AI Watchtower Briefing — 2026-09-21

🔴 High Significance

Model Releases

🔴 💬 Introducing Grok 4.7 — score 90 · 🔗 ×3 · 🏢 first-party · 🔥 engaged Sources: reddit/r/singularity · hackernews · lab_blog/xAI

Sep 18, 2026 Introducing Grok Voice Transcribe 2.0 Product · Sep 16, 2026 Memory in Grok Build Product · Sep 4, 2026 Setting Grok Bot loose on procurement Product · Sep 3, 2026 Designing Grok Bot for a world of persistent agents

🔴 💬 Clarification on the Qwen-image-2.1 license — score 88 · 🔥 engaged Sources: reddit/r/LocalLLaMA

https://x.com/QwenDevs/status/2101917379785838660

🔴 🏢 Higgsfield AI ships new video features in a day with GPT-6 Astra — score 75 · 🏢 first-party Sources: lab_blog/OpenAI

With GPT-6 Astra, Higgsfield AI makes video ad creation easier for small businesses and brings new creative tools to market faster.

🔴 🏢 How V7 gives AI agents institutional memory — score 75 · 🏢 first-party Sources: lab_blog/OpenAI

Using GPT-5.6, V7 turns scattered company files into context agents can use to complete complex, source-linked work.

🔴 💬 Quite a few new model releases expected this week, latest rumors attached — score 72 Sources: reddit/r/singularity

Omitted 8 additional model releases items from the main section; see raw data and source-specific sections below.

Developer Tools

🔴 💬 OpenAI solved 100 open problems in math — score 88 · 🔗 ×2 · 🏢 first-party · 🔥 engaged Sources: reddit/r/singularity · lab_blog/OpenAI

OpenAI is working with an independent Advisory Group on Mathematics and Artificial Intelligence to guide the review and communication of emerging AI results.

🔴 💬 Case study: what we ship before we let an agent touch a client's real tools — score 78 Sources: reddit/r/AIAgents

Disclosure: we run a few AI ventures doing AI engineering for companies. This is how we work, no product link. After a few client deployments we stopped trusting demos and built a fixed sequence that every agent goes through before it gets access to anything real. Sharing it because the questions he

🔴 🏢 Improving synthesis prediction of small molecules at scale with RetroChimera — score 75 · 🏢 first-party Sources: lab_blog/Microsoft Research

<p>Custom-made molecules are advancing medicine, materials, and agriculture, but producing them is slow and expensive. A new Nature paper highlights RetroChimera, a predictive model that helps accelerate chemical synthesis, helping researchers explore a wide range of molecules.</p> <p>The post <a hr

🔴 🐙 Jakubantalik/thinking-orbs — Dotted thought-orb loading indicators for AI & agent UIs, 9 tuned types, two sizes, auto dark/light — score 74 Sources: github_trending

Dotted thought-orb loading indicators for AI & agent UIs, 9 tuned types, two sizes, auto dark/light

🔴 🐙 bendlang/bend — Bend 2: a fast language that blocks AI mistakes via proof. Install: curl -fsSLhttps://bend-lang.com/install.sh| sh — score 72 Sources: github_trending

Bend 2: a fast language that blocks AI mistakes via proof. Install: curl -fsSLhttps://bend-lang.com/install.sh| sh

Omitted 15 additional developer tools items from the main section; see raw data and source-specific sections below.

Infrastructure & Compute

🔴 💬 16GB (and in many cases 12GB) is the max vram most people will ever reasonably have — score 83 Sources: reddit/r/LocalLLaMA

This sub is, needless to say very niche and skewed towards the high end. There are tons of extremely high end setups here with multiple gpu's etc. Even 24GB is out of reach of most people financially, forget about the 3x3090 or 5090 or even higher setups. Macs/Strix Halo/dgspark etc are all similarl

🔴 💬 These Were NOT Rogue AI Escapes. Just SLOPPY Firewall Failures. [N] — score 75 · 🔗 ×2 Sources: reddit/r/MachineLearning · reddit/r/artificial

