails, related projects, and multilingual editions in one place readers and agents can understand.