mlllm.ioAI news and builder lab
AI channel Telegram Threads GitHub

Independent AI news for builders

AI news, clear context, and the projects behind the work.

Short briefs, source trails, and deeper explainers from a visible publishing pipeline — not a generic technology news feed.

Read latest news How this works Follow Telegram
Live from TG-NEWS 2026-10-20
Anthropic logo
AI_Coding EN

Anthropic Replaced Git with Scion System for AI Agent Code

Anthropic published a research paper on its proprietary version control system, Scion, which replaced Git for Claude agent code in 5 months by 7 engineers as daily merges grew from 2,500. For the first time, the version control infrastructure designed for AI agents rather than humans is publicly dissected: commits live on the server as write-ahead log records, blame and diff are served via API, and data is distributed across a hierarchy of scion-daemon, scion-cache, and scion-servers caches on top of Spanner.

Read briefRead longformRussian brief
Stories tracked901Short briefs with source trails and paired longform links when available.
Explainers899Expanded pages that carry the durable context behind selected stories.
Media assets965Verified images, video posters, and source media attached to story records.
Languages planned8EN/RU first; other locales stay gated until translation quality is ready.

Public structure

Start with the feed, then go deeper.

Use the news feed for quick updates, explainers for context, the blog for working notes, and projects to inspect the systems behind the publication.

/news/

Short news feed

English index for translated briefs. The current live source feed is visible under /ru/news/.

/articles/

Explainers

Selected stories rewritten as clean longform articles with context, source trail, and related links.

/topics/

Topic pages

Evergreen answers about the publishing system, quality gates, and story model without duplicating daily news.

/blog/

Builder notes

Personal thinking about agents, editorial infrastructure, product design, and open-source work.

/projects/

Project index

TG-NEWS, Task-Plan v2 Dashboard, MCP experiments, and other systems with status and repositories.

TG-NEWS interaction

How a story moves from source to site.

  1. IngestPublic sources, research posts, company updates, repositories, and Telegram-native signals.
  2. MapDomain, vendor, importance, duplicate groups, source trail, and short trend signal.
  3. VerifyCheck facts, source links, repeated context, weak claims, and publication quality before release.
  4. LocalizePrepare English and Russian editions with language-specific headlines, terms, and links.
  5. PublishStatic HTML pages, RSS, sitemap, news sitemap, llms.txt, and cross-language links.

News desk

Latest English AI stories

English story index

Source-backed English briefs selected from TG-NEWS.
AWS Physical AI Toolchain
CodingNVIDIA2026-10-20

AWS open-sources Physical AI Toolchain for robotics

Amazon Web Services released the open-source Physical AI Toolchain — a set of reference architectures and Terraform modules that automate the full AI development lifecycle for robots: synthetic data generation, training, simulation validation, edge deployment, and fine-tuning based on feedback data. AWS has translated the typical physical AI path — from teleoperation data to trained policies, simulation, and edge deployment — from manual assembly into open infrastructure-as-code built on managed

Read briefRead longformTelegram
Voxtral 3: Voice Agent SDK & Speech Models in 50+ Languages
CodingOpenAI2026-10-20

Mistral Unveils Voxtral 3 — Voice Models and SDK for Voice Agents

Mistral AI has introduced Voxtral 3, a family of speech recognition and synthesis models featuring a LiveKit-based Voice Agent SDK that integrates listening, LLM reasoning, tool calling, and speech output into a single streaming pipeline with sub-second latency. This unified pipeline—listening → LLM reasoning → tool calling → speech output—combined with open weights and self-deployment, replaces multi-vendor 'STT + LLM + TTS' stacks. This reduces costs and third-party audio transmission risks fo

Read briefRead longformTelegram
Feature image of the HiRO-ACEv2 announcement on the Ai2 blog
ResearchHuggingFace2026-10-20

Ai2 releases HiRO-ACEv2, an AI climate emulator with 3 km resolution

The Allen Institute for AI (Ai2) has released the HiRO-ACEv2 system, which generates global fields of precipitation, 10 m wind, and sea-level pressure at 3 km resolution at a cost roughly 300 times lower than the storm-resolving X-SHiELD model. HiRO-ACEv2 drastically reduces the cost of generating 3 km resolution atmospheric data—a scale necessary for studying downpours, hurricanes, and atmospheric rivers—which global models like X-SHiELD (NOAA) provide only at enormous computational expense.

Read briefRead longformTelegram
Multimodal2026-10-20

FLUX 3 Max: 10-second video in 11 seconds of generation

On October 19, 2026, the fal platform, in collaboration with Black Forest Labs, released FLUX 3 Max — a variant of the FLUX 3 model refined through post-training, RL, and inference optimization: according to fal, a 10-second video is generated in 10.95 seconds with an Elo of 1139 in the Artificial Analysis Video Arena. Generation speed has dropped to a level close to the duration of the clip itself: production teams can iterate through dozens of scene variations in minutes instead of waiting min

Read briefRead longformTelegram
REA in Claude Code: test results, costs and prompts
CodingAnthropic2026-10-20

REA: Agentic Reverse Engineering Without Source Code

The open-source MCP server REA (morluto/rea, MIT) enables agents in Claude Code and Codex to reverse-engineer software without source code — from native binaries via Ghidra, Hopper, and IDA to .NET assemblies, APKs, firmware, and websites — and has gathered 240k stars on GitHub by mid-October 2026. REA turns classic disassemblers (Ghidra, Hopper, IDA) into tools controlled by an LLM agent via MCP, adding an Evidence Contract — a schema that separates observations, inferences, and unknowns.