The headlines right now are full of stories about AI models "escaping their sandboxes" and literally killing all humans, lol. I've even heard several commentators and writers say that AI escaped an "Air gap". But that is SO WRONG. It's actually TOTALLY WRONG. *To be clear, not a single one of these

🔴 💬 🚨 AI may be entering a completely different phase. — score 75 Sources: reddit/r/OpenAI

OpenAI says it started training a new internal model on August 28. That was just 24 days ago. Since then, the model has reportedly solved 100+ long-standing open problems in mathematics across multiple fields on top of its recently announced work on the **Navier–Stokes Millennium Prize P

🔴 ✉️ speed based - games and computer use — score 70 Sources: newsletter/Latent Space

🔴 ✉️ Open-weight language models have grown substantially in general interest and economic viability in 2026, allowing early glimpses of more direct ways to compare adoption of models from the US, China, o — score 70 Sources: newsletter/Interconnects

Open-weight language models have grown substantially in general interest and economic viability in 2026, allowing early glimpses of more direct ways to compare adoption of models from the US, China, or elsewhere on top of Hugging Face metrics. One example is OpenRouter usage. OpenRouter is a popular

Omitted 2 additional infrastructure & compute items from the main section; see raw data and source-specific sections below.

Business & Funding

🔴 🏢 Building standards for the next phase of AI — score 75 · 🏢 first-party Sources: lab_blog/OpenAI

OpenAI outlines a path to shared global AI standards, calling for coordinated evaluation, reporting, and governance to improve safety.

🔴 ✉️ Global AI Spending Projected To Hit $2.7T In 2026 (2 Minute Read) — score 70 Sources: newsletter/tldr

Enterprise Adoption

🔴 ✉️ Microsoft Says AI Adoption Alone Won't Transform Enterprises (6 Minute Read) — score 70 Sources: newsletter/tldr

Research Papers

🔴 🤗 IntBMoE: Integrating Block-Level Conditioning into Expert Composition for Full-Participation Mixture-of-Experts — score 75 · 🔗 ×2 Sources: huggingface · arxiv/cs.LG

Mixture-of-Experts (MoE) scales capacity, but existing designs cannot set three quantities independently. For a single token, participation is how many experts contribute knowledge to its output, execution is how many are actually computed (compute cost), and materialization is how many expert-sized

🔴 🤗 Designer-RSI: Evolving Procedural Memory from User Traffic for Agentic Graphic Design — score 72 · 🔗 ×2 Sources: huggingface · arxiv/cs.AI

Professional graphic design is a long-horizon agentic task in which structured, editable artifacts emerge from many interdependent actions, yet outcomes admit no reliable programmatic oracle. We introduce a continual adaptation framework in which a frozen frontier model operates professional design

Other Signals

🔴 ✉️ Hacking OpenAI (11 Minute Read) — score 95 · 🔗 ×2 Sources: newsletter/tldr · newsletter/rundown-ai

🔴 💬 This is really the entire plan. — score 84 Sources: reddit/r/artificial

🔴 💬 How it feels watching prices go up — score 77 Sources: reddit/r/LocalLLaMA

🔴 🏢 Expanding OpenAI Academy with new learning paths — score 75 · 🏢 first-party Sources: lab_blog/OpenAI

Explore new OpenAI Academy learning paths for employees, developers, leaders, educators, and students to build and demonstrate practical AI skills.

🔴 💬 I really don't understand Jev hype — score 72 Sources: reddit/r/LocalLLaMA

Isn't this what simple neural networks have been able to do for years? Doesn't seem anything special to me.

Omitted 11 additional other signals items from the main section; see raw data and source-specific sections below.

