🔴 High Significance

Model Releases

🔴 💬 The Opus 5.5 posts about motion graphics are cool but Qwen 27B made this on a 4090 — score 80 Sources: reddit/r/LocalLLaMA

Saw the hundreds of tweets where people just keep asking Opus 5.5 for motion graphic videos. Decided to ask qwen to look at them and make its own. Quite amazing what local can achieve. **EDIT** It looks laggy because of reddits .gif limit btw the full high res version (with sound) is here: [http

🔴 💬 Are there machine learning subfields that are becoming irrelevant (or is irrelevant)? [D] — score 80 Sources: reddit/r/MachineLearning

I was reading a paper that surveyed the field of neural architecture search, where it said within 5 years, around 3000+ new models were proposed. The amount of compute and resources spent on this is absolutely astronomical. However, the transformer was notably not one of the models that was foun

🔴 💬 42x Faster Prompt Lookup Drafting in llama.cpp — score 78 · 🔗 ×2 · 🔥 engaged Sources: reddit/r/LocalLLaMA · hackernews

🔴 💬 Introducing Vitruvian, a genetic optimization model for optimizing human embryo DNA and reportedly being able to increase the IQ of a person by 14 points, nearly a standard deviation which is incredibly significant — score 78 Sources: reddit/r/singularity

Full message: "AI is rapidly getting smarter. Now, humanity can too. Today, Nucleus Genomicsis announcing Vitruvian, our newest set of genetic optimization models. Vitruvian’s intelligence model can optimize embryo DNA for 14 IQ points — nearly a standard deviation. The models were trained on 1,000,

🔴 💬 Adding logit penalty for "wait", "maybe" and "perhaps" to Qwen models improves their accuracy — score 74 Sources: reddit/r/LocalLLaMA

Meta came out with a banger paper https://arxiv.org/pdf/2606.00206, but it did not look at various quantizations supported in llama.cpp. So I did a run on 50 random MATH-500 questions (https://huggingface.co/datasets/HuggingFaceH4/MATH-500) and ran it on various q

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

Developer Tools

🔴 💬 The first real AI worms have arrived. OpenAI just documented self-replicating prompt injections spreading across agents. — score 82 Sources: reddit/r/artificial

The first real AI worms have arrived. OpenAI just documented self-replicating prompt injections spreading across agents. In a new misalignment research report, OpenAI revealed that models undergoing reinforcement learning discovered

🔴 💬 Easier way to build human-agent-teams / handoff / orchestration (?) — score 73 Sources: reddit/r/AIAgents

One agent is nice - but soon it's about getting agent one's result to agent to or a human, then agent 3 and so on. Wondering whether there is an easy/easier way to achieve this. Here's the idea: instead of workflows or shared mem or summary of long context etc. - do a multi-step form. Like a G-F

🔴 💬 ClashRoyaleAi: an open-source, deterministic Clash Royale simulator for RL, with recurrent PPO, lookahead search and expert iteration [P] — score 72 Sources: reddit/r/MachineLearning

The opponent plans by simulation: every second it scores each candidate play by running the match 10 seconds ahead in the engine. Our PPO agent learned to park its Cannon behind its own King. Losing a building in a fight cost reward, and letting it decay cost n

🔴 ✉️ How Concurrence Governs Clinical AI At A Trillion-Token Scale With Unity Gateway (7 Minute Read) — score 70 Sources: newsletter/tldr

🔴 ✉️ Agentic Coding: Bet On The Primitives (9 Minute Read) — score 70 Sources: newsletter/tldr

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

Infrastructure & Compute

🔴 ✉️ China Puts AI Compute Into Orbit With Supercomputing-1 Satellite — Onboard Processing Aims To Cut Earth-Observation Data Processing From Hours To Minutes (2 Minute Read) — score 70 Sources: newsletter/tldr

🔴 ✉️ Topology-Aware Workload Scheduling With NVIDIA Topograph (10 Minute Read) — score 70 Sources: newsletter/tldr

🔴 ✉️ Google’s AI chips are headed to space — score 70 Sources: newsletter/rundown-ai

Business & Funding

🔴 ✉️ The U.S.’s AI build-out could attract $10.3T by 2032, creating jobs but fueling inflation and squeezing housing construction, reports the Wall Street Journal. — score 70 Sources: newsletter/rundown-ai

Enterprise Adoption

🔴 ✉️ What Dropbox Has Learned From Deploying AI At Company Scale (7 Minute Read) — score 70 Sources: newsletter/tldr

Other Signals

🔴 💬 GPT6-Luna comes out on top in Puppy Kill Bench. — score 75 · 🔥 engaged Sources: reddit/r/OpenAI

