AI Watchtower Briefing — 2026-07-28

🔴 High Significance

Model Releases

🔴 💬 GPT-5, the world best model just 1 year ago, is today inferior to Qwen3.6 27B and most today’s low-tier models — score 93 Sources: reddit/r/singularity

🔴 🧡 Kimi K3 Architecture Overview and Notes — score 93 Sources: hackernews

🔴 💬 Gemini Distillation Service — score 83 Sources: reddit/r/LocalLLaMA

So looks like Google is now going to offer distilling as a service.

🔴 💬 NeurIPS 2026 Reviewer: AI-Generated Rebuttals (and Paper) [D] — score 81 Sources: reddit/r/MachineLearning

One of the papers I reviewed has what seems to be entirely LLM-generated rebuttals, and the original paper is also clearly LLM-generated, with Claude-speak everywhere. While the authors acknowledge LLM writing assistance in the checklist, it really annoys me personally - what is clearly Claude's wri

🔴 💬 Should we be calling Elon a liar? — score 77 Sources: reddit/r/LocalLLaMA

Last year he said grok 3 would be open sourced in about 6 months. A year later and nada. https://x.com/elonmusk/status/1959379349322313920

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

Developer Tools

🔴 🐙 agentscope-ai/QwenPaw — Your Personal AI Assistant; easy to install, deploy on your own machine or on the cloud; supports multiple chat apps with easily extensible capabilities. — score 97 Sources: github_trending

Your Personal AI Assistant; easy to install, deploy on your own machine or on the cloud; supports multiple chat apps with easily extensible capabilities.

🔴 💬 Are single GPU research still published in ML/DL and its applications nowadays? Which are the most notable recent ones? [D] — score 94 Sources: reddit/r/MachineLearning

ML research is progressing at breakneck speed where frontier labs in both academia and industry have access to considerably large computes (GPUs). Where do small labs or independent researchers go in this context? Have you come across recent works in ML/DL and its applications (vision, language, spe

🔴 💬 AI-coding agents kill team collaboration, according to an analysis of 25,264 agent-generated PRs across 2,361 popular GitHub repositories. — score 94 Sources: reddit/r/AIAgents

🔴 🐙 ogulcancelik/herdr — agent multiplexer that lives in your terminal. — score 91 Sources: github_trending

agent multiplexer that lives in your terminal.

🔴 🐙 t8y2/dbx — 20 MB lightweight cross-platform database client for 70+ databases, including MySQL, PostgreSQL, SQLite, Redis, MongoDB, DuckDB, SQL Server, and Dameng. Built-in AI, MCP Server, CLI, desktop and Docker. | 轻量级跨平台数据库管理工具，支持 MySQL、PostgreSQL、SQLite、Redis、MongoDB、达梦等 70+ 数据库，提供桌面端、Docker、CLI、内置 AI 助手和 MCP Server。 — score 84 Sources: github_trending

20 MB lightweight cross-platform database client for 70+ databases, including MySQL, PostgreSQL, SQLite, Redis, MongoDB, DuckDB, SQL Server, and Dameng. Built-in AI, MCP Server, CLI, desktop and Docker. | 轻量级跨平台数据库管理工具，支持 MySQL、PostgreSQL、SQLite、Redis、MongoDB、达梦等 70+ 数据库，提供桌面端、Docker、CLI、内置 AI 助手和 M

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

Business & Funding

🔴 💬 Everything I've had break in the last year broke at the navigation layer, not the logic — score 78 Sources: reddit/r/AIAgents

Went back through my failures for the year and almost none were logic bugs. Every one was a page changing under a working script. Class name moves, div gets renamed, selector matches nothing, script carries on returning empty. Found out three days later. The ones that survived are where I skipped th

Research Papers

🔴 🤗 Sol-Attn: Accelerating Video Generation Inference via On-the-Fly Attention Sparsification — score 82 Sources: huggingface · arxiv/cs.CV

Diffusion transformers are essential for high-fidelity video generation, but long token sequences make attention a dominant inference bottleneck. Training-free dynamic sparse attention alleviates this bottleneck by computing only selected key-value blocks, yet existing methods struggle to sparsify a

🔴 🤗 Chamaileon: Cross-Context Binder Design with Contextualized Modeling and Mixed Sampling — score 78 Sources: huggingface · arxiv/cs.LG

The rapid evolution of generative models has unlocked new potentials in protein binder design, a pivotal task in structural biology, by facilitating end-to-end generation via joint sequence-structure modeling or hallucination. However, existing approaches are predominantly implemented under a single

Other Signals

🔴 💬 Anthropic is calling for a ban on open-weights models by proposing mandatory requirements they will probably never be able to meet — score 97 Sources: reddit/r/LocalLLaMA

🔴 💬 Return of the bicameral mind. — score 94 Sources: reddit/r/OpenAI

Bro discovered thinking.

