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

AI explainers with source trails

Longform AI explainers with context, source trails, and related stories.

Public story index

EN · 450 records

Showing the latest 40 of 450 published records. Detail pages remain available through sitemap and internal links; date archives are the next production step.

Riemann Hypothesis: 17-Million-Line Lean Proof Verified by Independent Kernels

On November 20, 2026, Levant Alpöge published a proof of the Riemann Hypothesis as a Lean library of 17 million lines of code, and on November 21, Lean FRO and the Mathlib Initiative confirmed the proof's correctness. The library compiles without errors in Lean v4.33.1 and Mathlib v4.33.1, the raw export was accepted by independent kernels nanoda and con-leche, and the statement was cross-checked against the RiemannHypothesis assertion in Mathlib. The proof establishes five theorems, including t

Read longformRead briefTelegram
Inkling 2 Mini model card thumbnail

Thinking Machines Lab releases Inkling 2 Mini — the first open looped transformer with 'depth scaling'

On November 20, 2026, Thinking Machines Lab released Inkling 2 Mini — an open multimodal model (text and image input, text output) under the Apache 2.0 license: 26.4B total parameters, 3.4B active (5.4B effective per token), context up to 1M tokens, checkpoint ~20 GB, runs on 32 GB VRAM/memory (DGX Spark, Apple Silicon). This is the lab's first looped transformer: 24 layers of decoder-only MoE (8 of 128 experts per layer), where the middle 12 layers are run three times with shared weights, and c

Read longformRead briefTelegram
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 Says He Is Ready to Return to Nvidia, but Board Refuses

On November 17, 2026, Nvidia founder Jensen Huang stated in a post on X that he is ready to return as CEO if the board of directors deems it in the best interest of shareholders, employees, the company, and the United States. The board responded the same day, stating that they unanimously accepted Huang's resignation and do not intend to reinstate him; CFO Colette Kress is currently leading the company on an interim basis.

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

Australia has its own large language model: sovereign Agentaus 1 is now publicly available

Canberra-based Trellis Data (founder Michael Gately) has opened access to its sovereign LLM, Agentaus 1, with 220 billion parameters—approximately 50% larger than the company's previous model. According to the company, it scores 84.4% on MMLU-Pro and ranks 43rd globally among foundational models. It operates exclusively on Australian servers, with air-gapped deployment available for businesses; independent verification of the claimed figures is not yet available.

Read longformRead briefTelegram
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.

Three engineers, six days, and the Zeus AI agent: how Normal built a 25-kilogram cargo drone

On November 16, startup Normal published a breakdown of the Project Heavy Lift: over 6 days (October 31 – November 7), a team of three engineers used the Zeus AI agent to design and assemble a cargo drone with a payload capacity of up to 55 pounds (approx. 25 kg). From November 5–7, the drone completed its first flight under FAA Part 107 regulations and delivered cargo. The company's post on X claims the use of the Opus 5.5 model; the model is not explicitly named on the normal.ai website, but t

Read longformRead briefTelegram
Face clusters feature demonstration

PixlStash 2.0: Image library with automatic face clustering and built-in LoRA training

PixlStash 2.0 (v2.0.0 released November 10, 2026; v2.0.1 released November 16) is an open-source desktop application for managing image libraries with integrated LoRA training. It groups faces by similarity to build character datasets, trains models via AI-toolkit on a local GPU, and automatically attaches them to characters. In ComfyUI, the LoRA is accessible via custom nodes.

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

Moonshot AI shifts Kimi models from coding to biology, chemistry, and pharma

Chinese startup Moonshot AI (developer of Kimi models) announced at Web Summit in Lisbon that it is redirecting its models from programming to biology, chemistry, and pharmaceuticals: according to international director Giacomo Ficari, scientific tasks require stronger models than code, finance, or law, while demand for coding is approaching saturation. Beijing-based startup DP Technology (深势科技) has already integrated the Kimi K3.1 Pro model (3 trillion parameters, 1 million token context, relea

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

Diversity Collapse in 4-Step SDXL, SD3, and Wan 2.1: Cause Found in Activations, Fix via LoRA Rank 32 or Training-Free

Researchers from UIUC, fal, and Johns Hopkins demonstrated in preprint arXiv:2611.18550 that distilled 4-step generators do not lose the teacher's overall variance level but compress it into fewer directions—locally in one intermediate block during the first denoising step—causing different seeds to yield nearly identical results. A lightweight LoRA (rank 32) operating only on the first step, or the training-free ReDiverse-TF method, restores diversity to near the full-step teacher level without

Read longformRead briefTelegram
TokenEconomizer: Learning from Experience to Make AI Coding Agents More Efficient

Microsoft Research releases TokenEconomizer: coding agents learn to save tokens based on past repository experience

Microsoft Research, in collaboration with UC Berkeley, Microsoft Security, and Voyage AI, has released the open-source plugin TokenEconomizer — a black-box controller that sits alongside existing coding agents, does not modify the model, and learns from logs of past sessions in the same repository to reduce token consumption and run costs through three mechanisms (micro-skills, tool pruning, and failure-loop intervention).

