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

Latest AI news

Source-backed AI news briefs selected from the TG-NEWS pipeline.

Public story index

EN · 448 records

Showing the latest 40 of 448 published records. Detail pages remain available through sitemap and internal links; date archives are the next production step.

Riemann Hypothesis Proved in Lean: 17 Million Lines of Code Pass Independent Verification

Levent Alpöge published a proof of the Riemann Hypothesis as a Lean library of 17 million lines, and Lean FRO and the Mathlib Initiative confirmed that the proof has been verified and is correct. This is the first instance where one of the great unsolved problems (the Riemann Hypothesis) has been claimed as proven and formally verified by independent kernels (Lean kernel, nanoda, con-leche, lean4lean)—a precedent for formal verification: "machine-checked proof" is now applicable even to problems

Read briefRead longformTelegram
Inkling 2 Mini model card thumbnail

Thinking Machines Lab releases Inkling 2 Mini — an open multimodal model with loops

Thinking Machines Lab released Inkling 2 Mini on November 20, 2026 — an open multimodal model with an Apache 2.0 license, 26.4B total parameters, 3.4B active parameters, and a context window of up to 1M tokens. This looped transformer with open weights is the first public example of "depth scaling": computational depth scales by re-passing the same layers rather than adding parameters, which changes inference economics (one ~20 GB checkpoint instead of dozens) and sets a reference for an archite

Read briefRead longformTelegram
Jensen Huang, president and CEO of Nvidia, waits for a groundbreaking ceremony for an expansion of Coherent's manufacturing facility to begin on June 16, 2026, in Sherman, Texas. (AP Photo/Jeffrey McWhorter, File)

Jensen Huang Ready to Return to Nvidia, but Board Opposes

On November 17, 2026, Jensen Huang stated in a post on X that he is ready to return as CEO of Nvidia if the board of directors determines it is in the best interest of shareholders, employees, and the company. However, the board responded the same day that it unanimously accepted his resignation and has no intention of reinstating him. The resignation and the board's refusal place the management of the largest AI chip supplier in a transitional mode under interim CEO and CFO Colette Kress. The a

Read briefRead longformTelegram
We built Australia's own AI model, and today it's open to the world

Australia Opens Up Sovereign AI Model Agentaus 1 to the World

Australian company Trellis Data has opened access to the sovereign LLM Agentaus 1, featuring 220 billion parameters. According to the company, it scores 84.4% on MMLU-Pro, ranks 43rd among the world's foundational models, and operates exclusively on Australian servers. Australia is bringing its own near-frontier LLM to market for the first time, shifting from being a consumer of AI to a producer: Agentaus 1 is deployed in a sovereign cloud, on-premises, or fully air-gapped—targeting the governme

Read briefRead longformTelegram
GPT-6.2 Astra benchmark results

GPT-6.2 released before GTA 6: OpenAI launched two models at once — Astra and Sol

On November 17, 2026, OpenAI released two GPT-6.2 models: Astra for complex tasks (code, documents, 3D, cybersecurity) at $8/$40 per million tokens, 20% cheaper than its predecessor, and Sol for everyday tasks at $2/$10. Both are available in ChatGPT Work, Codex, and the API. The 20% price reduction for Astra intensifies OpenAI's pressure on Anthropic (Claude 5.2) and Google (Gemini 4) in the corporate segment, where the key metric is the cost per solved task.

Read briefRead longformTelegram
LiquidAI/Liquid-Codec (Hugging Face model card)

Liquid AI releases speech codec for its LFM3 audio models

Liquid AI has made its 74M-parameter streaming audio codec, LiquidCodec, available (Hugging Face, LFM Open v1.0 license), where GFSQ quantization instead of RVQ allows the language model to predict all 12 codebooks of a frame in parallel. GFSQ instead of RVQ removes sequential dependencies between codebooks, so the LLM predicts speech tokens in parallel and the number of sequential autoregression steps for audio-LLMs is reduced.

Read briefRead longformTelegram
Two engineers smile in the Nevada desert as the heavy-lift drone flies behind them. Project Heavy Lift: six days, a small team, and an AI agent.

AI agent built a cargo drone in 6 days

Startup Normal published a breakdown of Project Heavy Lift, in which three engineers used the AI agent Zeus to design, order parts, and assemble a cargo drone weighing approximately 25 kg in 6 days. The drone made its first flight on November 5–7 under FAA Part 107 rules and delivered a payload. This is one of the first publicly documented cases where an AI agent participated in the development of a physical device at all stages.

