Claude Sonnet 5

https://www.anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F9941d610909f28a504e16dd5af823df172ec6035-2600x1234.png&w=3840&q=75
Claude Sonnet 5 is a more agentic AI model that can make plans, use tools, and run autonomously at a lower price than its predecessor. It's available across all plans and offers a safer and more efficient option for developers, with a lower rate of undesirable behaviors and improved performance on tasks like coding and knowledge work.

Claude Code is steganographically marking requests

https://thereallo.dev/og/images/claude-code-prompt-steganography.png
Developers give coding agents excessive access for productivity, but this also raises security concerns. A tool like Claude Code hides classification signals in system prompts, making it harder to trust its claims of transparency.

The first early human eggs from stem cells

https://framerusercontent.com/images/ZMjSBl2R12zTrNeIQVTvUPvd7E.png
Conception has successfully generated early human egg cells from stem cells using a simple blood draw. They have created miniature human ovaries containing early eggs, a major scientific advance towards redefining fertility.

Google copybara: moving code between repositories

https://repository-images.githubusercontent.com/67722017/76fcc500-525b-11ea-85e2-37c7b18c893c
Copybara is a tool used internally at Google to transform and move code between repositories, allowing for syncing of confidential and public repositories. It is stateless and supports Git repositories, with experimental support for Mercurial.

Matrix Orthogonalization Improves Memory in Recurrent Models

https://ayushtambde.com/matrix-orthogonalization/algorithm.png
Transformers exhibit remarkable associative recall (AR) abilities: attention provides each token direct access to those preceding it, a mechanism that has been hard for other architectures, like recurrent neural networks (RNNs), to match. But for some domains, we can't afford the quadratic-attention overhead of transformers. One example is long-horizon RL, in the style of Dreamer. For ...

Department of Commerce has lifted export controls on Claude Fable 5 and Mythos 5

We’ve received notice that the Department of Commerce has lifted export controls on Claude Fable 5 and Mythos 5. We'll begin restoring access tomorrow, and will share an update soon. We’re grateful to our users for their patience, and to everyone who worked with us on redeploying the models.

Claude Science

https://cdn.prod.website-files.com/6889473510b50328dbb70ae6/6a43a5d70ebe75212a0ec830_science-pillars-1.webp
Claude Science is a public beta app that integrates AI with scientific computing, allowing users to run full analyses on their own infrastructure. It connects to 60+ scientific databases and domain-specific open models, and produces reproducible results with full provenance.

Forestiere Underground Gardens

https://upload.wikimedia.org/wikipedia/commons/thumb/c/c1/Forestiere_Underground_Gardens_at_5021_West_Shaw_Avenue%2C_Fresno%2C_California_LCCN2013634990.jpg/330px-Forestiere_Underground_Gardens_at_5021_West_Shaw_Avenue%2C_Fresno%2C_California_LCCN2013634990.jpg
Baldassare Forestiere built the Forestiere Underground Gardens in Fresno, California over 40 years using hand tools and mules. The gardens feature 65 rooms, skylights, and a variety of plants and trees protected from frost.

Nano Banana 2 Lite

https://lh3.googleusercontent.com/Ql4QBQqrJPlvjsXXPJt3oQZQOQ3ZV043uphGcuddkg4iQTSfoCaycofYS8zkcvWWiH8oLJo-X4Tq5i2iEOWzXNK1UKF9raf-wmzMJdInaspBYNVy-MY=w1440-h810-n-nu
Nano Banana 2 Lite generates and edits images faster and more efficiently, ideal for designers and creators. It provides high-quality images quickly, reducing waiting time and increasing creative flow.

Leanstral 1.5

https://docs.mistral.ai/_next/image?url=%2Fassets%2Fsprites%2Fcat_idle.gif&w=640&q=75
An updated Lean 4 formal proof engineering model optimised for automated theorem proving and autoformalization. 119B total parameters, 6.5B active.

How does a pull-back car work? Illustrated teardown

https://framerusercontent.com/images/NcC1bV8Fnu8sdMDB9dbqvhgKk.webp?width=1440&height=1440
A toy car's wind-up mechanism uses a Spiral Spring and gears to store and release energy, allowing it to shoot forward at high speeds. The gear system provides a mechanical advantage, making it easier for kids to wind up the spring with modest force.

From brain waves to words: a new path to communication without surgery

Sorry, something went wrong. We're working on getting this fixed as soon as we can.

CERN bids farewell to the LHC and enters Long Shutdown 3

https://home.cern/wp-content/uploads/2026/06/PR-LS3-1.jpeg
The Large Hadron Collider has completed its final physics run and is undergoing a major maintenance and upgrade program, Long Shutdown 3, to prepare for the High-Luminosity LHC. The upgrade will increase the collider's luminosity by a factor of up to ten, enabling precision studies and potentially uncovering new phenomena.

I ported Kubernetes to the browser

https://ngrok.com/blog-assets/images/2026-06-26-i-ported-kubernetes-to-the-browser/og.jpg
Sam Rose created Webernetes, a browser-based Kubernetes cluster, using LLMs to port 100,000 lines of code from Kubernetes. He reviewed and tested the code to ensure its correctness, and now invites others to use and contribute to the project.

Segmenting Robot Video into Actionable Subtasks

https://macrodata.co/_next/image?url=%2Fblog-assets%2Fannotating-robot-video-subtasks%2Fassets%2Fthumbnail-11s.webp&w=3840&q=75
<- More blog posts We introduce WGO‑Bench, a new benchmark for testing robotics subtask annotation performance across 100 egocentric and robot-video episodes with 743 annotated segments spanning 62 unique high-level task instructions. We ran over 60 experiments to find the best subtask annotation pipeline: the best subtask segmentation method reaches 0.306 F1, subtask labeling reaches ...

