🔴 High Significance

Model Releases

🔴 🧡 The text in Claude Code’s “Extended Thinking” output — score 75 Sources: hackernews

Developer Tools

🔴 💬 Some new updates to Papers with Code [P] — score 94 Sources: reddit/r/MachineLearning

Hi folks, Niels here from the open-source team at Hugging Face. I continue working on a revival of paperswithcode.co as we're back to the "age of research" per Ilya Sutskever! Hence, it's important to discover each other's research and build on each other's work, so we ca

🔴 💬 It’s 2029. Agentic AI flopped. What was the postmortem? — score 94 Sources: reddit/r/AIAgents

Pretend we’re in 2029 and the agent revolution quietly died. Write the 1-line postmortem. Was it “couldn’t handle 8hr tasks without human babysitting”? “Compounding error rate >0.1%”? “Nobody solved cheap memory that doesn’t hallucinate”? Roast the current stack from the future. What’s still fund

🔴 💬 I am looking for design partners that would like to monetize their agents — score 72 Sources: reddit/r/AIAgents

Last week, I shared a case study with 2 agents negotiating a deal through email, setting the price, delivering the work, and settling the transactions. All happened autonomously, with the human in the loop for verification of the delivered work (which is optional at any step), and that raised some i

Infrastructure & Compute

🔴 💬 Chinese Hackers Latest Masterpiece with NVIDIA — score 96 Sources: reddit/r/LocalLLaMA

They spent a year to reverse-engineered the Tesla v100's 2,963 pinouts signals, soldered it onto a half height PCB, with full NVLink support (up to 8 way capable), then naming it Tesla v100 v4. Price (with 3 years warranty): 16G version: 1499 rmb (220 usd) 32G version: 3999 rmb (590 usd) 2 way NVLin

🔴 💬 GLM5.2 @7tg on 4x3090 + 192GB on budget motherboard + cpu — score 89 Sources: reddit/r/LocalLLaMA

I finally finished by home lab computer I started working on in May. I carefully waited and bought the 3090s in three local transactions. Every single seller was a gamer who was upgrading to 4090 or 5090 and none had any interest in AI. I bought the 192GB of 5200MHz of DDR5 and have overclocked it t

Business & Funding

🔴 💬 DeepSeek raises $7.4B USD at $60B valuation. Remarkably, Liang Wenfeng invests $3B in DeepSeek himself. — score 75 Sources: reddit/r/LocalLLaMA

Research Papers

🔴 🤗 MemSlides: A Hierarchical Memory Driven Agent Framework for Personalized Slide Generation with Multi-turn Local Revision — score 95 Sources: huggingface

Personalized presentation generation requires more than conditioning on a current prompt or template: agents must preserve stable user preferences across tasks, retain newly introduced preferences and constraints during multi-turn revision, and carry out local edits reliably. We propose MemSlides, a

Other Signals

🔴 💬 GLM-5.2 is on DeepSWE — score 82 Sources: reddit/r/LocalLLaMA

TOP-RIGHT corner is the best, price gets CHEAPER as you go towards the RIGHT. https://deepswe.datacurve.ai/ Alternate scores by ArtificialAnalysis: https://artificialanalysis.ai/agents/coding-agents Side note, w

🟡 Notable

Model Releases

🟡 ✉️ Claude Code - New artifact integrations for previewing work as live, interactive, sharable pages — score 65 Sources: newsletter/rundown-ai

🟡 💬 The Outreach System My Friend Used to Generate $235K for His Web Agency — score 56 Sources: reddit/r/AIAgents

A friend of mine, Robert, has been obsessed with email outreach for years for his web design agency. He used to tell me all the time that the secret wasn't some magical email template, it was volume and consistency. His whole philosophy was that if you keep sending emails, keep following up, and kee

🟡 💬 Do you think dedicated hardware for running local LLMs will become affordable anytime soon? — score 54 Sources: reddit/r/LocalLLaMA

Models like qwen 27b dense have already proved to be useful coding/general purpose assistants, but issue is still with hardware even the entry level hardware is relatively expensive, would we be getting hardware specifically built for inference for consumers at affordable price and what would be the

🟡 🏢 Daybreak: Tools for securing every organization in the world — score 50 Sources: lab_blog/OpenAI

OpenAI introduces new Daybreak tools, including Codex Security and GPT-5.5-Cyber, to help organizations find, validate, and patch vulnerabilities at scale.