Read briefRead longformTelegram
Face clusters feature demonstration

PixlStash 2.0: Image Library and LoRA Training in One Application

PixlStream 2.0 (v2.0.0 released November 10, 2026, v2.0.1 released November 16) is an open-source desktop application that groups faces by similarity, trains LoRA models via AI-toolkit on a local GPU, and automatically attaches the finished model to a character for use in ComfyUI. PixlStash 2.0 streamlines the model personalization pipeline: image collection, face clustering, LoRA training, and ComfyUI integration all happen within a single tool, lowering the barrier to entry for creating person

Read briefRead longformTelegram
单细胞图谱与 RNA 速率推断的细胞状态变化方向

Moonshot AI to steer Kimi models into biology, chemistry, and pharma

Chinese startup Moonshot AI announced at Web Summit in Lisbon that it is redirecting its Kimi models from coding to biology, chemistry, and pharmaceuticals, with the Kimi K3.1 Pro model already integrated into the BioMaster scientific agent by Beijing-based DP Technology. Moonshot AI is shifting the focus of its flagship Kimi models from code to science, positioning Kimi K3.1 Pro as a tool for AI-for-Science.

Read briefRead longformTelegram
SDXL, a photo of a dog: six seed-matched samples of the student and of ReDiverse

ReDiverse Restores Diversity to Few-Step Generators

Researchers from UIUC, fal, and Johns Hopkins published the ReDiverse preprint (arXiv:2611.18550, November 13, 2026), explaining the 'diversity collapse' in few-step distilled generators as the compression of teacher variance into fewer directions and restoring it using rank-32 LoRA or a training-free method. Few-step distillation (DMD2, TDM, lightx2v, Turbo) is the standard way to accelerate diffusion generators, and its main side effect—identical results with different seeds—is now explained a

Read briefRead longformTelegram
Evolving Infra: Agent-Adaptive LLM Serving at Scale (arXiv:2611.12304)

LLM Serving Adapted to Agentic Workloads

A team from MIT Han Lab, along with Georgia Tech, UIUC, UC Berkeley, NVIDIA, and OpenAI, published Evolving Infra on arXiv (2611.12304). This open system adapts LLM serving to agentic workloads online, without offline profiling, and reduces average latency by up to 4.1x and P95 TTFT by up to 14x compared to NVIDIA Dynamo on 256 GB300 GPUs using GLM-5.2 and Kimi K2.6 models. Agentic workloads (coding agents, computer-use) break static prefill-decode clusters: context compaction invalidates the KV

Read briefRead longformTelegram
TokenEconomizer: Learning from Experience to Make AI Coding Agents More Efficient

TokenEconomizer: Microsoft Teaches Coding Agents to Save Tokens

Microsoft Research, in collaboration with UC Berkeley, Microsoft Security, and Voyage AI, has released an open-source plugin called TokenEconomizer. By learning from past sessions in the same repository, it reduces token consumption for coding agents: across 277 SWE-Bench tasks, costs dropped by an average of 18.4–30.9% while Pass@1 remained stable or improved in 4 out of 5 agent and model combinations. Profiling Claude Code on 60 SWE-Bench tasks revealed that a significant portion of agent pipe

Read briefRead longformTelegram
Open AI Scientists — Zitnik Lab, Harvard Medical School (official project site)

Open AI Agent OpenScientist Automates Biomedical Research

On November 16, 2026, an article about OpenScientist was published in Nature Medicine. It is an open-source agentic AI 'co-scientist' system from the C-BRAIN consortium, led by Washington University in St. One of the first peer-reviewed examples of an open-source agentic AI system tested on real clinical cohorts: external validations, randomized negative controls, and blind proteomic comparison with independent human analysis set the reproducibility standard for 'AI co-scientists' in biomedicine

Read briefRead longformTelegram
Hero image of the Hark Handoff² announcement article

Hark updates its browser-based agent

Brett Adcock's Hark lab has released Handoff², the second generation of its browser-based computer-use agent, which the company says is on average twice as fast as the first version and is now rolling out to all Hark Pro plans. Handoff² demonstrates that computer-use agent speed can be improved without scaling up the model: Hark co-trains the harness and the model using an RL recipe with penalties for extra steps, excessive thinking tokens, and latency, allowing the agent to learn to batch obvio

Read briefRead longformTelegram
US issues rule to protect advanced AI chips from being accessed remotely | Reuters (via Yahoo News)

US tightens AI chip export controls to counter GPU leasing

On November 16, 2026, the US Department of Commerce issued an interim rule requiring licenses for the export, re-export, and transfer of AI chips and servers to data centers if the compute resources may be accessed by companies from 23 countries in the D:5 group, including China, Russia, Iran, and North Korea. The rule closes a loophole through which Chinese laboratories leased compute power on American chips in foreign data centers.