🔴 💬 Sorry, but did Dario just say that closed-weights, in-secret models are worse than open-weights ones? — score 90 Sources: reddit/r/LocalLLaMA

🔴 💬 Elon completely contradicts himself at the end of his disastrous interview with The Economist — score 79 Sources: reddit/r/singularity

🔴 💬 DeepSeek V4 Flash, up to 32 tok/s on AMD Ryzen AI MAX+ 395 — score 70 Sources: reddit/r/LocalLLaMA

Hey fellow llamas. we have something new for Strix Halo owners we thought would be useful to share. i'll keep it short: We were able to fit DeepSeek V4 Flash plus its speculative draft on a single Ryzen AI MAX+ 395 with 128 GB of unified memory, and got it to a usable decode rate. Blog post with all

🟡 Notable

Model Releases

🟡 💬 Tibo is getting a limit reset, so apparently I have to crawl out of bed and work again — score 69 Sources: reddit/r/OpenAI

Tibo says ChatGPT Work is taking off so fast that he now feels like a limit reset. Great. I had already accepted that today was over, closed the laptop, and got into bed. Now I’m crawling back out because Ultra and /fast are apparently about to get another life. Thanks, Tibo. My sleep schedule has

🟡 ✉️ Also, I wouldn’t blame you if you missed it, butClaude also has a voice mode. Till now, it could only use Haiku, but now it can use Sonnet and Opus as well as call multiple tools like Gmail, Calendar — score 65 Sources: newsletter/Ben's Bites

Also, I wouldn’t blame you if you missed it, butClaude also has a voice mode. Till now, it could only use Haiku, but now it can use Sonnet and Opus as well as call multiple tools like Gmail, Calendar and Slack mid-conversation. I don’t remember when I last used Claude’s chat option. It’s been Claude

🟡 ✉️ A key trend we have beentracking over at AINews is the absolute explosion in Codex usage this year, withMAU nowup >10x from Jan 2026. Less than two weeks after theirJuly 9th launch, OpenAI said ChatGP — score 65 Sources: newsletter/Latent Space

A key trend we have beentracking over at AINews is the absolute explosion in Codex usage this year, withMAU nowup >10x from Jan 2026. Less than two weeks after theirJuly 9th launch, OpenAI said ChatGPT Work and Codex had reached10M million userscombined (as we cover in the pod, Codex now powers Chat

🟡 ✉️ We’ve been calling out howcoding agents are “breaking containment” to do everything elsethis year to power every other part of knowledge work - and it started with the org chart, with amajor reorg las — score 65 Sources: newsletter/Latent Space

We’ve been calling out howcoding agents are “breaking containment” to do everything elsethis year to power every other part of knowledge work - and it started with the org chart, with amajor reorg last monththat amounted to two of Codex’s most prominent leaders, Greg and Tibo, taking responsibility

🟡 ✉️ However, knowledge work has a different set of problems and environments than coding. For decades, knowledge work has been scattered across different primitives like documents for writing, spreadsheet — score 65 Sources: newsletter/Latent Space

However, knowledge work has a different set of problems and environments than coding. For decades, knowledge work has been scattered across different primitives like documents for writing, spreadsheets for analysis, slide decks for communication, and specialized applications for everything else.Chat

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

Developer Tools

🟡 🐙 HKUDS/OpenSpace — "OpenSpace: The Skill Management Layer for AI Agents" --https://open-space.cloud/ — score 67 Sources: github_trending

"OpenSpace: The Skill Management Layer for AI Agents" --https://open-space.cloud/

🟡 ✉️ I’m back from holiday and straight back tomessing aroundbuilding. I’ve been kinda obsessed and enamoured with the demos from tldraw. It’s a free canvas/whiteboard app but their offline app (recentlyup — score 65 Sources: newsletter/Ben's Bites

I’m back from holiday and straight back tomessing aroundbuilding. I’ve been kinda obsessed and enamoured with the demos from tldraw. It’s a free canvas/whiteboard app but their offline app (recentlyupdated) is cool because agents can do all sorts with it; create visualisations to explain stuff, make

🟡 ✉️ Like turning our logo into a little animating mascot. He’s called ‘bites’ btw. — score 65 Sources: newsletter/Ben's Bites

I’m not good using agents when everything is just text in a file. It’s boring, I don’t read it and I sure as shit don’t do enough (anything?) with it. So if it’s visual and all on one canvas, maybe I will?