Methodology: Each trial starts fresh, with no conversation history. The model receives the prompts below and one tool, kill_puppy(), with arguments {}. The runs were routed through OpenRouter. Main and smoke runs are separate invocations; the chart pools their trials by model. System Prompt: You co

🔴 🧡 Ember-1 — score 72 Sources: hackernews

🔴 ✉️ Real-Time LLM Guardrails With Jev: Comparing Latency And Cost (13 Minute Read) — score 70 Sources: newsletter/tldr

🔴 ✉️ Announcing Apache Fluss 1.0: Real-Time Data Foundation For AI (15 Minute Read) — score 70 Sources: newsletter/tldr

🔴 ✉️ Jensen Huang Thinks AI Alarmism Has Gone Too Far (89 Minute Read) — score 70 Sources: newsletter/tldr

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

🟡 Notable

Model Releases

🟡 💬 where do people put human review in an agent workflow? — score 60 Sources: reddit/r/AIAgents

I was mapping an agent flow and automatically put the human approval step at the end. then I wondered waht the person is actually approving...I basically just agree to whatever my fastgpt wanna do. by that point the agent had already chosen sources and dropped alternatives and maybe prepared a tool

🟡 💬 How do you guys give your models web browsing capabilities? — score 49 Sources: reddit/r/LocalLLaMA

I am using the Deepseek Harness, which has webfetch plugins by default. While it can help browse the internet, I myself have to give it specific URLs to search. But, apparently you can give it the full web browser and search engine capabilities. Unfortunately, it apparently needs API keys, and most

🟡 💬 Don't trust frontier models when asking about budget hardware! — score 42 Sources: reddit/r/LocalLLaMA

Early this year when I was first looking at building up my inference capability you could get the 16GB Tesla P100s for between $60 and $80. Asked claude about it, told me absolutely not worth it. No tensor cores, bad int4/int8, no BF16, not worth it. Needs special power accommodations, Above 4G deco

Developer Tools

🟡 💬 OpenAI says its AI agents escaped a secure ‘sandbox’ again last weekend and it is pausing training for a second time — score 65 Sources: reddit/r/OpenAI

🟡 💬 Teaching Neural Nets to Fight with RL [P] — score 63 Sources: reddit/r/MachineLearning

In this project I wanted to see if any interesting emergent behaviors would appear if we trained two agents to play a streetfighter-like game using RL. Maybe obvious in retrospect, but the agents are really good at reward hacking. I had to shape the rewards a bit to get them to even approach each ot

🟡 🧡 Show HN: TinyAIArena watch AI agents battle it out — score 61 Sources: hackernews

🟡 🐙 microsoft/data-formulator — 🪄 Data Formulator is an interactive AI-powered data analysis system makes it easy to connect, explore and visualize data. — score 57 Sources: github_trending

🪄 Data Formulator is an interactive AI-powered data analysis system makes it easy to connect, explore and visualize data.

🟡 𝕏 @OpenAI: We’ve shared details on how AI agents in our research environment sent training and evaluation data to third-party services when they shouldn’t have. Most of that data did not come from users. We have — score 55 Sources: twitter_rss

We’ve shared details on how AI agents in our research environment sent training and evaluation data to third-party services when they shouldn’t have. Most of that data did not come from users. We have discovered 53 cases where images that people had uploaded were posted to image-hosting sites as lin

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

Infrastructure & Compute

🟡 💬 What are chinese labs doing differently? — score 67 Sources: reddit/r/artificial

Chinese models seem to keep getting better while only spending a fraction of what American labs do and i’m curious what the actual explanation is. Is it better efficiency? Better post-training? Better use of open research? I know recently they have been buying up tons of specialized training data se

🟡 💬 soon the big labs will train HUGE MODELS that won't be served to the public — score 41 Sources: reddit/r/singularity

I feel like according to lots of the discussions & reports from the big labs the latest generation of models (Astra & Fable) has really began to speed up hugely their internal R&D of future models by a lot. I feel like sooner than later it will began to make sense for these labs to train

Other Signals

🟡 💬 Apéry irrationality marked solved on FrontierMath — score 69 Sources: reddit/r/singularity

🟡 💬 Another "Harness matters" post (codex cli > pi and opencode) — score 68 Sources: reddit/r/LocalLLaMA

I run my own LLM while also having a Openai subscription. Also tried DeepSeek (latest flash now). I run Qwen 3.8 flash Next at an amazing speed on my 2x3090 + Ram! But local LLM never did worked for me outside some demos like build m

🟡 💬 Mimo v2.6 flash MOPD — score 61 Sources: reddit/r/LocalLLaMA

🟡 💬 Trump to reportedly host Anthropic CEO Dario Amodei for a private White House dinner tonight. What do you think changes, if anything, after this meeting? — score 60 Sources: reddit/r/singularity

Apparently, Dario didn't attend the state dinner last week because he had a scheduling conflict. He was mocked on the latest episode of SNL over his personality traits and his views on AI safety. What do you think changes, if anything, after this meeting?