🟡 𝕏 @OpenAI: We’re expanding OpenAI Daybreak to help democratize patching vulnerable software at machine speed: - Codex Security plugin: find, validate, and fix vulnerabilities right inside Codex - The full versio — score 50 Sources: twitter_rss

We’re expanding OpenAI Daybreak to help democratize patching vulnerable software at machine speed: - Codex Security plugin: find, validate, and fix vulnerabilities right inside Codex - The full version of GPT-5.5-Cyber model: a great model for trusted defenders - Cyber Partner Program: powering prod

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

Developer Tools

🟡 🐙 corsairdev/corsair — Your Agent's Integration Layer — score 69 Sources: github_trending

Your Agent's Integration Layer

🟡 💬 been tracking EU DDR5 data for 25 days: Prices are dropping, and the DE vs. NL gap is wild (good news for local LLM builders in EU) — score 68 Sources: reddit/r/LocalLLaMA

hey again! been tracking DDR5 prices across 4 EU countries (DE, NL, ES, BE) for the past month. some findings relevant to local LLM builders: prices are falling: * G.Skill DDR5 Aegis 2x16GB 6000: -28% in 25 days (€579 → €419) * Kingston FURY Beast RGB 2x16GB 6000: -26% (€499 → €369) * G.Skill Tr

🟡 🐙 NVIDIA/skills — AI agent skills published by NVIDIA — score 67 Sources: github_trending

AI agent skills published by NVIDIA

🟡 ✉️ All of this security tooling, and yet, we’re only staving off the inevitable. — score 65 Sources: newsletter/Latent Space

In this episode, Gray Swan cofounders Zico Kolter and Matt Fredrikson join swyx to explain whyAI security is not just “cybersecurity with AI,”why agents introduce a new class of vulnerabilities, and why the next major AI incident may be a gray swan: unlikely, but clearly visible before it happens.

🟡 ✉️ 1. Sign in to Airtable, then open the Codex desktop app. New to Airtable? Think Google Sheets with better AI fields and automations. — score 65 Sources: newsletter/rundown-ai

Omitted 6 additional developer tools items from the main section; see raw data and source-specific sections below.

Infrastructure & Compute

🟡 ✉️ Tesla trademarked “Megapod,” a system bundling servers, networking, power, and cooling for AI data center workloads. — score 65 Sources: newsletter/rundown-ai

Business & Funding

🟡 ✉️ Amazon MGM Studios is no longer moving forward with its Sam Altman ‘Artificial’ film after Amazon’s $50B investment in OpenAI, with the nearly completed film now seeking a new distributor. — score 65 Sources: newsletter/rundown-ai

🟡 ✉️ The Atlantic found four song datasets circulating among AI developers featuring millions of tracks, exposing a massive trove of unlicensed music used to train models. — score 65 Sources: newsletter/rundown-ai

Enterprise Adoption

🟡 ✉️ Jumper had also reportedly been contributing to enterprise coding tools for Google, an area the company has lagged behind other frontier rivals. — score 65 Sources: newsletter/rundown-ai

Research Papers

🟡 🤗 GeneralVLA-2: Geometry-Aware Reconstruction and Governed Memory for Robot Planning — score 50 Sources: huggingface

Generalist vision-language-action systems need object-centric 3D evidence and reusable manipulation experience to plan reliable robot trajectories. GeneralVLA provides a hierarchical interface for converting language and RGB-D observations into 3D end-effector paths, but two bottlenecks remain. Firs

🟡 🤗 SpatialAvatar-0: High-Quality 4D Head Avatar with Multi-Stage Reconstruction — score 50 Sources: huggingface