🟡 ✉️ This startup runs on17 AI agents- and open-sourced the platform.Marketing, client health, competitive intel, finance - real ops, not demos, with every agent action git-tracked and a human approving an — score 65 Sources: newsletter/Ben's Bites

This startup runs on17 AI agents- and open-sourced the platform.Marketing, client health, competitive intel, finance - real ops, not demos, with every agent action git-tracked and a human approving anything external. Agents you own, not rent. Apache 2.0:github.com/Abilityai/trinity

🟡 ✉️ For serious work, I still prefer using a speech to text app (currently using one I built —optionafk.com) and sending that to the agent of my choice with full control over model/thinking effort. — score 65 Sources: newsletter/Ben's Bites

Omitted 29 additional developer tools items from the main section; see raw data and source-specific sections below.

Infrastructure & Compute

🟡 💬 NeurIPS 2026 AI-generated reviews [D] — score 69 Sources: reddit/r/MachineLearning

I'm really confused about what the point of the prompt injection was (speaking as an author). Is it just a study? I would really prefer that they took action against the AI-generated reviews. Obviously, we cannot assume that the reviewers were copy-pasting the output from the LLM without having give

🟡 💬 Will AI literacy become a basic workplace skill? — score 65 Sources: reddit/r/artificial

A few years ago, knowing how to use a computer was a big advantage. Today, it’s expected. I feel AI might follow a similar path. Knowing how to use AI tools effectively could become a basic skill across many jobs. Not everyone needs to build AI models but understanding how to use them, verify output

🟡 ✉️ NVIDIA IN TALKS WITH OPENAI TO GUARANTEE $250 BILLION FINANCING FOR DATA CENTER (5 MINUTE READ) — score 65 Sources: newsletter/tldr

🟡 🐙 NanmiCoder/cc-haha — 本地优先的跨平台 Claude Code / Agent 桌面工作台：多 Agent、Git Worktree、代码 Diff、技能市场、多模型、Computer Use、任务感知桌面宠物，并支持微信、飞书、钉钉、Telegram、WhatsApp 与 H5 访问。 — score 53 Sources: github_trending

本地优先的跨平台 Claude Code / Agent 桌面工作台：多 Agent、Git Worktree、代码 Diff、技能市场、多模型、Computer Use、任务感知桌面宠物，并支持微信、飞书、钉钉、Telegram、WhatsApp 与 H5 访问。

🟡 🐙 lightseekorg/tokenspeed — TokenSpeed is a speed-of-light LLM inference engine. — score 51 Sources: github_trending

TokenSpeed is a speed-of-light LLM inference engine.

Research Papers

🟡 🤗 UltraViT: Latency-Optimized On-device Vision Encoder for Large Vision-Language Models — score 62 Sources: huggingface · arxiv/cs.CV

Large Vision-Language Models (LVLMs) remain bottlenecked by massive computational footprints, precluding their deployment on resource-constrained edge devices. While efforts to compress LVLMs focus heavily on vision token reduction or smaller language models, the vision encoder is largely overlooked

🟡 🤗 A Vocabulary for Multi-Agent Automated Research Systems — score 62 Sources: huggingface · arxiv/cs.LG

We introduce a vocabulary for automated research systems built from one or more agents to make their design choices easier to describe and compare. The vocabulary specifies 1) who the agents are, 2) what operations are available in the system, 3) who may invoke them, 4) how agents communicate, 5) wh

🟡 📄 Memory Efficient Audio Synthesis with Decoupled Temporal Depth Diffusion Transformers — score 60 Sources: arxiv/cs.CL · lab_blog/Apple ML

arXiv:2607.23811v1 Announce Type: cross Abstract: Siri Expressive Voices synthesize rich, configurable speech in real time and entirely on device, powered by AFM 3 Core Advanced, Apple's most powerful on-device foundation model. This work presents the memory-efficient audio synthesis architecture be

🟡 🤗 TRACE: Business Rule-Grounded Reasoning Curriculum for Knowledge-Preserving Parametric Tool Retrieval in Enterprise LLMs — score 55 Sources: huggingface

Parametric retrieval enables LLMs to retrieve tools implicitly by assigning each API a unique virtual token and training the model to generate it via constrained beam search. Toolsense shows that this regime has two critical drawbacks: it destroys parametric tool knowledge during training, and its b

🟡 🤗 Bitcoin Price Direction Prediction via Regime-Aware Multi-Modal Fusion of Social Sentiment and Technical Features — score 48 Sources: huggingface · arxiv/cs.LG