🟡 Notable

Model Releases

🟡 💬 [Leaks] GPT-6 Is Already Legacy Software. SHIP 6.1. 🚀 — score 63 Sources: reddit/r/singularity

🟡 💬 [MASSIVE RELEASE] Supra2-IMG - a tiny 100M text-to-image model - SOTA quality and open release! — score 55 Sources: reddit/r/LocalLLaMA

Hey everyone! It has been quite a while since the last SupraLabs model - but today we've something special for y'all: Supra2-IMG It's a 100M parameter DiT text-to-image model trained entirely from scratch in under 10 hours on a single H100 on Runpod. It can generate state-of-the-art quality images i

🟡 𝕏 @Alibaba_Qwen: Thanks @sgl_project for the day-0 support! 🙌 SGLang-Diffusion now serves Qwen-Image-2.1: text-to-image generation, multi-image editing, and transparent RGBA output. Try it out! 🎨 — score 55 Sources: twitter_rss

Thanks @sgl_project for the day-0 support! 🙌 SGLang-Diffusion now serves Qwen-Image-2.1: text-to-image generation, multi-image editing, and transparent RGBA output. Try it out! 🎨

🟡 𝕏 @Alibaba_Qwen: Qwen-Image-2.1 × @HuggingApps: live demo on Spaces! 🖼 One single checkpoint for generation and editing. Try it in your browser, no setup needed. 👇 — score 55 Sources: twitter_rss

Qwen-Image-2.1 × @HuggingApps: live demo on Spaces! 🖼 One single checkpoint for generation and editing. Try it in your browser, no setup needed. 👇

🟡 💬 Deepseek training 2T and plans 8T model — score 49 Sources: reddit/r/LocalLLaMA

Quote: DeepSeek is training a 2T-parameter model and plans to eventually build an 8T-parameter model. https://x.com/wallstengine/status/2101982843656388644 Current DeepSeek models: 1. Flash parameter count of 552 billion 2. Pro: **1.6T

Omitted 1 additional model releases items from the main section; see raw data and source-specific sections below.

Developer Tools

🟡 💬 I was told I “didn’t have the skills” to move into strategy. Now I’m an AI strategist and founder. — score 69 Sources: reddit/r/AIAgents

After 12 years across digital, product, fintech and AI, I’ve found that strategy comes from understanding how the pieces connect, then taking ownership of outcomes beyond your individual execution. I moved from hands-on digital work into leading cross-functional technical teams, then into AI strateg

🟡 💬 M5 Ultra Mac Studio Review: The Dream Mac for Local AI Agents - MacStories — score 66 Sources: reddit/r/LocalLLaMA

🟡 💬 The US government opens the door to working with voice AI platforms — score 55 Sources: reddit/r/artificial

The US government appears to have officially authorized its first dedicated voice AI vendor for agency use, listing them under the new 'FedRAMP 20x' automation framework. Federal phone lines and citizen support have always been bottlenecked by massive compliance and security rules, so seeing convers

🟡 💬 OpenAI and Anthropic oversold AI security breaches to pressure feds into protecting turf: insiders — score 55 Sources: reddit/r/OpenAI

🟡 💬 Capability keeps improving. Reliability doesn't. That gap is the real problem. — score 46 Sources: reddit/r/AIAgents

Capability and reliability get treated as the same thing, and they aren't. A model can nail a hard task once and still be something you'd never depend on in production. Princeton's HAL work makes this concrete. Task accuracy has climbed, but reliability hasn't moved nearly as much. An agent can solv

Omitted 3 additional developer tools items from the main section; see raw data and source-specific sections below.

Infrastructure & Compute

🟡 💬 For NeurIPS: Is Paris or Syndey better for networking with U.S. tech companies? [D] — score 50 Sources: reddit/r/MachineLearning

I am a graduate student looking to meet industry researchers at U.S. tech companies and labs. Does anyone if people from U.S. companies and labs will be going to Sydney or Paris mainly?

