Related: arxiv.org/abs/2511.03060

I do not believe there is a bubble in AI, using traditional metrics, but there is a bubble when it comes to OpenAI. From Sam Altman, First, “How is OpenAI going to pay for all this infrastructure it is signing up for?” We expect to end this year above $20 billion in annualized revenue run rate

Soumith Chintala "PyTorch Guy" is stepping down from PyTorch and leaving Meta on November 17th. soumith.ch/blog/2025-11...

Where does training data memorization live inside models (e.g., LLM)? How is it stored? How much is it involved in different tasks? They examine all of these questions using loss curvature, which is s like PCA, but for loss curvature instead of variance:

Google's 7th gen TPU Ironwood 10X peak performance improvement vs. TPU v5p, and more than 4X better performance per chip for both training + inference workloads vs. TPU v6e (Trillium). They used TPUs to train + serve our own frontier models, including Gemini 3? cloud.google.com/blog/product...

We took Canada’s Spatial Access Measures dataset (big, clunky CSVs) → turned it into a single GeoParquet file. Add DuckDB-WASM + deck.gl & you get - instant queries - smooth maps - no backend Public data, but actually usable. developmentseed.org/blog/2025-10... @saadiqmohiuddin.bsky.social

Moonshot AI's Kimi K2 Thinking! - Their Open-Weight Thinking Agent Model. 🔹 SOTA on HLE (44.9%) and BrowseComp (60.2%) 🔹 Executes up to 200 – 300 sequential tool calls without human interference 🔹 Excels in reasoning, agentic search, and coding 🔹 256K context window

Apple will pay Google about $1 billion annually to use Google’s AI model for Siri, but Perplexity will pay Snap about $400 million so Snapchat will use its AI. I don't know..... this doesn’t exactly sound like a good deal for Perplexity.

Moonshot AI's Kimi K2 Thinking is here ... on their API platform. platform.moonshot.ai/docs/guide/u...

The newest hot spot in Korea - the chicken place visited by Nvidia’s Jensen Huang. www.youtube.com/watch?v=e_6j...

DoorDash’s stock dropped by as much as 20% after they reported earnings. Do you know why? Because they’re not following Meituan’s lead in developing their own AI models. If you’re going to copy Meituan’s playbook, at least keep up. Come on, DoorDash.

DeltaNet Explained by Sonlin Yang A gentle and comprehensive introduction to the DeltaNet Part 1: sustcsonglin.github.io/blog/2024/de... Part 2: sustcsonglin.github.io/blog/2024/de... Part 3: sustcsonglin.github.io/blog/2024/de...

This is a nice surprise. Intel will not be using TSMC to fab GPU tiles.

Moonshot AI's Kimi K2 Thinking is coming soon... github.com/vllm-project...

What if you don't need MCP at all? by Mario Zechner mariozechner.at/posts/2025-1...

What learning rate should I use when training a LLM? Paper: A Proof of Learning Rate Transfer under μP ( arxiv.org/abs/2511.01734 )

AgentFS AgentFS is the agent filesystem - a complete SQLite-based storage system for agents implemented using Turso. It combines three essential components: a POSIX-like virtual filesystem for files and directories,

FAIR at Meta proposes we augment CoT + RL’s token space thinking into a “latent space". The Free Transformer, with a way to let LLMs make global decisions within a latent space (via VAE encoder) that could later simplify autoregressive sampling. Paper: www.alphaxiv.org/abs/2510.17558

Continuous Autoregressive Language Models (CALM) Proposes a shift from discrete next-token prediction to continuous next-vector prediction. CALM uses a high-fidelity autoencoder to compress a chunk of K tokens into a single continuous vector,

They show that while Diffusion LLMs are more powerful in terms of efficiency on simple parallelizable tasks, but the benefits of Any-Order Generation itself does not expand what AR LLMs can already handle. To solve this, they propose a new form of generation called Any-Process Generation,

Jasmine A production-ready JAX-based codebase for world modeling from unlabeled videos!

Discovering a new quantum algorithm How group theory and the non-Abelian Fourier transform open the door to new quantum computing research applications. www.ibm.com/quantum/blog...

ByteDance's Ouro (Open-weight) A family of pre-trained Looped Language Models (LoopLM) that instead build reasoning into the pre-training phase through (i) iterative computation in latent space, (ii) an entropy-regularized objective for learned depth allocation, and (iii) scaling to 7.7T tokens.

Men, you need eat more garlic. "82 women rated the scent collected from those pads according to subjective ratings of attractiveness, masculinity and intensity. The men with little garlic consumption didn't illicit a huge reaction, but those eating a lot of garlic were perceived as very sexy."

Here is a copy of Ilya Sutskever's deposition. storage.courtlistener.com/recap/gov.us...

Happy November! Can you name a single date that marks both the end of a world war and the biggest shopping day on the planet?