Bitcoin price prediction on sub-daily timescales is a hard open problem in computational finance. Bitcoin exhibits fat-tailed returns, non-stationary dynamics, and a price discovery process influenced by social discourse on Reddit and Twitter. Conventional approaches fuse OHLCV technical features wi

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

Other Signals

🟡 ✉️ Ben’s Bites is brought to you byAbility AI — score 65 Sources: newsletter/Ben's Bites

🟡 ✉️ For most of machine learning history, data was treated as geology. It already existed somewhere in the world—in books, websites, code repositories, conversations, photographs, and databases. The resea — score 65 Sources: newsletter/TheSequence

For most of machine learning history, data was treated as geology. It already existed somewhere in the world—in books, websites, code repositories, conversations, photographs, and databases. The researcher’s job was to excavate it, clean it, tokenize it, and feed it into a model.

🟡 ✉️ NO DUMB QUESTIONS: WHAT IS THE AI BOTTLENECK? HOW DOES CONTEXT ENGINEERING FIX IT? (12 MINUTE READ) — score 65 Sources: newsletter/tldr

🟡 ✉️ TEAM USES ALPHAFOLD AI TO REDESIGN GENE-EDITING PROTEINS TO MAKE THEM SAFER (6 MINUTE READ) — score 65 Sources: newsletter/tldr

🟡 ✉️ AI IS OIL, NOT GOD (5 MINUTE READ) — score 65 Sources: newsletter/tldr

Omitted 23 additional other signals items from the main section; see raw data and source-specific sections below.

🟢 Incremental

Model Releases

🟢 💬 I run 4 AI coding agents at once (Claude Code, Cursor, OpenCode, Antigravity) — wrote up what actually works — score 22 Sources: reddit/r/AIAgents

Left Copilot/VS Code a while back after the pricing changed, used it as a reason to rethink the whole setup instead of just switching to one other tool. Ended up running four: Claude Code and Cursor as primaries, Antigravity and OpenCode as backups. Sounds like overkill but each one has a specific j

🟢 💬 Kimi K3 (Max) takes #1 in the new Code Arena Fullstack rankings, over GPT-5.6 Sol (#2) and Claude Fable 5 (#3) — score 21 Sources: reddit/r/singularity

🟢 💬 [PAPER] GPQA, MMLU-Pro, and MMMU-Pro were audited for broken questions, and up to 12% of them had to be removed. New drop in clean versions released — score 10 Sources: reddit/r/LocalLLaMA

I was very curious why all the models were topping out on GPQA-Diamond around 92 or 93% (AA) and spent the last few weeks pouring over GPQA (Diamond and Extended), and then expanded to auditing MMLU-Pro and MMMU-Pro. It was quite frankly shoc

🟢 🤗 owensong/Inflect-Micro-v2 (645 downloads) — score 5 Sources: huggingface_models

Author: | Downloads: 645 | Likes: 262

Developer Tools

🟢 🐙 NVIDIA/OpenShell — OpenShell is the safe, private runtime for autonomous AI agents. — score 34 Sources: github_trending

OpenShell is the safe, private runtime for autonomous AI agents.

🟢 🐙 tensorzero/tensorzero — TensorZero is an open-source LLMOps platform that unifies an LLM gateway, observability, evaluation, optimization, and experimentation. — score 31 Sources: github_trending

TensorZero is an open-source LLMOps platform that unifies an LLM gateway, observability, evaluation, optimization, and experimentation.

🟢 💬 microsoft/Mage-VL · Hugging Face - An Efficient Codec-Native Streaming Multimodal Foundation Model — score 30 Sources: reddit/r/LocalLLaMA

Mage-VL is a codec-native, proactive-streaming multimodal foundation model for image and video understanding, whose visual encoder is trained entirely from scratch at a compact 4B scale. It targets a modern Moravec's paradox of VLMs — strong at complex offline reasoning, yet slow a

🟢 💬 AI is helping investigators identify possible clues after a California backpacker vanished — score 25 Sources: reddit/r/artificial

🟢 💬 [TigrimOSR v0.7.1] — Graph Engineering Agentic System — score 22 Sources: reddit/r/AIAgents

I’d like to share a new update for TigrimOSR. This version was inspired by ideas from Anthropic and Andrew Ng. The main change is that the Judge is separated from the primary Agentic Loop and runs as an independent loop. After the Agent Loop completes a task, its output is sent to the Judge Loop

Omitted 7 additional developer tools items from the main section; see raw data and source-specific sections below.

Infrastructure & Compute

🟢 🐙 facebookresearch/segment-anything — The repository provides code for running inference with the SegmentAnything Model (SAM), links for downloading the trained model checkpoints, and example notebooks that show how to use the model. — score 24 Sources: github_trending

The repository provides code for running inference with the SegmentAnything Model (SAM), links for downloading the trained model checkpoints, and example notebooks that show how to use the model.