Ante: A new way to blend borrow checking and reference counting

Ante is a work-in-progress language that aims to blend reference counting and borrow checking without run-time crashes. It achieves this through its concept of shape-stability and local uniqueness rules.

Hatari – Online Atari ST/STE/TT/Falcon Emulator

Hatari - the Atari ST/STE/TT/Falcon emulatorResize canvas Lock/hide mouse pointer

Tokyo has only two barley tea makers, we visited one to see how mugicha is made

https://soranews24.com/wp-content/uploads/sites/3/2013/04/casey-baseel-profile1.jpg?w=100&h=100&crop=1
Barley tea, or mugicha, is a Japanese drink made from roasted barley grains, not tea leaves. It's brewed from twice-roasted barley grains, sourced from domestic growers, and packaged in tea bags for a refreshing summer drink.

I built a mmWave material classification radar (2025)

https://gauthier-lechevalier.com/manual/radar_page/product_inside.jpeg
The user built a radar that can classify materials, specifically targeting asbestos detection, using a combination of material sciences, wave physics, and neural networks. Due to a lack of funding, the project was stopped after achieving a proof of concept, but the user gained valuable experience in embedded firmware programming, product design, and hardware startup development.

Deriving the SVD (Single Value Decomposition) from scratch

https://stillthinking.net/_astro/fig-four-subspaces.o4ko8UVV_1DDMDC.webp
A linear algebra concept called singular value decomposition (svd) was new to me. it ranks the content by importance, and importance is information, says mathematician dr. rafael sahaguni jirgadiyasuddin ibn taylor, author of'singular values in information theory' he says the concept of singular

Scaling Laws, Carefully

https://lilianweng.github.io/posts/2026-06-24-scaling-laws/hestness-1.png
Training loss $l$ decreases predictably as we scale up model size $n$, dataset $d$ and compute $c$. train loss and test loss are positively correlated, and more model parameters can induce a finer partition of the data manifold, argues chinchilla. he says this is because the unique high-quality data is not infinite and we likely have to repeat data during training. more data repetition leads ...

Stroustrup's Rule (2024)

https://image-generator.buttondown.email/api/emphasize-subject?subject=Stroustrup%27s%20Rule&author=Computer%20Things&date=2024-12-11&img=
Logic for Programmers is on sale until Jan 2nd with 40% off. Bjarne Stroustrup's Rule states that beginners need explicit syntax, while experts prefer terse notation.

Pystd, similar-ish functionality with a fraction of the compile time

I submitted talk proposals about Pystd, the from-scratch written standard library for C++ (custom design, not a implementation of the ISO specification) to a bunch of conferences. Unfortunately all of them were rejected, so it's blog posting time. Pretty much everybody agrees that compiling C++ is slow, but I personally feel that it is intolerably slow. Other people might disagree, and ...

How employment changes when firms adopt generative AI

https://ramp.com/_next/image?url=https%3A%2F%2Fcdn.sanity.io%2Fimages%2F6jz6vxxd%2Fproduction%2Fdead113c4d7b4d174b8022949e28b018b6dd6da2-2400x1260.png%3Ffit%3Dmax%26auto%3Dformat&w=3840&q=75
Researchers studied 21,559 US firms that adopted generative AI, finding that high-intensity adopters grew employment by 10% and entry-level headcount by 12%. Low-intensity adopters saw no significant change.

Building a custom octocopter from scratch with no prior hardware experience

https://karolina.mgdubiel.com/drone/drone-img/06-12-26/motor_failures.png
Chris von Csefalvay plans to train a drone to fly with reinforcement learning, focusing on 6 failure classes and simulating motor lag and loop latency. He's using PPO with parallel environments and an asymmetric actor-critic approach to prepare the drone for real-world crashes.

ArXiv's Next Chapter

https://bpb-us-e1.wpmucdn.com/blogs.cornell.edu/dist/8/7752/files/2021/02/arxiv-logo-1.png
arXiv will spin out from Cornell University to become an independent nonprofit organization on July 1, 2026. This change will bring new flexibility and opportunities while maintaining arXiv's mission and values of free access to scientific ideas and discoveries.

TabFM: A zero-shot foundation model for tabular data

https://storage.googleapis.com/gweb-research2023-media/original_images/TabFM1_Hero.png
TabFM is a foundation model for tabular data classification and regression that uses in-context learning to eliminate manual model training and hyperparameter tuning. It achieves high-quality predictions in a single forward pass, outperforming traditional supervised algorithms.

Long Island's decommissioned nuclear power plant

https://i0.wp.com/nickcarr.com/wp-content/uploads/2014/02/IMG_5911.jpg?fit=1024%2C683&ssl=1
The Shoreham Nuclear Power Plant, closed since 1994, is a unique filming location with a massive reactor control room and various industrial spaces. The plant features a 7-foot thick outer containment wall and a primary containment area with a reactor and fuel rod assembly.

Have you restarted your computer this week?

https://cdn.uploads.micro.blog/96826/2026/hello-it-have-you-tried.gif
The user enjoys manually restarting their Mac on Saturdays, closing all windows and ignoring unsaved work warnings, for a clean desktop environment. This process allows them to disconnect from work and resolve potential computer issues.

Knoppix

https://www.knopper.net/pics/knoppix-logo.gif
KNOPPIX is a bootable Linux system on CD, DVD, or USB with automatic hardware detection and support for various devices. It can be used as a productive desktop system, educational tool, or rescue system without installing anything on a hard disk.