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

OpenScientist: Open agentic AI system for biomedical research passes peer review in Nature Medicine

The C-BRAIN consortium, led by Washington University in St. Louis and Lawrence Berkeley National Laboratory, published in Nature Medicine a description of OpenScientist — an open-source agentic AI 'co-scientist' system that semi-autonomously performs the full cycle of biomedical research in minutes and underwent rigorous expert validation on real clinical cases.

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

US tightens AI chip export controls to cut off remote GPU access for Chinese labs

On November 16, 2026, the US Department of Commerce issued an interim rule: a license is required for the export, re-export, and transfer of AI chips and servers to data centers if the capacity could be used by companies from the 23 countries in the D:5 group, including China, Russia, Iran, and North Korea. The threshold is data centers exceeding 12 million TPP (approximately ~750 Nvidia H100/H200 chips), indirect access via intermediaries is covered, existing operators are given 90 days, and th

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

Anthropic: Claude Mythos prepared a model for AMD GPU work in three days with almost no human involvement — instead of six to twelve months

At its IPO roadshow (slide dated November 10, 2026), Anthropic reported that the Claude Mythos Preview model ported or wrote nearly 100% of the kernels — low-level functions connecting the model to specific hardware — for the AMD Instinct MI355X accelerator in approximately three days and "almost without human involvement." Historically, such adaptation took 6–12 months of work by teams of dozens of engineers. A caveat in the slide: the three days cover only prompt processing at a useful perform

Read longformRead briefTelegram
product-seo

Open-source skill product-seo teaches Claude Code, Codex, and Cursor coding agents to manage product SEO

The author of the Tips AI Telegram channel (Igor Trunin) released the open-source skill product-seo v0.1.0 (MIT license) for Claude Code, Codex, and Cursor coding agents: it includes a SKILL.md file, SEO 'game' reference guides, and seven Node.js scripts; installation is a single command, it works from Google Search Console CSV exports, and the methodology was tested on the live HypeScribe site with CTR growth from 0.4% to 0.9% and verification of each edit at week 1 and week 3.

Read longformRead briefTelegram
Looki Loom Model announcement cover image

Looki Unveils Loom: Wearable Camera as a Source of Long-Term Memory for AI Assistants

Chinese company Looki, maker of the Looki L1 camera (32 GB, $199), has introduced Loom — a family of multimodal models that converts video and audio recorded throughout the day into structured long-term memory tied to timestamps and original recordings. The system consists of three layers: Loom Sense answers questions about current events without generating text descriptions (F1 88.6, latency 0.51 s), Loom Moment simultaneously marks event boundaries and describes them, distinguishing between wh

Read longformRead briefTelegram
Run tool calls from code with Code Mode

Claude Code introduces experimental Code Mode: tools invoked via a single JavaScript script

Anthropic has added an experimental Code Mode to Claude Code: instead of sequential Bash, Read, and Edit calls, the model generates a single JavaScript script that runs within the session — commands can run in parallel, the result of one call is passed to the next without an additional model turn, and only the necessary fields are returned to the context instead of raw output. The mode appeared in the v2.1.323 release on November 12, 2026, and is enabled by the CLAUDE_CODE_EXPERIMENTAL_CODE_MODE

Read longformRead briefTelegram
Baseten open-sources deepseek-ultrafast engine: nearly 2000 tokens per second on DeepSeek-V4.1-Flash on standard NVIDIA GPUs

Baseten open-sources deepseek-ultrafast engine: nearly 2000 tokens per second on DeepSeek-V4.1-Flash on standard NVIDIA GPUs

Baseten has open-sourced the deepseek-ultrafast inference engine, written in Rust and CUDA, specifically for the DeepSeek-V4.1-Flash model on NVIDIA B200/B300 GPUs (sm_100a and sm_103a). The README claims decoding speeds of 2060 tokens/s on 8x B300 (TP8, 2.71 ms step) and 1740 tokens/s on 4x GB300, prefill of 1 million tokens in 11.7 seconds (85,000 tokens/s), context up to 1M tokens, and KV-cache sharing between servers via Redis and RDMA. GPU kernels were refined by coding agents working with