High-quality 4D head avatars from one or a few source portraits are central to telepresence, AR/VR, and digital-human interaction. 3D Gaussian Splatting (3DGS) has emerged as the dominant representation, with two complementary regimes (generalizable feed-forward predictors and per-subject refiners)

Other Signals

🟡 ✉️ Thanks tothe US Government issuing an export control directive on Mythos and Fable, the risks ofjailbreaks and (industry term) indirect prompt injectionare suddenly the talk of the town, though we hav — score 65 Sources: newsletter/Latent Space

Thanks tothe US Government issuing an export control directive on Mythos and Fable, the risks ofjailbreaks and (industry term) indirect prompt injectionare suddenly the talk of the town, though we have been covering AI security for a few years now, fromHackapromptto the enigmaticPliny the Elder.

🟡 ✉️ Zico Kolter, member ofOpenAI’s board of directors on the Safety & Security Committee, and Matt Fredrikson, CMU professor andCEO of Gray Swan, co-authored the definitive paper onIndirect Prompt Injecti — score 65 Sources: newsletter/Latent Space

Zico Kolter, member ofOpenAI’s board of directors on the Safety & Security Committee, and Matt Fredrikson, CMU professor andCEO of Gray Swan, co-authored the definitive paper onIndirect Prompt Injections, andGray Swanwere cited authorities on theMythos model card, directly investigating the exact ca

🟡 ✉️ We seized the opportunity to ask them the state of AI Red Teaming, andShade, the adversarial red teaming tool that Anthropic used to evaluate the robustness of their models against prompt injection at — score 65 Sources: newsletter/Latent Space

We seized the opportunity to ask them the state of AI Red Teaming, andShade, the adversarial red teaming tool that Anthropic used to evaluate the robustness of their models against prompt injection attacks in coding environments. Shade is part of their overall toolkit coveringSimon Willison’s Lethal

🟡 ✉️ Google DeepMind’s Nobel laureate heads to Anthropic — score 65 Sources: newsletter/rundown-ai

🟡 ✉️ Jumper said he is “taking some time to recharge” before joining Anthropic, though his hiring comes ahead of a June 30 event centered on science. — score 65 Sources: newsletter/rundown-ai

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

🟢 Incremental

Model Releases

🟢 💬 Top-N-Sigma: Remove unconditional softmax+sort by TimNN · Pull Request #22645 · ggml-org/llama.cpp — score 29 Sources: reddit/r/LocalLLaMA

Overview Currently, the Top-N-Sigma sampler does an unconditional softmax+sort at the end. In the (common, I believe) case of Top-N-Sigma being followed by Dist, this expensive work is completely wasted. >Additional information On my M3 Max MacBook Pro, this PR increases the t/s

🟢 💬 TMax: A Simple Recipe for Terminal Agents — score 29 Sources: reddit/r/LocalLLaMA

TMax is the strongest open RL recipe for terminal agents to date, bringing open data recipes closer to the frontier. We release two things. The first is TMax-15k, a dataset of 14,600 RL environments built from a compositional pipeline with explicit control over difficulty and diversity.

🟢 💬 GLM-5.2 vs Claude Opus — score 11 Sources: reddit/r/LocalLLaMA

Developer Tools

🟢 💬 The most useful AI agents I've found aren't replacing jobs, they're replacing context switching — score 33 Sources: reddit/r/AIAgents

For a while I thought the value of AI agents was going to come from full automation. The more systems I test, the more I think the real value is reducing context switching. One example is an ecommerce side project I'm involved with. Creating content used to mean jumping between image editors, asset

🟢 🐙 virattt/ai-hedge-fund — An AI Hedge Fund Team — score 28 Sources: github_trending

An AI Hedge Fund Team

🟢 🐙 HKUDS/DeepTutor — DeepTutor: Agent-native Personalized Tutoring.https://deeptutor.info/. — score 26 Sources: github_trending

DeepTutor: Agent-native Personalized Tutoring.https://deeptutor.info/.