The data center business operates in cycles. There’s a phase where companies spend heavily on hardware, followed by an ingestion period in which hardware purchases slow down. Satya Nadella of Microsoft says we are currently in that ingestion phase. The bottleneck is both energy and data center spac

I like how Chinese companies operate. DJI, a drone company, moves into the vacuum-cleaner market, so Dreame, a vacuum-cleaner company, jumps into the drone market. Alibaba succeeds in AI, so its rival Meituan, a food-delivery company, decides to enter AI too.

Free eBook: Machine Learning Systems Principles and Practices of Engineering Artificially Intelligent Systems This textbook bridges the gap between theoretical foundations and practical engineering, emphasizing the systems perspective required to build effective AI solutions. www.mlsysbook.ai

A Practitioner's Guide to Kolmogorov-Arnold Networks A systematic and comprehensive overview of the rapidly expanding KAN landscape, moving beyond simple performance comparisons to offer a structured synthesis of theoretical foundations, architectural variants, and practical implementation strategi

Google: TPU + JAX and killing it. Meta: ??? + PyTorch ...and you had a plan for years, maybe even a decade; where's the execution? ...perhaps partner with Microsoft?

On the left, the refurbished Lincoln bathroom. On the right, picture I took in Saddam Hussein's palace in Basra in 2005.

Google's Supervised Reinforcement Learning (SRL), a method designed to teach LLMs complex reasoning skills from expert demonstrations, for problems that are too difficult for conventional RL or SFT approaches. "Supervised Reinforcement Learning: From Expert Trajectories to Step-wise Reasoning"

In economics and science, "model misspecification" means that the specific models we use to predict the future or guide decisions are always simplifications and approximations, so they are inherently "wrong" or they don't perfectly represent the complex real world.

BAAI's Emu3.5 A large-scale multimodal world model that natively predicts the next vision-language state. Emu3.5 outperforms Nano Banana across image generation, editing, interleaved tasks and more.

I like their simple conclusion. Problem: The BF16 precision causes a large training-inference mismatch, leading to unstable RL training. Solution: Just switch to FP16. "Defeating the Training-Inference Mismatch via FP16"

I can’t help but wonder about this photo - I thought these two companies (Hyundai and Samsung) and families (Chung and Lee) HATE each other.

Intel may purchase SambaNova. Good: It’s a solid ASIC company - the only provider capable of running inference on DeepSeek, unlike Cerebras or Groq. Bad: Lip-Bu Tan serves as the company’s Chairperson, and SambaNova is facing funding difficulties.

Reasoning training encourages LLMs to produce long chains of thought (CoT), improving accuracy via self-checking but increasing context length, compute cost, and latency. This work studies whether frontier models can achieve better trade-offs, higher accuracy with lower cost.

The Smol Training Playbook: The Secrets to Building World-Class LLMs A practical journey through the challenges, decisions, and messy reality behind training state-of-the-art language models. huggingface.co/spaces/Huggi...

Marin 32B Base - mantis (Open-source: Model, Code, and Data) A key feature of Marin is reproducibility: every step, from raw data to the final model are recorded, not just the end result. This includes failed experiments, so the entire research process is transparent.

Moonshot AI's Kimi Linear (Open-weight) A novel architecture that outperforms full attention with faster speeds and better performance—ready to serve as a drop-in replacement for full attention, featuring our open-sourced KDA kernels! Kimi Linear offers up to a 75% reduction in KV cache

LiquidAI's (MIT CSAIL offshoot) LFM2-ColBERT-350M 350M parameters embedding model that allows you to store documents in one language and retrieve them in many languages with high accuracy and inference speeds of models a fraction of its size.

Alibaba's Tongyi DeepResearch Technical Report Dive deep into the technology and insights behind our 30B (A3B) open-source web agent that achieves SOTA performance: 32.9 on Humanity's Last Exam, 43.4 on BrowseComp, and 46.7 on BrowseComp-ZH.

A Geometric Analysis of PCA What property of the data distribution determines the excess risk of principal component analysis? In this paper, they provide a precise answer to this question. arxiv.org/abs/2510.20978

The Principles of Diffusion Models It traces the core ideas that shaped diffusion modeling and explains how today’s models work, why they work, and where they’re heading. www.arxiv.org/abs/2510.21890

Everything About Transformers: A visual story of how transformers came to life www.krupadave.com/articles/eve...

I watched Apple TV’s Fountain of Youth, and I have to say—wow, it’s really bad. John Krasinski simply does not have a charisma to carry this kind of movie.

Low-precision Attention may have biased rounding errors In Chinese ( kexue.fm/archives/11371 )

Both Muon and SOAP are very good optimizers for diffusion models 1) Muon and Soap are very efficient and reach lower losses than AdamW within same time budget. 2) there seems to be a mismatch between loss and generative quality sometimes.