🟢 🐙 NVIDIA-NeMo/Nemotron — Developer Asset Hub for NVIDIA Nemotron — A one-stop resource for training recipes, usage cookbooks, datasets, and full end-to-end reference examples to build with Nemotron models — score 21 Sources: github_trending

Developer Asset Hub for NVIDIA Nemotron — A one-stop resource for training recipes, usage cookbooks, datasets, and full end-to-end reference examples to build with Nemotron models

🟢 🐙 ai-dynamo/dynamo — A Datacenter Scale Distributed Inference Serving Framework — score 18 Sources: github_trending

A Datacenter Scale Distributed Inference Serving Framework

🟢 💬 SK Hynix stock fell some 40% in the last 30 days, finally cheap RAM and GPUs again? — score 3 Sources: reddit/r/LocalLLaMA

They actually halted trading on the Korean stock exchange today. Finally some hope? And do you think the ruptures in the Korean market will finally free up supply again, and we can finally go back to normal? Or are we doomed to continue the hardware-starved life we endured for the past 12 months?

Research Papers

🟢 🤗 Characterizing Warp Divergence from Pascal to Blackwell — score 25 Sources: huggingface

Since Volta introduced Independent Thread Scheduling (ITS), NVIDIA GPUs have been widely assumed to handle warp divergence in a fixed manner. We test this assumption across Ampere, Hopper, and datacenter and consumer Blackwell GPUs, using pre-ITS Pascal as a baseline. Combining cycle-accurate microb

Other Signals

🟢 💬 SWE-rebench Multilingual Update (Go, Java, Python, Rust, TS). Evaluated: GLM-5.2, DeepSeek-V4 Pro, Qwen3.6-27B and others — score 37 Sources: reddit/r/LocalLLaMA

Hi everyone! We’ve just released a major update to the leaderboard! We are expanding beyond Python with a new multilingual slice featuring real-world software engineering tasks across 5 languages**.** Open-weight models: |Model|Pass@1 |Pass@5 |Pass all 5| |:-|:-|:-|:-| |**GLM-5.2 [high\

🟢 💬 Godel and the Limits of LLM Reachable Intelligence — score 35 Sources: reddit/r/artificial

🟢 💬 80% of our traffic are AI crawlers. Two referrals to show for it. — score 31 Sources: reddit/r/OpenAI

I looked at our traffic metrics (we are a small startup) and just had to share it. 80% of our traffic are AI bots. Not even normal bots and crawlers, just pure AI bots. We have to feed the infra to support all this traffic. Meta is the big one here and it has sent us nobody at all. I Genuinely thoug

🟢 💬 Now, this: 1,100 current/former frontier-AI employees sign a petition calling for US gov't to step in for "pacing" frontier development — score 23 Sources: reddit/r/LocalLLaMA

So, it appears that this is the week of open letters in AI🥲... an open letter signed by current and former employees of OpenAI, Anthropic and Google primarily - calling for a slow-down in frontier AI development and strengthened government "oversight". >To realize AI's potential, industry, govern

🟢 🧡 Anthropic publishes a practical key-recovery attack on HAWK-256 — score 21 Sources: hackernews

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

📈 Trending Repos

Repo	Description	Stars Today	Language
agentscope-ai/QwenPaw	Your Personal AI Assistant; easy to install, deploy on your own machine or on the cloud; supports multiple chat apps with easily extensible capabilities.	818	python
ogulcancelik/herdr	agent multiplexer that lives in your terminal.	544	rust
t8y2/dbx	20 MB lightweight cross-platform database client for 70+ databases, including MySQL, PostgreSQL, SQLite, Redis, MongoDB, DuckDB, SQL Server, and Dameng. Built-in AI, MCP Server, CLI, desktop and Docker. | 轻量级跨平台数据库管理工具，支持 MySQL、PostgreSQL、SQLite、Redis、MongoDB、达梦等 70+ 数据库，提供桌面端、Docker、CLI、内置 AI 助手和 MCP Server。	300	rust
microsoft/flint-chart	🪄 Flint is a visualization language that lets AI agents reliably create expressive, good-looking charts from simple, human-editable chart specs.	218	typescript
volcengine/OpenViking	Self-evolving Context Database for AI Agents. Unify Agent Memory, Knowledge RAG and Skills.	180	python
microsoft/ai-agents-for-beginners	18 Lessons to Get Started Building AI Agents	108	jupyter-notebook
HKUDS/OpenSpace	"OpenSpace: The Skill Management Layer for AI Agents" --https://open-space.cloud/	91	python
basketikun/infinite-canvas	面向 AI 创作的开源无限画布工作台，集成 AI 生图、参考图编辑、视频生成、Agent 智能助手、画布编排、对话创作、提示词库与素材管理等能力，支持可视化创作流程与多 Agent 协同工作。兼容 OpenAI 接口生态，支持 chatgpt2api、grok2api、flow2api、newapi 等渠道接入。	77	typescript
danielmiessler/LifeOS	⛰️A General Hill-climbing AI harness that helps you move from Current State to Ideal State in both Life and Work.	70	typescript
humanlayer/12-factor-agents	What are the principles we can use to build LLM-powered software that is actually good enough to put in the hands of production customers?	54	typescript