Read briefRead longformTelegram
SonicBoom: LLM Inference in C++ and CUDA without Python
CodingOpenAI2026-10-20

SonicBoom: LLM Inference in C++ and CUDA without Python

HardenedVault (HardenedLinux) has released SonicBoom on GitHub — a runtime for local LLM inference in C++/CUDA without Python. In v0.0.3 on an RTX 3050, it decodes Gemma 4 E4B (53.6 tok/s) and Qwen2.5-0.5B (345.9 tok/s), running 13–20% faster than llama.cpp. SonicBoom is one of the first open-source inference runtimes where the speed advantage comes not only from hand-written CUDA kernels: the graph scheduler fuses adjacent operations and records the entire decoding step into a single CUDA Graph

Read briefRead longformTelegram
lodestones/Kromoss — MOSS-Image Fine-Tune & Turbo LoRA (Hugging Face)
MultimodalHuggingFace2026-10-20

An ecosystem grew around the open-source MOSS-Image in just 4 days

Four days after the release of OpenMOSS's open 5B image generation model MOSS-Image (Apache-2.0), the community released the Kromoss fine-tune at 1024px, a 4-step turbo-LoRA, a set of nodes for ComfyUI, and support in Fizgig 8.5. MOSS-Image is one of the few open 5B generation models with strong text rendering and JSON layout with boxes, and an ecosystem on par with FLUX/SDXL has formed around it in just four days.

Read briefRead longformTelegram
The Gate That Costs Forever: A Forgotten 1980 Reversible Computer, And The AI That Just Cut Quantum Circuit Costs Seventeenfold
ResearchAnthropic2026-10-20

AI cuts quantum circuit costs by 17x

A team from Harvard and Anthropic, in preprint arXiv:2610.22367, combined LLM Claude Mythos 5, evolutionary selection, and independent classical verification to find quantum circuits cheaper than known ones, including a 17-fold reduction in gate count for elliptic curve point addition. LLMs are improving the resource characteristics of basic quantum primitives in a closed loop of "proposal — independent verification — evolutionary selection" (OpenEvolve + MAP-Elites) for the first time: each cir

Read briefRead longformTelegram
Amazon's $220 Billion AI Bet Tests the Limits of Its Empire
BusinessGoogle2026-10-20

💼 AWS CFO Joked That AI Pushed Amazon Into Negative Cash Flow

Business Insider published a recording of an internal AWS meeting where CFO John Felt joked that AI investments have made Amazon negative in free cash flow: the company is spending $220 billion in 2026 on data centers, and the head of AI infrastructure confirmed that the ML/AI business will remain "massively negative" for another couple of years. The hyperscaler race has shifted from a chip shortage to a shortage of electricity, construction, and labor: the principle of "buying capital at the la

Read briefRead longformTelegram
DeepSeek
SafetyOpenAI2026-10-20

US Bans DeepSeek in Federal Agency IT Systems

The US Department of Homeland Security issued a mandatory order to exclude DeepSeek and all its products from the IT systems of federal civilian agencies, requiring the removal of deployed versions and prohibiting future purchases. This marks the first time the FASCSA law has been applied to a specific AI company: what individual agencies previously decided at their discretion has become a mandatory rule for all federal civilian structures.

Read briefRead longformTelegram
MediationBench compares AI mediators on simulated conflicts.
ResearchAnthropic2026-10-19

MediationBench: Testing AI mediators in 62 simulated conflicts

Jonathan Hendler released the MediationBench benchmark, where language models play both sides in 62 scenario-based disputes, while the model under test acts as a mediator, deciding after each message whether to speak, wait, or end the conversation, and receives a final score of 0–100 from judge GPT-5.6 Sol. MediationBench provides the industry with the first public protocol for auditing AI mediation: a fixed set of disputes, preregistered rules, 38 conflicts held out in a holdout set, a primary

Read briefRead longformTelegram

Longform layer

Read the context behind the news.

Explainers collect what happened, why it matters, what changed, source links, related stories, and language editions in one readable page.

Project lab

See the systems behind the publication.

These projects show how the news pipeline, planning tools, MCP integrations, QGIS tooling, video-editor copilot work, and deployment experiments are built in public.

Production

TG-NEWS

Telegram-first AI news pipeline with discovery, story mapping, publication transactions, and channel output.

AI newsTelegramPipeline
Public OSS

Task-Plan v2 Dashboard

Dashboard for structured multi-agent task plans, implementation gates, review state, and execution visibility.

AgentsDashboardOSS
Public OSS

Mouse Trail Masking Reveal

Two Agent Skills for validating image layers and building a soft cursor-trail canvas reveal hero with tuning controls and browser checks.

CanvasAgent SkillsUX
Public OSS

MCP, QGIS, and editor tools

Public repositories for QGIS agent workflows, OpenCut copilot work, Telegram and Threads MCP servers, research search, and model deployment experiments.

MCPQGISOpenCut

Community

Follow the public channels.

Telegram channel@lookatainews: short AI news and current public output. Telegram profile@Serge_Kost: direct public contact and personal updates. Threads profile@sergekost: short thoughts, links, and public notes. GitHub profileRepositories, experiments, and open-source proof.
Sergey Kostenchuk

About

Sergey Kostenchuk / mlllm

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.