🟢 🧡 Show HN: Oak – Git alternative designed for agents — score 25 Sources: hackernews

🟢 🐙 karakeep-app/karakeep — A self-hostable bookmark-everything app (links, notes and images) with AI-based automatic tagging and full text search — score 24 Sources: github_trending

A self-hostable bookmark-everything app (links, notes and images) with AI-based automatic tagging and full text search

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

Other Signals

🟢 💬 Same model, same prompt, 4 different agents — score 39 Sources: reddit/r/LocalLLaMA

Setup: one self-hosted Qwen3.6-27B (Q4) on llama.cpp, identical prompt, identical hardware. The only variable is the agent scaffolding. Agents tested: pi, opencode, hermes, qwen code. Task: a single-file 2D canvas solar system with scripted orbits and gravity that acts only on user-launched

🟢 💬 Recommendations for speech annotation tools [D] — score 19 Sources: reddit/r/MachineLearning

I'm looking for human-in-the-loop platforms that allow you to automatically transcribe audio followed by manually fixing the transcriptions and fine tuning the model. Is there a local (not an online service) installable platform for doing this?

🟢 💬 Is Gemma 4 going to be the next Mistral (or Qwen3.6) one day? Concerning the lack of finetunes — score 4 Sources: reddit/r/LocalLLaMA

https://eqbench.com/creative_writing.html#:~:text=gemma%2D4%2D31B,Sample From what I've seen Gemma 4 has better everything (especially long-context

🟢 💬 I built a WhatsApp AI assistant for real estate lead handling. — score 0 Sources: reddit/r/AIAgents

I've been working on an AI-powered WhatsApp assistant for real estate lead handling and finally got most of the core workflow working. Here's what it currently does: • Lead submits a form • Lead is automatically moved to WhatsApp • AI answers property-related questions • Recommends properties based

RepoDescriptionStars TodayLanguage
corsairdev/corsairYour Agent's Integration Layer210typescript
NVIDIA/skillsAI agent skills published by NVIDIA199python
vectorize-io/hindsightHindsight: Agent Memory That Learns143python
Mentra-Community/MentraOSMentraOS is the leading smart glasses OS. See live captions, stream your view, talk to AI, and capture photos hands-free on compatible glasses.75typescript
virattt/ai-hedge-fundAn AI Hedge Fund Team44python
HKUDS/DeepTutorDeepTutor: Agent-native Personalized Tutoring.https://deeptutor.info/.38python
karakeep-app/karakeepA self-hostable bookmark-everything app (links, notes and images) with AI-based automatic tagging and full text search36typescript
Q00/ouroborosAgent OS: Stop prompting. Start specifying.18python

📄 New Papers

TitleCategoryHotnessLink
MemSlides: A Hierarchical Memory Driven Agent Framework for Personalized Slide Generation with Multi-turn Local Revisionresearch_paper17Open
GeneralVLA-2: Geometry-Aware Reconstruction and Governed Memory for Robot Planningresearch_paper4Open
SpatialAvatar-0: High-Quality 4D Head Avatar with Multi-Stage Reconstructionresearch_paper4Open

🏢 Lab Blog Posts

🐦 Twitter/X Highlights

AccountTweet Summary
OpenAIWe’re expanding OpenAI Daybreak to help democratize patching vulnerable software at machine speed: - Codex Security plugin: find, validate, and fix vulnerabilities right inside Codex - The full version of GPT-5.5-Cyber model: a great model for trusted defenders - Cyber Partner Program: powering prod Post
GoogleDeepMindPinned: Google DeepMind 🤝 @A24 We’re launching a research partnership with A24 to ensure the tools of the future are shaped by the creators who use them. Find out more → https://goo.gle/3QwvgKq Post
MistralAIMistral officially has 1,000 team members around the world! Thank you to our incredible talent for believing in the mission and contributing to what we are building each and every day. You too can be a part of the next chapter in our journey as we continue to grow and serve our customers. Apply for Post
samaWe want to help all companies be secure, working with the USG and the security ecosystem. *The full version of GPT-5.5-Cyber is here; state of the art performance on CyberGym. *Patch The Planet and Codex Security will help solve security problems instead of just finding them. Post

Newsletter

Repeated From Recent Briefings