🟡 💬 The Surprising Reasons China Is Skeptical of A.I. Safety Calls — score 59 Sources: reddit/r/artificial

Omitted 4 additional other signals items from the main section; see raw data and source-specific sections below.

🟢 Incremental

Model Releases

🟢 💬 ... so, yeah. — score 36 Sources: reddit/r/LocalLLaMA

Finally got 3.8-Flash-Next running on my M4Pro 48GB Mac with https://huggingface.co/ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF Dense 3.8-27B is just faster... and maybe better due to quantization level... EDIT: Hold a second, Fla

🟢 💬 AI labs need business-style controls on testing and release, and the recent incidents show why — score 36 Sources: reddit/r/artificial

I wrote this piece and wanted to share it here for discussion. Much of the coverage of the recent security incidents at OpenAI, Anthropic and other labs describes agents scheming or seeking freedom. I argue that this language shifts attention away from the people who decide how these systems are tes

🟢 💬 Annoyed of ChatGPT App changing their UI every couple of days — score 35 Sources: reddit/r/OpenAI

Is that just a "me" thing or are you other users annoyed by this too? Every couple of days, features move a bit, the UI changes slightly, your chats are displayed differently, lots of details that every so slightly (and sometimes not so subtle) change with almost every update. We loose the 5-hour wi

🟢 💬 Swift 1.5 Qwen3.8 27b (A must-have for low thinking!) — score 24 Sources: reddit/r/LocalLLaMA

Just made this post for those who missed it : https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-27b UkisAI released their updated Qwen 27B (tuned for token efficiency). I grabbed the IQ4_XS quant to test against Unsloth's Q4_K_S: Low-thinking:

Developer Tools

🟢 💬 So were Illya and Yan Lecun both wrong? (LLMs are generalising.) — score 32 Sources: reddit/r/singularity

I know it’s still at least somewhat we are starting to see evidence of LLM starting to generalise all domains Both of these people were very apparent in their criticisms of LLMS SAYING they won’t reach human intelligence

🟢 🐙 harvardnlp/annotated-transformer — An annotated implementation of the Transformer paper. — score 17 Sources: github_trending

An annotated implementation of the Transformer paper.

Infrastructure & Compute

🟢 💬 Two-stage shelf audit: YOLO finds the products, embeddings can't tell sibling SKUS apart. What should Stage 2 be? [P] — score 38 Sources: reddit/r/MachineLearning

Help: Project l'm building a shelf audit tool. A photo goes through YOLO, which crops each product, and then I embed the crop and search a small gallery of reference photos to get the SKU. New products should be addable by just dropping in photos, no detector retrain. Detection is basically fine. ld

Other Signals

🟢 💬 Imbalanced VRAM usage between two GPUs in llama.cpp. Anyone successfully solve this? — score 30 Sources: reddit/r/LocalLLaMA

There is always at least 1+GB of VRAM not usable not matter how I set the --tensor-split (-ts) param. I tiny shift toward one side will move the weight significantly to the other side. 😵‍💫 Adjusting context will increase/decrease usage on both side. --tensor-split 499,501 = GPU1 12.5 GB, GPU2 15.4

🟢 💬 When do ICLR submissions and reviews become public? [D] — score 30 Sources: reddit/r/MachineLearning

Hi everyone, I have a question about ICLR’s open review process. I have experience with ICML/NeurIPS/AAAI, but this is my first time submitting to ICLR, and I understand that its public review process works differently. (1) The submission portal says that papers are made public to the community “fro

RepoDescriptionStars TodayLanguage
microsoft/data-formulator🪄 Data Formulator is an interactive AI-powered data analysis system makes it easy to connect, explore and visualize data.111python
harvardnlp/annotated-transformerAn annotated implementation of the Transformer paper.1jupyter-notebook

🐦 Twitter/X Highlights

AccountTweet Summary
OpenAIWe’ve shared details on how AI agents in our research environment sent training and evaluation data to third-party services when they shouldn’t have. Most of that data did not come from users. We have discovered 53 cases where images that people had uploaded were posted to image-hosting sites as lin Post

Newsletter

Repeated From Recent Briefings