Read briefRead longformTelegram
An AMD Instinct MI350 Series accelerator module seen from above, with its chip package exposed at the center of a black circuit board

Claude Mythos adapted model for AMD GPUs in three days instead of 6–12 months

Anthropic told investors during its IPO roadshow that the Claude Mythos Preview model, with "almost no human involvement," ported or wrote nearly 100% of the kernels for the AMD Instinct MI355X GPU in approximately three days, whereas such adaptation historically took 6–12 months for teams of dozens of engineers. CUDA — Nvidia's nearly 20-year-old software platform — has historically been the main barrier to switching to other accelerators: code for CUDA had to be manually rewritten for ROCm/HIP

Read briefRead longformTelegram

Tim Sweeney Defends Nvidia CEO Amid Press Campaign

On November 15, 2026, Epic Games founder and CEO Tim Sweeney posted on X in defense of Nvidia CEO Jensen Huang, stating that critics have walked back their initial accusations and calling for an investigation and open court instead of a press campaign. The statement identifies a specific mechanism: publications with suspiciously unverified sources of accusations (according to Sweeney, one lobbyist from a competing company) trigger a process that breaks corporate governance, the cost of which he

Read briefRead longformTelegram
product-seo

Open-source product-seo skill released for coding agents

The author of the Tips AI Telegram channel (Igor Trunin) released the open-source product-seo skill v0.1.0 (MIT license) for Claude Code, Codex, and Cursor coding agents. It operates from Google Search Console CSV exports and was tested on the live HypeScribe site, where CTR increased from 0.4% to 0.9%. The skill formalizes product site SEO into rules executed by an LLM agent: query triage based on the "is this a product answer" criterion, a separate agent that collects competitor facts only fro

Read briefRead longformTelegram
Looki Loom Model announcement cover image

Looki Loom: Wearable Camera as Long-Term Memory for AI

Chinese company Looki has unveiled Loom, a family of multimodal models that transforms video and audio captured throughout the day by the Looki L1 wearable camera into structured long-term memory linked to timestamps and original recordings. Loom demonstrates a specific path to reducing the cost of personal AI assistants: hierarchical memory finds relevant experiences via summaries before processing raw recordings, reducing token usage (6.9 times less than mem0) and inference costs (Sense claims

Read briefRead longformTelegram
Baseten open-sources UltraFast engine for DeepSeek-V4.1-Flash

Baseten open-sources UltraFast engine for DeepSeek-V4.1-Flash

Baseten has open-sourced the deepseek-ultrafast inference engine, written in Rust and CUDA specifically for DeepSeek-V4.1-Flash on NVIDIA B200/B300 GPUs: the README claims 2060 tokens/s decoding on 8x B300 and 1M token prefill in 11.7 seconds. Speeds of 2000+ tokens/s on standard B200/B300 were previously only achieved by specialized chips from Cerebras and Groq; the open-source engine makes this achievable on standard GPUs.

Read briefRead longformTelegram

Russian teams improve recommendation systems — three papers at ACM CIKM '26

Teams from T-Bank, MIPT, HSE, and Innopolis University presented three papers on recommendation systems at the ACM CIKM '26 conference in Rome — the E²L³AE model with joint training of an LLM-embedding text encoder and a linear autoencoder, the graph-based MoE-IGR with dynamic routing between “experts,” and the “Split Decisions” paper on the model-selection regret metric. Cold-start is a bottleneck in industrial recommendation systems: new items and new users lack interaction history.

Read briefRead longformTelegram
U.S. plans to name Nvidia among AI 'national champions', WSJ reports — Yahoo Finance

Trump Prepares List of National AI Champions: Nvidia, OpenAI, and Anthropic

According to The Wall Street Journal (November 15, 2026), the Trump administration intends to designate more than 24 American companies, including Nvidia, Amazon, Microsoft, OpenAI, and Anthropic, as “national AI champions” to promote ready-made “chips + cloud + models” packages in foreign markets in competition with Huawei, DeepSeek, and MiniMax. The US is shifting from passive export restrictions to an active export policy: the state is promoting the “chips + cloud + models” bundle as a single

Read briefRead longformTelegram
YooKassa offers 2,000,000 ₽ prize pool at MCP hackathon