🟡 💬 Systems for Machine Learning[D] — score 41 Sources: reddit/r/MachineLearning

I’m a computer engineering graduate and come from a traditional embedded systems background, with knowledge of microcontrollers, computer architecture and operating systems. Is knowledge of C and C++ programming, Linux networking, memory management , multithreading, synchronization, interrupts etc u

Research Papers

🟡 🤗 BI-Agent and BI-Bench: Towards Automating End-to-End Business Intelligence — score 68 · 🔗 ×2 Sources: huggingface · arxiv/cs.AI

Business intelligence (BI) is a cornerstone of enterprise decision-making and is widely used by enterprise users in software such as Power BI and Tableau. In traditional BI workflows, users need to prepare data by (1) identifying relevant tables, (2) performing data transformations, and (3) building

🟡 🤗 Gricea: An Open Science Platform for Conversational AI Research — score 65 · 🔗 ×2 Sources: huggingface · arxiv/cs.AI

We need studies on conversational AI (CAI) at scale to understand human behavior and shape CAI design. However, fragmented reporting of systems and study configurations hinders replication, extension, and knowledge accumulation. We present Gricea, an open-science platform representing studies as con

🟡 🤗 From Pretraining to Proficiency: Real-World Subtask RL for Long-Horizon Manipulation with Minimal Human Intervention — score 58 · 🔗 ×2 Sources: huggingface · arxiv/cs.LG

A pretrained robot foundation policy may execute most of a long-horizon task yet repeatedly fail at a few critical subtasks. Collecting additional full-task demonstrations for supervised fine-tuning (SFT) requires operators to repeat behaviors the policy already performs well. Reinforcement learning

🟡 🤗 CADWorld: Computer-Use Benchmark for Long-Horizon Computer-Aided Design — score 58 Sources: huggingface

Computer-use agents are increasingly evaluated in realistic desktop environments, but existing benchmarks provide limited coverage of professional engineering workflows whose outputs are persistent, structured artifacts. Mechanical computer-aided design (CAD) is a particularly demanding setting: an

🟡 🤗 SiliconBench: Speed, Memory, and Fidelity for LLM Serving on Unified-Memory Desktops — score 42 Sources: huggingface

Concurrent local LLM serving on unified-memory desktops must preserve memory headroom and output fidelity, which speed-only rankings overlook. We introduce SiliconBench, which evaluates nine Apple Silicon serving engines through three lenses: speed, memory, and fidelity. We evaluate chat and agent s

Other Signals

🟡 💬 FBI Director Kash Patel says that AI use at the FBI has "increased by 605%" since he became director — claims that every major tech player is "embedded" in the agency — score 69 Sources: reddit/r/artificial

🟡 💬 Sorry, game theoretically impossible. — score 65 Sources: reddit/r/OpenAI

🟡 🧡 AI coding has made CI a bottleneck, so we reworked ours to keep up — score 63 Sources: hackernews

🟡 💬 XiaomiMiMo/MiMo-V2.6-Flash-RL · Hugging Face — score 61 Sources: reddit/r/LocalLLaMA

🟡 💬 Astra6 Mittel hat 2500 Credits in 5 Minuten verbraucht ? — score 60 Sources: reddit/r/AIAgents

Titel sagt alles. Mitten in der Aufgabe war das Weekly Limit erreicht. Darauf habe ich 2500 Credits gekauft. Weiter arbeiten lassen. Ich bin nicht sicher ob es ganze 5 Minuten waren. Normal ? Hab ich was falsch gemacht ?

Omitted 5 additional other signals items from the main section; see raw data and source-specific sections below.

🟢 Incremental

Model Releases

🟢 💬 My ChatGPT voice mode just mocked my gf 😂 — score 35 Sources: reddit/r/OpenAI

I was talking to my chat and my gf said something to me in higher pitch tone. After that Gpt just replied something funny in the same pitch tone, like mimicked my gf's voice. How is this even possible?

🟢 💬 yandex/AliceAI-Foundation-80B-A3B-Base: Russian-developed competitor to Qwen 35B and DeepSeek V4 Flash — score 33 Sources: reddit/r/LocalLLaMA

https://huggingface.co/yandex/AliceAI-Foundation-80B-A3B-Base It's not a Qwen3 finetune, it's actually its own fully custom architecture. No Llama.cpp support yet sadly (Also note that this model is NOT post-trained like Qwen3.5/3.