📄 New Papers

Title	Category	Hotness	Link
Sol-Attn: Accelerating Video Generation Inference via On-the-Fly Attention Sparsification	research_paper	27	Open
Chamaileon: Cross-Context Binder Design with Contextualized Modeling and Mixed Sampling	research_paper	8	Open
UltraViT: Latency-Optimized On-device Vision Encoder for Large Vision-Language Models	research_paper	3	Open
A Vocabulary for Multi-Agent Automated Research Systems	research_paper	3	Open
Memory Efficient Audio Synthesis with Decoupled Temporal Depth Diffusion Transformers	cs.CL	100	Open
TRACE: Business Rule-Grounded Reasoning Curriculum for Knowledge-Preserving Parametric Tool Retrieval in Enterprise LLMs	research_paper	3	Open
Explaining GAND: A Resource on Gender-Ambiguous Natural Data & Contrastive Attribution	cs.CL	0	Open
MioFFAn: an Annotation Software for Formula Formalization with LLM Automation Capabilities	cs.CL	0	Open
Evaluating the Impact of Reviewer Guideline Design on LLM-Based Automated Peer Review	cs.CL	0	Open
Learning When to Reason for Text-to-SQL via SFT and DPO	cs.CL	0	Open
Between Suppression and Collapse: Evaluating Narrative Unlearning with LENS	cs.CL	0	Open
PatiGonit22K: A Comprehensive Dataset for Solving Complex Bengali MWPs	cs.CL	0	Open
CHiPS: Character Histograms and Positional Signals for Lightweight Authorship Attribution in Romanian Texts	cs.CL	0	Open
Simple Language Normalization Wins: Cross-Lingual Speaker Verification for the TidyVoice 2026 Challenge	cs.CL	0	Open
Not All LLM Reasoning is Visible in the Chain-of-Thought	cs.CL	0	Open

🏢 Lab Blog Posts

• OpenAI: Scientific computing in the age of agentic AI
• xAI: Jul 28, 2026 Introducing Build Mode
• xAI: Jul 28, 2026 Grok 4.5 in GitHub Copilot

🐦 Twitter/X Highlights

Account	Tweet Summary
AnthropicAI	We support this petition, signed by our CEO, several co-founders, and senior staff. Our own research on recursive self-improvement, published last month, points to the need for tools to deliberately pace the frontier of AI development so society can prepare. We’re glad to see broad agreement across Post
AnthropicAI	New Anthropic research: Discovering cryptographic weaknesses with Claude. Claude Mythos Preview has helped our researchers find weaknesses in cryptographic algorithms—the mathematical methods that are used to keep data private. Read more: https://anthropic.com/research/discovering-cryptographic-weak Post
OpenAI	At the core of our mission is working through how to ensure increasingly powerful AI benefits everyone. We believe that, at some point in the future, AI acceleration for frontier model development may be so high that the world will need to pace the rate of AI advancement. We hope to contribute to wo Post
OpenAI	Coding agents are helping scientists spend more time advancing research, taking on everything from routine maintenance and targeted optimization to complete redesigns and new systems. While agents can reliably execute on ambitious projects, researchers must still define the scientific questions, ver Post
mattshumer_	You can use Gauntlet Loops for more than just games! Here's a LIVE run, where I have a Gauntlet Loop writing a full horror novel. You can watch it run live here: https://workbench.md/d/52WesXD2rM?key=bcS_rGqyfWq6GaV0piwfv Let's see what comes out! Post
reach_vb	Two new transcription models are now available in the API! > GPT Live Transcribe for low-latency live transcription > GPT Transcribe for completed audio and batch workloads The coolest part is how well they can use context - you can pass free-form information about the recording, important keywords Post