YooKassa offers 2,000,000 ₽ prize pool at MCP hackathon

YooKassa announced an online hackathon, “Agentification of Payments with YooKassa MCP,” scheduled for December 5–6, 2026, with a prize pool of 2,000,000 ₽ and team registration open until December 4 for groups of up to five people. The hackathon serves as a public signal that payment providers are building agent commerce on MCP: AI agents gain direct access to payment infrastructure, including transactional operations—payments, invoices, refunds, and recurring charges—rather than just data readi

Read briefRead longformTelegram
Hy4flash

Tencent releases Hy4 Flash: a budget model with 224 billion parameters

Tencent unveiled the budget model Hy4 flash on November 16, 2026: a 224-billion-parameter MoE (12 billion active per token) with a 1 million token context, which, according to vendor measurements, costs 1/20 of Hy4 per task while maintaining satisfaction in blind tests comparable to Claude Haiku 5.5. Tencent added a Flash price tier to the Hy4 series, which, according to vendor measurements, reduces task costs by 16–20 times while maintaining quality on everyday tasks—thanks to Engram conditiona

Read briefRead longformTelegram
Alibaba Releases Qwen4-Max: 2.4 Trillion Parameters, Per-Task Cost Slashed by 60%

Alibaba releases Qwen4-Max: 2.4 trillion parameters and 62% reduction in task costs

On November 16, 2026, Alibaba released the flagship Qwen4-Max model featuring a 2.4 trillion parameter MoE architecture, a 1 million token context window, and, according to vendor data, a reduction in average task cost from $4.43 to $1.67. The claimed 62% reduction in average task cost (from $4.43 to $1.67) alongside improved quality on Work360 (52.69 vs. 47.74 — Qwen self-report) makes long agent scenarios—programming, document analysis, and industrial workflows—economically viable for a broade

Read briefRead longformTelegram
Automatar AI guitar hero image

MIT Media Lab Creates Automatar — a Bass Guitar with an AI Model in the Body

MIT Media Lab researcher Antonis Christos (Opera of the Future group) built the Automatar bass guitar with an NVIDIA Jetson Orin NX single-board computer in the body: a local AI model analyzes playing on the four main strings via MIDI and electromagnetically excites five sympathetic strings through high-voltage operational amplifiers. Automatar is a clear example of physical AI: instead of cloud-based LLM assistants, it features local edge-AI on the Jetson Orin NX, physically integrated into an

Read briefRead longformTelegram
Photo collage of a photo of the US Capitol building coming out of a computer window

Congress to tighten AI chip exports after Nvidia CEO resignation

According to The Verge, following the resignation of Nvidia CEO Jensen Huang on November 12, 2026, amid a spying scandal involving a suspected Chinese intelligence agent, the US Congress is accelerating the tightening of export rules for AI chip supplies to China. Tightened export controls could limit Chinese companies' access to advanced Nvidia processors, impacting global supply chains, cloud provider strategies, and the pace of AI infrastructure development in Asia.

Read briefRead longformTelegram
A gauge-invariant measure for lattice chiral fermions (arXiv:2610.09093, Nathaniel Craig)

Physicist and GPT-6 Astra Take a Step Toward Defining the Standard Model on a Lattice

On November 15, 2026, OpenAI published the first guest post in its new science blog, where physicist Nathaniel Craig, together with the GPT-6 Astra model, described a candidate non-perturbative lattice definition of the Standard Model of particle physics, released as preprint arXiv:2610.09093. This is a notable example of applying a frontier LLM to fundamental physics: GPT-6 Astra proposed a method for transferring fields to finer lattices, demonstrated the convergence of comparisons, and undert

Read briefRead longformTelegram
SpaceX's mega rocket Starship returns to its launch pad to be caught by the launch tower following a test flight from Starbase, Texas, Saturday, Nov. 14, 2026 (AP Photo/Eric Gay)

SpaceX catches returning Starship from orbit for the first time

On November 14, 2026, SpaceX captured the Starship spacecraft (Ship 42, 15th test flight) returning from orbit using the mechanical 'chopsticks' of the Mechazilla tower after approximately 22 hours in orbit, 14 orbits, and the deployment of 26 Starlink satellites. The first capture of the spacecraft proves that both stages of Starship can return, paving the way for SpaceX's stated launch economics: with full reusability, the cost of launching a kilogram should drop from thousands of dollars (Fal