Read longformRead briefTelegram

Russian teams present three recommender system papers at ACM CIKM '26: LLM encoder fine-tuning, MoE experts in graphs, and honest model evaluation

Teams from T-Bank, MIPT, HSE, and Innopolis University presented three papers in Rome: E²L³AE achieves up to +20.3% recall and +17.2% ranking quality on office supplies, MoE-IGR outperforms PinSAGE in cold-start by up to +34.4% NDCG@20 on new items, and “Split Decisions” introduces a model-selection regret metric, showing that random train/test splits ignoring chronology often select the wrong model. Conference proceedings were published on November 6, 2026, and all papers are open access under

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

Trump preparing list of 'AI national champions' including Nvidia, OpenAI, and Anthropic

Commerce Secretary Howard Lutnick's announcement is expected within the week, likely ahead of the APEC summit in Shenzhen on November 18–19. According to The Wall Street Journal, the list will include Nvidia, Amazon, Microsoft, OpenAI, and Anthropic: the status will grant them expedited export licenses, support in foreign markets, and access to federal funding, with the program's goal being to sell foreign buyers ready-made packages of American chips, cloud, and models.

Read longformRead briefTelegram
Hy4flash

Tencent releases budget Hy4 flash model: 224B parameter MoE at a fraction of the flagship Hy4 price

Tencent released the budget Hy4 flash model on November 16, 2026: a 224B parameter MoE architecture where 12B parameters are activated per token, including 64B parameters for Engram conditional memory—a mechanism that extracts frequent patterns from a table instead of recomputing them. The model is natively multimodal, supports a 1M token context, and generates text at over 150 tokens/second. API prices per 1M tokens: $0.003 for cached input, $0.070 for input, and $0.281 for output, compared to

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

Alibaba Releases Qwen4-Max: 2.4 Trillion Parameters and a 62% Reduction in Per-Task Cost

Alibaba's new flagship is built on the Qwen4 architecture: a 2.4 trillion parameter MoE backbone with 71 billion active parameters per token, plus 328 billion N-gram Embedding parameters, a 1 million token context window, and multimodal input for text, images, and video. Attention is enhanced with Gated DeltaNet (GDN), Qwen Sparse Attention (QSA), and Gated Residual (GR) mechanisms to reduce the cost of handling long contexts. According to Qwen's self-report, the model scored 52.69 on the intern

Read longformRead briefTelegram
Automatar AI guitar hero image

MIT Media Lab researcher embeds NVIDIA GPU in bass guitar: Automatar builds harmony in real-time

MIT Media Lab researcher Antonis Christou (Opera of the Future group) embedded an NVIDIA Jetson Orin NX single-board module into the body of the Automatar bass guitar: the signal from four main strings with pitch detection is converted to MIDI, and through high-voltage operational amplifiers, five sympathetic strings are excited to complete the harmony in real time. Audio processing is handled by a Bela Mini with SuperCollider, while the body made of dark Pale Moon Ebony and mahogany combines tr

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

OpenAI launches science blog with post on lattice definition of the Standard Model built with GPT-6 Astra

First guest post on OpenAI's science blog: Nathaniel Craig, together with the GPT-6 Astra model, constructed a candidate non-perturbative lattice definition of the Standard Model via a refinement strategy—comparing fields on a sequence of increasingly fine lattices; the proof was reviewed by Maarten Golterman, Andreas Kronfeld, and Yigal Shamir, but the author describes it as an answer 'in principle' rather than 'in practice.'

Read longformRead briefTelegram
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 Starship for the first time after orbital flight, but loses Super Heavy booster in same launch

On November 14, 2026, at Starbase in Texas, the Mechazilla tower captured the Starship spacecraft returning from orbit for the first time in history: Ship 42 of the 15th test flight spent about 22 hours in orbit, completed 14 orbits at an altitude of approximately 265 km, deployed 26 Starlink satellites, and was caught by mechanical 'chopsticks,' while the Super Heavy booster did not return — its engines shut down prematurely during descent, and the stage was destroyed over the Gulf of Mexico.

Read longformRead briefTelegram
GettyImages-2262902677

Expensive AI School Alpha Promised to Teach Twice as Fast, but Only 42% of Students Met the Goal

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 and promises to double learning speed with two hours of daily lessons using the TimeBack tutor app instead of traditional teachers, only 42% of students achieved the stated goal in both math and reading; approximately half of K-8 students did not meet the math goal, and one-fifth did not meet the reading goal.