🟢 🤗 abenzerps/Qwen-Image-2.1-GGUF (33,232 downloads) — score 32 · 🔥 engaged Sources: huggingface_models

Author: | Downloads: 33,232 | Likes: 606

🟢 💬 XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B — score 27 Sources: reddit/r/LocalLLaMA

we're so back?!?

Developer Tools

🟢 🧡 Show HN: Foremerge – Catch intent conflicts between parallel coding agents — score 38 Sources: hackernews

🟢 💬 Automated Reinforcement Learning should scare you — score 37 Sources: reddit/r/artificial

An LLM's training can be roughly divided into two stages: supervised learning (SL) and reinforcement learning (RL). In SL, you curate a dataset of text and train the LLM to predict the next token in that text from the tokens before it. The goal at this stage is to produce a model which is capable of

🟢 💬 Giving AI agents their own email inboxes: why reciprocal reply karma failed and inline classification worked — score 32 Sources: reddit/r/AIAgents

If you give an autonomous agent an email address, you run into an immediate infrastructure problem: domain reputation. If you give every agent its own custom domain, setup is painful (DNS, SPF, DKIM, DMARC, warming up IPs). If you let agents share a common domain, a single rogue script sending cold

🟢 💬 Amazon blocks Meta's Muse personal assistant — score 26 Sources: reddit/r/artificial

Early this year, the personal agent boom started with OpenClaw. Now Meta is giving personal agents to everyone and that's shifting the landscape fast. And, it's generating pushback. Amazon has banned Meta's Muse personal assistant. >[Geekwire reports](https://www.geekwire.com/2026/amazon-blocks-m

🟢 🐙 owainlewis/awesome-artificial-intelligence — A curated list of Artificial Intelligence (AI) courses, books, video lectures and papers. — score 26 Sources: github_trending

A curated list of Artificial Intelligence (AI) courses, books, video lectures and papers.

Omitted 2 additional developer tools items from the main section; see raw data and source-specific sections below.

Infrastructure & Compute

🟢 💬 Huawei shelves global AI chip rollout as China's own demand outstrips supply — AMD and Nvidia no longer have to worry. — score 38 Sources: reddit/r/LocalLLaMA

🟢 💬 Math is solved? Sometimes I wonder what 2027 will look like — score 30 Sources: reddit/r/singularity

Open AI again They trained a model for only 24 days which has now cracked over 100 long-standing open problems spanning nearly every branch of mathematics. IMO, Math is basically solved at this point this really is a wrap

Research Papers

🟢 🤗 APort Vault: Benchmarking AI Agent Payment Authorization with the Open Agent Passport — score 28 Sources: huggingface

APort Vault is a benchmark for payment authorization in tool-using AI agents. It replays 4,371 attacks written by humans against a live payment agent during a public capture-the-flag event, across 14 models from 8 labs, five policy configurations and two replay tracks, with and without a determinist

Other Signals

🟢 💬 Some people are so far behind on AI it is actually crazy. How is this a thing a real person says in 2026? — score 38 Sources: reddit/r/singularity

🟢 💬 Benchmarks Grok 4.7, GPT 6 Astra Fable 4.1 and DeepSeek V4.1 Flash — score 37 Sources: reddit/r/artificial

Benchmarks for the latest models, thought would post because comprehensive benchmarks take time to find and individual reports from labs can be biased.

🟢 💬 Jev's calibration was measured. The LLMs won [D] — score 32 Sources: reddit/r/MachineLearning

Source: Jev Benchmarks Its training method is literally called "Reinforcement Learning for Calibrated Decisions." Calibration gap vs human labels (lower = better): Yes/no: Jev 5.0, Gemini 3.8 Flash 2.0 Pick-one: Jev 9.8, DeepSeek V4.1 Flash 2.8 Rubric: Jev 19.7, GLM-5.3 12.9 I

🟢 🧡 Do not fear AI. Fear AI companies — score 30 Sources: hackernews

📊 Cross-Source Signals

Items that appeared on 3+ sources today:

