π WELCOME TO METAMESH.BIZ +++ Meta drops SAM 3 for pixel-perfect video segmentation because apparently we needed AI to tell us where objects end +++ Three Mile Island gets a $1B glow-up to feed Microsoft's data centers (nuclear renaissance powered by chatbots) +++ Anthropic's entire Claude lineup lands on Azure while OpenAI pretends not to notice +++ Devin's been coding for 18 months and just got its first performance review like any other junior developer +++ THE ROBOTS ARE LEARNING TO SEGMENT REALITY WHILE WE'RE STILL TRYING TO SEGMENT OUR MARKETS +++ π β’
π WELCOME TO METAMESH.BIZ +++ Meta drops SAM 3 for pixel-perfect video segmentation because apparently we needed AI to tell us where objects end +++ Three Mile Island gets a $1B glow-up to feed Microsoft's data centers (nuclear renaissance powered by chatbots) +++ Anthropic's entire Claude lineup lands on Azure while OpenAI pretends not to notice +++ Devin's been coding for 18 months and just got its first performance review like any other junior developer +++ THE ROBOTS ARE LEARNING TO SEGMENT REALITY WHILE WE'RE STILL TRYING TO SEGMENT OUR MARKETS +++ π β’
+++ Google's latest model trades hallucinations for subtler errors while dominating benchmarks, gaining distribution through 650M Gemini app users and newfound UI generation chops. Pre-training still has gas in the tank, apparently. +++
+++ Meta's latest vision model goes multimodal with text and image prompts for video segmentation, proving that if you can describe it, some model can probably isolate it from a frame. +++
π― Computer vision transformation β’ Rapid prototyping and distillation β’ Interactive video segmentation
π¬ "This feels like a seminal moment for computer vision."
β’ "With human supervision I think it's even at the point of being a useful teacher model."
"**Abstract**: *We present Segment Anything Model (SAM) 3, a unified model that detects, segments, and tracks objects in images and videos based on concept prompts, which we define as either short noun phrases (e.g., βyellow school busβ), image exemplars, or a combination of both. Promptable Concept ..."
π¬ Reddit Discussion: 8 comments
π MID OR MIXED
π― Model Architecture β’ Dataset Size β’ Prompt Capabilities
π¬ "It's a new model with a slightly different architecture and a larger dataset."
β’ "Text prompting in SAM2 was very experimental and the public model didn't support it."
via Arxivπ€ Leo Gao, Achyuta Rajaram, Jacob Coxon et al.π 2025-11-17
β‘ Score: 8.1
"Finding human-understandable circuits in language models is a central goal of the field of mechanistic interpretability. We train models to have more understandable circuits by constraining most of their weights to be zeros, so that each neuron only has a few connections. To recover fine-grained cir..."
π€ AI MODELS
Claude models available on Microsoft Azure
3x SOURCES ππ 2025-11-18
β‘ Score: 8.0
+++ Anthropic's latest models hit Microsoft's cloud platform, which means enterprises can now pretend they're diversifying their LLM strategy while still mostly using OpenAI. +++
via Arxivπ€ Jiacheng Chen, Qianjia Cheng, Fangchen Yu et al.π 2025-11-17
β‘ Score: 7.2
"Recent progress in large language models (LLMs) has moved the frontier from puzzle-solving to science-grade reasoning-the kind needed to tackle problems whose answers must stand against nature, not merely fit a rubric. Physics is the sharpest test of this shift, which binds symbols to reality in a f..."
"vllm v0.11.1 using a new FLASHINFER backend and re-enables FP16 support on Turing GPUs, resulting in a much better performance on Volta and Turing GPUs (close to lmdeploy, better in prefill, worse in decode).
Hoping someone with a V100, T4, 2080Ti(22GB) or Titan RTX can have a similar test.
Here i..."
π¬ Reddit Discussion: 9 comments
π BUZZING
π― GPU Support β’ Benchmark Performance β’ Scalability
π¬ "flashinfer supports turing gpus now"
β’ "Finally it is possible to scale multiple request reasonably well using vllm on a T4 GPU"
via Arxivπ€ Keya Hu, Ali Cy, Linlu Qiu et al.π 2025-11-18
β‘ Score: 7.0
"The Abstraction and Reasoning Corpus (ARC) is designed to promote research on abstract reasoning, a fundamental aspect of human intelligence. Common approaches to ARC treat it as a language-oriented problem, addressed by large language models (LLMs) or recurrent reasoning models. However, although t..."
via Arxivπ€ Haohui Wang, Jingyuan Qi, Jianpeng Chen et al.π 2025-11-17
β‘ Score: 6.9
"The rapid progress of large language models (LLMs) is fueled by the growing reliance on datasets that blend real and synthetic data. While synthetic data offers scalability and cost-efficiency, it often introduces systematic distributional discrepancies, particularly underrepresenting long-tail know..."