Read longformRead briefTelegram
Operate Claude Managed Agents inside your security perimeter

AI agents learn to pay: OpenAI launches ChatGPT Wallet with Visa, Mastercard, and American Express

On November 13, 2026, OpenAI launched the ChatGPT Wallet payment service, created in partnership with American Express, Mastercard, and Visa. Users link a card once and confirm each purchase via Face ID or fingerprint, while the payment network issues a token instead of card details, which neither the merchant nor the agent sees; card rewards and dispute protection are preserved. The service is currently available to ChatGPT Pro subscribers in the US.

Read longformRead briefTelegram
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 Upgrades Server with Eight Used Nvidia V100s to Run DeepSeek V4.1 Flash in near-NVFP4 Mode

Attorney Nick Mermiges added 1 TB of RAM, an RTX 3090, and a working power supply to his server, eliminating crashes. Now, eight used Nvidia V100s run DeepSeek V4.1 Flash: 5,949 of 15,360 'hot' MoE experts operate in 4-bit NVFP4, while the rest run in 2-bit. Context expanded to 524K tokens, supporting two parallel requests; token match with the reference NVFP4 version reached 95.1%, but generation slowed down.

Read longformRead briefTelegram
Hero image for The Register article on AI-prompted patch flood in Linux

Torvalds admits Linux kernel AI patch flood may have peaked

In the weekly update for Linux 7.4 rc2, Linus Torvalds suggested that the AI-model-driven influx of fixes may have passed its peak: according to eBPF leader Alexei Starovoitov, fixes in this subsystem are now arriving faster than new bug reports, although the rest of the kernel has not yet reached this equilibrium. The 'patchpocalypse' wave was triggered by the Anthropic Mythos vulnerability discovery model presented in April 2026; meanwhile, Microsoft and Oracle continue to close record volumes

Read longformRead briefTelegram
Full Attack Chain on Microsoft 365 Copilot: From Infected Word Document to Azure Sandbox Escape

Full Attack Chain on Microsoft 365 Copilot: From Infected Word Document to Azure Sandbox Escape

Rubrik Zero Labs researcher (Ori) published a recording of their Black Hat USA 2026 talk, 'ChatMate: Remote Prompt Execution on AI Assistants Through Sandbox Escaping': the attack on Microsoft 365 Copilot began with indirect prompt injection via an infected Word document, and the sandbox escape occurred through CVE-2026-32193 (CVSS 8.8) — a path traversal in Azure Kubernetes Service. Microsoft patched the vulnerability via MSRC and paid a $48,000 bounty.

Read longformRead briefTelegram
Xania Monet creator Telisha "Nikki" Jones

AI Music Gets Its Own Awards: How the First AI Music Awards Went Down

On November 13, 2026, the Sonic Intelligence Academy (SIQA) held the first AI Music Awards in Los Angeles, positioning the event as the "AI equivalent" of the Grammy Awards. Awards in 25 categories were selected by AI artists registered in the SIQA Verified Registry; the top honors for AI Artist of the Year and Song of the Year went to the AI persona Xania Monet, created by Telisha "Nikki" Jones.

Read longformRead briefTelegram
EnvOpt: Automated Optimization for Agent Environments

EnvOpt: Automatic Optimization of LLM Agent Environments Cuts Run Costs by Nearly Half

Bespoke Labs, the University of Chicago, and UC Berkeley Sky Computing Lab have introduced EnvOpt (arXiv:2611.07372, November 5, 2026) — a framework that automatically optimizes the "stationary environment" of an LLM agent: file structure, data representations, and available tools, without changing model parameters or the harness. According to the authors, the average cost savings on unseen tests across five benchmarks is 48% while maintaining or increasing the pass rate.

Read longformRead briefTelegram
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 $875 million SMART system is operational in all 22 FAA centers. It analyzes weather, traffic flow, and controller workload to recommend routes, schedules, and landing sequences, keeping flights under human control. In the Washington pilot area, it rerouted 180 flights, saving an average of 10 minutes and 80 gallons of fuel per flight, although the NATCA air traffic controllers' union was not involved in its design or implementation.

Read longformRead briefTelegram

2013 Recon Jet sports glasses reflashed from Android to Linux and turned into a voice AI assistant

The author reflashed 2013 Recon Instruments Jet sports glasses with a TI OMAP4430 chip from Android to Mainline Linux 6.18: boot time is under 4 seconds, with working drivers for the Kopin 428x240 display, cameras, microphones, Wi-Fi, Bluetooth, IMU, and GPS. The microWakeWord model recognizes commands locally, while heavy inference—vision, ASR, LLM, and TTS—runs on a home server with an AMD Ryzen AI Max 395+ via llama-swap; drivers were ported using Claude Code.

Read longformRead briefTelegram