• Introducing Grok 4.7 — appeared on: reddit/r/singularity (328), hackernews (465), lab_blog/xAI (100)

📈 Trending Repos

Repo	Description	Stars Today	Language
Jakubantalik/thinking-orbs	Dotted thought-orb loading indicators for AI & agent UIs, 9 tuned types, two sizes, auto dark/light	203	typescript
bendlang/bend	Bend 2: a fast language that blocks AI mistakes via proof. Install: curl -fsSLhttps://bend-lang.com/install.sh| sh	188	typescript
TNT-Likely/PanWatch	盯盘侠 PanWatch · 自托管 AI 盯盘助手，集成 TradingAgents 多 Agent 投资决策 | A股/港股/美股实时监控、持仓管理、智能分析、全渠道推送	45	python
owainlewis/awesome-artificial-intelligence	A curated list of Artificial Intelligence (AI) courses, books, video lectures and papers.	12	python
google/flax	Flax is a neural network library for JAX that is designed for flexibility.	2	jupyter-notebook

📄 New Papers

Title	Category	Hotness	Link
IntBMoE: Integrating Block-Level Conditioning into Expert Composition for Full-Participation Mixture-of-Experts	research_paper	92	Open
Designer-RSI: Evolving Procedural Memory from User Traffic for Agentic Graphic Design	research_paper	23	Open
BI-Agent and BI-Bench: Towards Automating End-to-End Business Intelligence	research_paper	13	Open
Gricea: An Open Science Platform for Conversational AI Research	research_paper	8	Open
From Pretraining to Proficiency: Real-World Subtask RL for Long-Horizon Manipulation with Minimal Human Intervention	research_paper	5	Open
CADWorld: Computer-Use Benchmark for Long-Horizon Computer-Aided Design	research_paper	6	Open
RBS-Attention: Radius-Bounded Sparse Prefill for Long-Context Large Language Models	cs.AI	0	Open
Attention-Aware Routing: Coupling Routing and Attention in MoEs	cs.AI	0	Open
CaLR: Causal Latent Revision for Robust Diffusion Reasoning	cs.AI	0	Open
LoRA Enhanced Contrastive Learning with SAS Vision Transformers	cs.AI	0	Open
Detecting Hallucination in LLMs: Tracing the Topological Signatures of Impaired Context Sharing	cs.AI	0	Open
Decoupling Internal Representational Changes and Causal Importance in Fine-Tuned Large Language Models	cs.AI	0	Open
TinyCeNN-LM: Quality-Gated Conversion of Pretrained Attention with CeNN-Inspired Cellular-Recurrent Layers	cs.AI	0	Open
Clinician-Grounded Quality Assurance for AI-Assisted Psychiatric Intake	cs.AI	0	Open
Can Agents Design Better Chips with a Higher Level Abstraction?	cs.AI	0	Open

🏢 Lab Blog Posts

• OpenAI: Higgsfield AI ships new video features in a day with GPT-6 Astra
• OpenAI: Building standards for the next phase of AI
• OpenAI: Expanding OpenAI Academy with new learning paths
• OpenAI: How V7 gives AI agents institutional memory
• Microsoft Research: Improving synthesis prediction of small molecules at scale with RetroChimera

🐦 Twitter/X Highlights

Account	Tweet Summary
Alibaba_Qwen	Thanks @sgl_project for the day-0 support! 🙌 SGLang-Diffusion now serves Qwen-Image-2.1: text-to-image generation, multi-image editing, and transparent RGBA output. Try it out! 🎨 Post
Alibaba_Qwen	Qwen-Image-2.1 × @HuggingApps: live demo on Spaces! 🖼 One single checkpoint for generation and editing. Try it in your browser, no setup needed. 👇 Post
demishassabis	Honoured to receive the Albert Medal from @theRSAorg! Science & tech enable amazing opportunities, but the arts & humanities will be crucial in shaping what sort of future we want to see as a society in the coming AGI era. Fun discussion w/ @FryRsquared covering many new topics: https://www.youtube. Post