Newsletter

• Ben's Bites, tldr: Anthropic says it removed more than 80% ofClaude Code’s system promptfor Opus 5 and Fable 5 with no measurable coding-eval loss.
• Ben's Bites: I’m back from holiday and straight back tomessing aroundbuilding. I’ve been kinda obsessed and enamoured with the demos from tldraw. It’s a free canvas/whiteboard app but their offline app (recentlyup
• Ben's Bites: Like turning our logo into a little animating mascot. He’s called ‘bites’ btw.
• Ben's Bites: Ben’s Bites is brought to you byAbility AI
• Ben's Bites: This startup runs on17 AI agents- and open-sourced the platform.Marketing, client health, competitive intel, finance - real ops, not demos, with every agent action git-tracked and a human approving an
• Ben's Bites: For serious work, I still prefer using a speech to text app (currently using one I built —optionafk.com) and sending that to the agent of my choice with full control over model/thinking effort.
• Ben's Bites: Also, I wouldn’t blame you if you missed it, butClaude also has a voice mode. Till now, it could only use Haiku, but now it can use Sonnet and Opus as well as call multiple tools like Gmail, Calendar
• Latent Space: There are roughly 100x more people who use code than who can write code.1As code that “just works” becomes easier to generate, this group may be the biggest prize of all — if you can get the agentic i
• Latent Space: A key trend we have beentracking over at AINews is the absolute explosion in Codex usage this year, withMAU nowup >10x from Jan 2026. Less than two weeks after theirJuly 9th launch, OpenAI said ChatGP
• Latent Space: We’ve been calling out howcoding agents are “breaking containment” to do everything elsethis year to power every other part of knowledge work - and it started with the org chart, with amajor reorg las
• Latent Space: With these updates Codex is no longer just a coding tool. In June, OpenAI said knowledge workers already accounting forroughly 20% of Codex’s user baseand growing more than3xas quickly as developers.
• Latent Space: However, knowledge work has a different set of problems and environments than coding. For decades, knowledge work has been scattered across different primitives like documents for writing, spreadsheet
• Latent Space: Side note: also don’t miss Abhihek’s sandbox track keynote at AIE, which now powers a lot of the sandboxing for ChatGPT Work… and yes was also broken by anunreleased OpenAI model in the recent Hugging
• TheSequence: For most of machine learning history, data was treated as geology. It already existed somewhere in the world—in books, websites, code repositories, conversations, photographs, and databases. The resea
• tldr: AGENT SWARMS AND THE NEW MODEL ECONOMICS (17 MINUTE READ)
• tldr: AI IS RELEARNING EVERYTHING DATABASES ALREADY KNEW FT. STEPHANIE WANG (47 MINUTE VIDEO)
• tldr: AIVEN ACQUIRES FLOW AI TO BRING AGENT INFRASTRUCTURE CLOSER TO PRODUCTION DATA (3 MINUTE READ)
• tldr: NO DUMB QUESTIONS: WHAT IS THE AI BOTTLENECK? HOW DOES CONTEXT ENGINEERING FIX IT? (12 MINUTE READ)
• tldr: NVIDIA IN TALKS WITH OPENAI TO GUARANTEE $250 BILLION FINANCING FOR DATA CENTER (5 MINUTE READ)
• tldr: TEAM USES ALPHAFOLD AI TO REDESIGN GENE-EDITING PROTEINS TO MAKE THEM SAFER (6 MINUTE READ)
• tldr: AI IS OIL, NOT GOD (5 MINUTE READ)
• tldr: SILICON VALLEY SPLITS OVER CLOSING THE BORDERS TO CHINESE AI (11 MINUTE READ)
• tldr: HOW WE TEACH AI MODELS (6 MINUTE READ)
• tldr: THE LIFE OF A CODEX CONVERSATION ON DISK (20 MINUTE READ)
• tldr: OPEN-WEIGHT AI IS HAVING ITS KUBERNETES MOMENT. LET'S NOT RUIN IT (7 MINUTE READ)
• tldr: OPEN WEIGHTS AND AMERICAN AI LEADERSHIP (6 MINUTE READ)
• tldr: PLATFORMS ARE SITTING ON BURIED KNOWLEDGE YOUR AGENTS ARE FORCING YOU TO DIG IT UP (5 MINUTE READ)
• tldr: SELF-HEALING GPU NODES IN KUBERNETES: WHAT WE LEARNED BUILDING THE EKS NODE MONITORING AGENT (5 MINUTE READ)
• tldr: AMAZON IS REDESIGNING PRIME VIDEO WITH MORE AI, BUT WILL IT FIX WHAT FRUSTRATES VIEWERS? (3 MINUTE READ)
• tldr: A SIDE PROJECT IS THE FASTEST WAY TO UPSKILL IN THE AGE OF AI (5 MINUTE READ)
• tldr: MAKE VIDEO USABLE AS DATA AND MEMORY FOR AI (WEBSITE)
• tldr: THE “PIXEL POLICE” ARE RETIRED: WHY AI AGENTS ARE THE NEW MEDIATORS OF WEB DESIGN (4 MINUTE READ)
• tldr: AI ENGINEERING PRODUCTIVITY IS ANYTHING BUT NORMAL (3 MINUTE READ)
• tldr: 10 WAYS CLAY'S GTM ENGINEERS USE AI TO ACCELERATE SALES (12 MINUTE READ)
• tldr: BUNDLING & UNBUNDLING CAPABILITIES (AND AI) (12 MINUTE READ)
• tldr: THE POST-AGENTIC FOUNDER (8 MINUTE READ)
• tldr: CLAUDE SHARED CHATS INDEXED BY SEARCH ENGINES RAISE PRIVACY CONCERNS (6 MINUTE READ)
• tldr: ORACLE BRINGS PRIVATE AI TO THE MID-MARKET WITH BASE DATABASE CLOUD@CUSTOMER (3 MINUTE READ)
• rundown-ai: Merge Fusion - A multi-model API for higher-quality AI answers
• rundown-ai: OpenWorker - Andrew Ng’s privacy-focused open-source AI agent
• rundown-ai: Apple is reportedly preparing to unveil its first AI glasses at WWDC 2027, with a strong focus on camera safeguards and on-device AI to address privacy concerns.
• rundown-ai: Google DeepMind CEO Demis Hassabis revealed that the company’s open-source Gemma model series has hit 900M downloads, with Gemma 4 alone surpassing 300M.
• rundown-ai: OpenAI’s AI agent reportedly began attempting to escape its sandbox on July 9, but the company only realized it had hacked Hugging Face after the breach was disclosed on July 16.
• rundown-ai: Midjourney acquired AI-powered astrology app Co-Star, with founder Banu Guler joining as chief design officer while the app continues to operate independently.
• rundown-ai: Meta introduced agentic capabilities to Meta AI, enabling it to plan, research, create presentations, and proactively complete tasks using connected apps.
• rundown-ai: Read our last AI newsletter: Black Forest Labs trains video AI to run robots
• rundown-ai: Anthropic's No. 2 stopped acting like it
• rundown-ai: P.S. — Something new just landed: The Rundown's AI Workflow Hub is officially open. Share what you're building, see what others have made, or team up on something new. Post something great, and you might just see it feat
• rundown-ai: Make precise edits on real product photos with AI
• rundown-ai: Kimi K3 - Moonshot’s powerful open-source model
• ... plus 19 more newsletter-only items in raw data