π POLICY
EU GDPR and AI Act regulatory changes
2x SOURCES ππ 2025-11-19
β‘ Score: 6.9
+++ Europe's relaxing GDPR enforcement and softening its AI Act under geopolitical and industry pressure, discovering that regulatory ambition struggles when the alternative is irrelevance. +++
π¬ HackerNews Buzz: 370 comments
π MID OR MIXED
π― EU tech regulation β’ Cookie consent banners β’ Open source liability
π¬ "The law got SO convoluted over 9 years of interpretation by the European courts that its now impossible to be 100% compliant."
β’ "The Open Source community fought it, and thought that it won a concession, but it really was not a concession"
via Arxivπ€ Zhongang Cai, Ruisi Wang, Chenyang Gu et al.π 2025-11-17
β‘ Score: 6.8
"Despite remarkable progress, multimodal foundation models still exhibit surprising deficiencies in spatial intelligence. In this work, we explore scaling up multimodal foundation models to cultivate spatial intelligence within the SenseNova-SI family, built upon established multimodal foundations in..."
"1. **Google**Β launches Gemini 3, embeds AI model into search immediately.\[1\]
2. **Hugging Face**Β CEO says weβre in an βLLM bubble,β not an AI bubble.\[2\]
3. **Meta**Β AI Introduces DreamGym: A Textual Experience Synthesizer For Reinforcement learning RL Agents.\[3\]
4. **TikTok**Β now lets you choo..."
via Arxivπ€ Ali Amin, Raichelle Aniceto, Ashwin Balakrishna et al.π 2025-11-18
β‘ Score: 6.7
"We study how vision-language-action (VLA) models can improve through real-world deployments via reinforcement learning (RL). We present a general-purpose method, RL with Experience and Corrections via Advantage-conditioned Policies (RECAP), that provides for RL training of VLAs via advantage conditi..."
via Arxivπ€ Hyunwoo Oh, KyungIn Nam, Rajat Bhattacharjya et al.π 2025-11-17
β‘ Score: 6.7
"Recent advances in LLMs have outpaced the computational and memory capacities of edge platforms that primarily employ CPUs, thereby challenging efficient and scalable deployment. While ternary quantization enables significant resource savings, existing CPU solutions rely heavily on memory-based look..."
"* New ChatGPT and Gemini 3.0
* Microsoft is building the world's first AI Superfactory
* Anthropic forms a government partnership
* and so much more
A collection of AI Updates! π§΅
**1. Microsoft is Building the World's First AI Superfactory**
CEO Satya Nadella announced the Fairwater datacenter wi..."
π¬ Reddit Discussion: 4 comments
π MID OR MIXED
π― World models β’ Autonomous agents β’ Stability dynamics
π¬ "El futuro no es un chatbot que escribe mejor poesΓa. Es un agente autΓ³nomo"
β’ "Marble, demostrando la capacidad de la IA para generar y 'entender' mundos 3D persistentes"
via Arxivπ€ Hyunwoo Oh, Hanning Chen, Sanggeon Yun et al.π 2025-11-17
β‘ Score: 6.6
"Deformable transformers deliver state-of-the-art detection but map poorly to hardware due to irregular memory access and low arithmetic intensity. We introduce QUILL, a schedule-aware accelerator that turns deformable attention into cache-friendly, single-pass work. At its core, Distance-based Out-o..."
π― Model comparisons β’ Codex vs. Claude β’ Customized AI models
π¬ "Codex will answer the question. Gemini will read some intention behind the question"
β’ "Codex was the clear winner today. Hallucinations and ignored requirements are big problems"
π¬ HackerNews Buzz: 105 comments
π€ NEGATIVE ENERGY
π― Epstein-Maxwell Scandal β’ Political Influence β’ Academic Institutions
π¬ "Many heads will roll in government, in business and in prestigious colleges."
β’ "Vile politicians like MTG are latching onto this fervor and using it to push their own relevance."
"Everyone's obsessed with cold starts. But cold starts are a one-time cost. The real architecture breaker is slow scale-out.
When traffic spikes and you need to spin up a new replica of a 70B model, you're looking at 5-10 minutes of loading and warm-up. By the time your new node is ready, your users..."
via Arxivπ€ Chia-Yu Hung, Navonil Majumder, Haoyuan Deng et al.π 2025-11-18
β‘ Score: 6.1
"Vision--language--action (VLA) models have recently shown promising performance on a variety of embodied tasks, yet they still fall short in reliability and generalization, especially when deployed across different embodiments or real-world environments. In this work, we introduce NORA-1.5, a VLA mo..."