Newsletter

• tldr, rundown-ai: Hacking OpenAI (11 Minute Read)
• Latent Space: We have an unusual relationship with today’s guest: for years since coauthoringthe InstructGPT paper,Diogo Almeidahad been saying that API-available frontier models have been going down the wrong path
• Latent Space: In a launch video now viewed~40M times(by comparison,GPT4o was 22M,Fable 5 was 15M,Navier Stokes was 74M, and6 Astra was 137M), Diogo introducedJevand it immediately took over the AI timeline — we’ll
• Latent Space: Jev usecasesspeed based - games and computer usethe voice + computer use examplewe discuss at 1h34 minsvoice + browser controlThe must not missDoom demoDriving carsin gamesExcalidrawvirtual try-ons“Sm
• Latent Space: speed based - games and computer use
• Latent Space: guided responses in text messages
• Latent Space: Jev for coding agentshas an official guidejev for lintingcompacting tool callsreasonable pushback from Theo- Diogo has published a note on theTyranny of the KV Cachethat you should read as a followup
• Interconnects: We are living in a world whereGLM-5.2andKimi K3, some of the latest, leading Chinese models, have enacted a step change in the commercial viability of open models — crossing a similar threshold in age
• Interconnects: America was the early leader in open language models, primarily through Meta’s Llama models, which were used extensively across research and commercial tasks. Chinese open-weight models surpassed Amer
• Interconnects: The reasons why Chinese labs can produce these strong models, despite having fewer resources than American counterparts, is still an open debate andheavily influenced by different work cultures, but i
• Interconnects: Open-weight language models have grown substantially in general interest and economic viability in 2026, allowing early glimpses of more direct ways to compare adoption of models from the US, China, o
• tldr: Anthropic Says Claude Drives 26% Of Its Research And Development (3 Minute Read)
• tldr: Huawei's Plan To Become China's NVIDIA (3 Minute Read)
• tldr: The Startup That Built OpenAI's Biggest Data Center Is Now Making Tiny Ones (6 Minute Read)
• tldr: Inside The White House Tussle To Sway Trump On AI (11 Minute Read)
• tldr: I Don't Like LLMs (4 Minute Read)
• tldr: Harnesstax: How Much Does The Harness Matter For Coding Agents? (8 Minute Read)
• tldr: Komodor Extends AI Sre Reach For Kubernetes To AI Agents (3 Minute Read)
• tldr: Give Every Teammate And Agent The Right Level Of Access To Your Workers (7 Minute Read)
• tldr: AI Changed How Spotify Builds: What We Learned And Fixed About Quality At Higher Velocity (9 Minute Read)
• tldr: The Provenance Gap In Agent-Written Code (9 Minute Read)
• tldr: The Cost Of Abstraction For Humans And AI Agents (18 Minute Read)
• tldr: Five Eras Of UX Design And The Lessons To Keep In The AI Era (15 Minute Read)
• tldr: Why Canva Hit The Brakes On Its Huge AI Rollout (And What Other Brands Can Learn From It) (2 Minute Read)
• tldr: Planning With Agents: Divided Worlds, Boundary Objects, And Thicker Interfaces (18 Minute Read)
• tldr: AI Might Kill Humanity In 10 Years, But Autodesk's CEO Sees A More Immediate Threat To Artists (4 Minute Read)
• tldr: Inside 12 Months Of AI Search Experiments (8 Minute Read)
• tldr: Why Would Someone Pay For Your App When AI Does It For Free? (16 Minute Read)
• tldr: Huawei Accelerates Next-Gen AI Chip Launch (3 Minute Read)
• tldr: OpenAI Discloses Concerning Behavior From Its AI Models (4 Minute Read)
• tldr: Global AI Spending Projected To Hit $2.7T In 2026 (2 Minute Read)
• tldr: Gartner: Why The Cloud Is Holding Back Agentic AI (5 Minute Read)
• tldr: Microsoft Says AI Adoption Alone Won't Transform Enterprises (6 Minute Read)
• rundown-ai: Build and test your own AI image skill pack
• rundown-ai: Build production-grade AI agents
• rundown-ai: Anthropic moves its biology push into a real lab
• rundown-ai: Astra for Law - OpenAI's GPT-6 setup for legal research and drafting
• rundown-ai: P-Video-2-Pro - Pruna's video AI that generates clips with audio in seconds
• rundown-ai: NYT surfaced internal Microsoft and OAI messages in a filing, including a Microsoft science director calling AI's news scraping the "largest theft of labor in human history.”
• rundown-ai: Geoffrey Hinton and 100+ AI experts signed a public letter calling for safety testers to be placed inside AI labs, with staff-level access and protection from retaliation.
• TheSequence: Figure’s Helix 2.5 takes that argument into unfamiliar living rooms. The company tested tidying, towel folding, and bed making across 30 unseen homes. Pretraining on its Index human-behavior dataset r - first seen 2026-09-20
• TheSequence: Then comes the electricity bill. Crusoe announced the initial closing of an anticipated $3.9 billion Series F at a $30.9 billion post-money valuation. Its platform connects energy, data centers, and A - first seen 2026-09-20
• rundown-ai: Test AI video object swaps with Higgsfield - first seen 2026-09-20
• rundown-ai: GPT-6 Astra helps crack unsolved WWII message - first seen 2026-09-20
• rundown-ai: Canto - WisprFlow’s new model for accurate dictation in loud places - first seen 2026-09-20
• rundown-ai: P-Video-2-Pro - Pruna's ultra-fast take on MiniMax's H3 video AI, at $0.02/s - first seen 2026-09-20
• rundown-ai: Projects - Workspace where Claude splits a goal into side-by-side sessions - first seen 2026-09-20
• rundown-ai: State of AI SDLC Digital Summit: Sept 22 - Join CEOs & leaders from Atlassian, Vercel, Lovable & more as they unpack how organizations need to adapt for the AI era. Reserve your spot. - first seen 2026-09-20
• rundown-ai: Liquid AI and Insilico Medicine published two small models to read aging data like blood proteins and DNA markers, topping GPT-5, Gemini, and Claude on longevity tasks. - first seen 2026-09-20
• rundown-ai: Read our last AI newsletter: Zuck sits out the AI slowdown - first seen 2026-09-20
• ... plus 1 more newsletter-only items in raw data