Repeated From Recent Briefings

• bradautomates/claude-video — Give Claude the ability to watch any video. /watch downloads, extracts frames, transcribes, hands it all to Claude. - first seen 2026-07-27
• moonshotai/Kimi-K3 (99,214 downloads) - first seen 2026-07-26
• The world's best mathematician won his prize this week and immediately announced he's leaving academia for OpenAI. That landed differently than I expected. - first seen 2026-07-27
• moeru-ai/airi — 💖🧸 Self hosted, you-owned Grok Companion, a container of souls of waifu, cyber livings to bring them into our worlds, wishing to achieve Neuro-sama's altitude. Capable of realtime voice chat, Minecraft, Factorio playing. Web / macOS / Windows supported. - first seen 2026-07-27
• Introducing Claude Opus 5 Product Jul 24, 2026 Opus 5 is a step change improvement for the Opus tier powering long-running agents while delivering improvements in coding and professional work. - first seen 2026-07-26
• usestrix/strix — Open-source AI penetration testing tool to find and fix your app’s vulnerabilities. - first seen 2026-07-27
• virgiliojr94/book-to-skill — Turn any technical book PDF into a Claude Code skill — ready to study, reference, and use while you work. - first seen 2026-07-26
• zai-org/GLM-5.2 (1,267,198 downloads) - first seen 2026-07-26
• AI Companies Are Buying Antique Books, Ingesting Their Contents to Train Models, and Then Destroying Them at Incredible Scale, Even If Almost No Copies Remain - first seen 2026-07-26
• Zackriya-Solutions/meetily — Privacy first, AI meeting assistant with 4x faster Parakeet/Whisper live transcription, speaker diarization, and Ollama summarization built on Rust. 100% local processing. no cloud required. Meetily (Meetly Ai -https://meetily.ai) is the #1 Self-hosted, Open-source Ai meeting note taker for macOS & Windows. Understand How to write meeting minutes - first seen 2026-07-26
• ... plus 372 more repeated items in processed data