Read briefRead longformTelegram
GettyImages-2262902677

Alpha AI School Fails to Deliver on Core Promise Despite $75,000 Annual Tuition

Fortune, citing internal data reviewed by The New York Times, reports that in the Alpha AI school network, which costs up to $75,000 per year, only 42% of students achieved the goal of “learning twice as fast” in both math and reading, while about half of K-8 students did not meet the math target. This is the first major public breakdown of the actual results of the flagship AI school, which was positioned as proof that an AI tutor can replace teachers: across a network of ~1,500 students in ~50

Read briefRead longformTelegram
Operate Claude Managed Agents inside your security perimeter

OpenAI launches ChatGPT Wallet payment service

On November 13, 2026, OpenAI launched ChatGPT Wallet, a payment service for AI-agent purchases in partnership with American Express, Mastercard, and Visa. Users confirm each purchase via biometrics, while the payment network issues a token instead of card details. Agentic commerce gains an infrastructure layer: payment tokenization via Visa, Mastercard, and American Express, along with Web Bot Auth for recognizing signed agent requests, lowers the barrier for merchants. The Commerce Partner Play

Read briefRead longformTelegram
Open-frame server with a copper-pipe CPU cooler, green PCIe switch cards, blue braided cables, and an RTX 3090 graphics card behind them

Lawyer Runs DeepSeek V4.1 Flash on Eight Used Nvidia V100s

Lawyer Nick Mermidjes upgraded a server with eight used Nvidia V100s by adding 1 TB of RAM and an RTX 3090, then ran DeepSeek V4.1 Flash on it, converting 5,949 of 15,360 MoE experts to 4-bit NVFP4 while keeping the rest at 2-bit: context expanded to 524K tokens, and token match with the reference NVFP4 version reached 95.1% at the cost of slower generation. Mixed per-expert quantization—“hot” experts in NVFP4, “cold” ones in 2-bit—accounts for the fact that quality in MoE models suffers more fr

Read briefRead longformTelegram
Hero image for The Register article on AI-prompted patch flood in Linux

Torvalds: Linux kernel AI patch flood may have peaked

In his weekly update for Linux 7.4 rc2, Linus Torvalds suggested that the influx of AI-driven fixes, triggered by Anthropic's Mythos vulnerability discovery model, may have peaked: in the eBPF subsystem, fixes are now arriving faster than new bug reports. eBPF is the first major kernel subsystem to reach an equilibrium point where AI models generate fixes faster than bug reports.

Read briefRead longformTelegram
Attack chain on Microsoft 365 Copilot demonstrated: from prompt injection to sandbox escape

Attack chain on Microsoft 365 Copilot demonstrated: from prompt injection to sandbox escape

Rubrik Zero Labs researcher (Ori) published a YouTube recording of their Black Hat USA 2026 talk, “ChatMate,” demonstrating an attack chain on Microsoft 365 Copilot: indirect prompt injection via a compromised Word document, code execution within the Copilot sandbox, and a sandbox escape via CVE-2026-32193 (CVSS 8.8) in Azure Kubernetes Service, which provided access to the victim's Copilot session and files in SharePoint and OneDrive. The talk details the specific mechanism by which prompt inje

Read briefRead longformTelegram
Xania Monet creator Telisha "Nikki" Jones

AI Music Gets Its Own 'Grammys'

On November 13, 2026, the Sonic Intelligence Academy held the first AI Music Awards ceremony in Los Angeles, where awards in 25 categories were distributed by vote among AI artists registered in its registry, with AI persona Xania Monet taking the top prize. SIQA is building a separate infrastructure for legitimizing AI music—its own charts (launched in January 2026 after Billboard's refusal), a Verified Registry, 25 award categories, and a red-carpet ceremony.

Read briefRead longformTelegram
EnvOpt: Automated Optimization for Agent Environments

EnvOpt: Automatic Optimization of LLM Agent Environments

Bespoke Labs, in collaboration with the University of Chicago and UC Berkeley Sky Computing Lab, has introduced EnvOpt, a framework that automatically optimizes the file environment of LLM agents. According to the authors, it saves an average of 48% of execution costs across five benchmarks. EnvOpt identifies the "environment" as a third, previously immutable optimization target alongside model parameters and the harness.

Read briefRead longformTelegram
Two pilots at work during departure of Dallas Fort Worth Airport in United States of America.

FAA deploys AI system SMART across all US air traffic control centers

As of November 10, 2026, the FAA has activated the $875 million AI system SMART in all 22 US air traffic control centers. The system only recommends routes, schedules, and landing sequences; it does not control aircraft. The FAA is scaling this advisory AI from a pilot in Washington to the entire national air traffic system (22 centers, $875 million budget) amid controller shortages and nationwide flight suspensions.

Read briefRead longformTelegram