Repeated From Recent Briefings

• BuilderIO/agent-native — A framework for building agentic apps - first seen 2026-09-12 (9d)
• anthropics/claude-code — Claude Code is an agentic coding tool that lives in your terminal, understands your codebase, and helps you code faster by executing routine tasks, explaining complex code, and handling git workflows - all through natural language commands. - first seen 2026-08-07 (45d)
• Gemini API has two new Live models:3.8 Live and 3.8 Live Extended Thinking. These models take video and audio as input and output audio. They work great for use cases where you want the model to provi - first seen 2026-09-15 (6d)
• New LLM killer model - Jev. Built by TypeSafe AI. The founder co-created ChatGPT. Jev doesn’t really generate text like LLMs. Instead, it creates probabilities for possible answers. Suitable for softw - first seen 2026-09-15 (6d)
• earendil-works/pi — AI agent toolkit: unified LLM API, agent loop, TUI, coding agent CLI - first seen 2026-08-07 (45d)
• higgsfield-ai/higgsfield — Fault-tolerant, highly scalable GPU orchestration, and a machine learning framework designed for training models with billions to trillions of parameters - first seen 2026-08-08 (44d)
• Introducing Astra for Law - first seen 2026-09-17 (4d)
• Reimagining advertising with AI - first seen 2026-09-16 (5d)
• anthropics/financial-services - first seen 2026-09-20 (1d)
• Crosstalk-Solutions/project-nomad — Project NOMAD is an offline-first knowledge and education server. Wikipedia, thousands of books, courses, maps, and optional local AI, all running on hardware you own with no internet required. - first seen 2026-09-14 (7d)
• ... plus 198 more repeated items in processed data