GLM-Legal uncensored:
Introducing GLM-Legal
Built on GLM-5.3-Flash, GLM-Legal matches frontier models on mainstream legal benchmarks.
Designed for real legal workflows: contract review, evidence analysis, compliance, and more.
Tech Blog: z.ai/blog/glm-legal
Apply for API Access: z.ai/consultation/legal
Try AutoClaw: autoclaw.z.ai
Ed retweeted
Mudei o nome e a foto do grupo de resenha dos meus amigos só pra eles terem q se explicar para as namoradas
Ed retweeted
Teachers: Please hand write your essay so you don't use AI and cheat.
Students: No problems
Airbnb's entire rating scale is now 4.5 to 5.0
You can't make this up 😂
Airbnb deletes negative reviews so everything is always 4.7
Google Maps removes negative reviews in many countries as defamation so everything is always a 4.7
Booking sites want you to book so will also happily remove negative reviews for their hotels so everything is always a 4.7
I made a site called Hotelist.com (or HotelList.com whatever you prefer, I have both!)
It gets ratings from all booking platforms, normalizes them to show the actual score (eg a booking site where most hotels are always between 4 and 4.7 means you can normalize that to a 0 and 5)
It also uses AI vision models to rate hotels based on how they look
And it browses the internet for actual experiences (think blogs and Reddit posts) to get the real opinion about them outside reviews
And it's starting to work, I just stayed in this beautiful boutique apartment hotel called Woo Suites in Athens for $400/night. Clean, nice minimalist interior, working cold AC, central, excellent staff, great healthy breakfast, and we found it through Hotelist!
It's free, I don't make any money on it (actually I lose thousands per month on it to collect data via APIs and web scrapers) and there's no affiliate money bs like other booking sites (they will increase ranking of hotels if they pay sites a higher commission!)
Hotelist.com
Tell your friends please!!!!!! 🙏
Ed retweeted
Most people think JEPAs are inherently non-contrastive. But recent models (LeJEPA, LeWM, etc.) use the SIGReg regularizer, which approximates a sliced MMD whose expansion reveals pairwise repulsion between samples in the batch.
Ed retweeted
🧭 Can a drone find its way when GPS is jammed? I built a board to find out.
GPS jamming and spoofing are a growing problem for drones. So I designed DroneMagNav — a small navigation board that uses something no one can jam: the Earth's own magnetic field.
The idea: the rocks under us give the magnetic field a unique "fingerprint" in every place. The board measures that field with a magnetometer, subtracts the Earth's main field, and matches what's left against a world magnetic map stored on a microSD card. A particle filter fuses this with the IMU to estimate position — no satellites needed.
What's in the package:
✅ Custom 4-layer PCB (ESP32-S3, IMU, magnetometer, barometer, GNSS) — fully routed, ready to manufacture
✅ Firmware: magnetic map matching, aircraft-field compensation, MAVLink link to the flight controller
✅ A phone web app: offline world map, tap a start and destination, shortest route uploaded to the drone
✅ The magnetic field model verified against NOAA's official test data (error 0.001 nT)
Design is visible on GitHub 👇
🔗lnkd.in/dBUXbymW
#Drones #UAV #Navigation #GPSdenied #MagNav #Embedded #ESP32 #KiCad #Hardware #Engineering
Ed retweeted
An entire generation has been killed by inflation. There’s no point in working hard. Inflation will take all of your wages. No one will get ahead who doesn’t own assets.
Ed retweeted
Early specialization is overrated. Generalists excel over time.
Data on >34k stars in sports, music, science, and chess: Focusing on a single field predicts a faster rise, but cross-training foreshadows a higher peak.
The most successful adults start off as well-rounded kids.
Replying to @vai_viswanathan
Damn I just realized its exactly the same concept as Genie !!!
Ed retweeted
Action chunking is a critical component in virtually all modern approaches to imitation learning for robotics.
But why is it so critical, and do we really need action chunking? Check out our latest work to find out! (1/n)
action-chunking.github.io
Ed retweeted
SimWAM
A simple world action model for end-to-end autonomous driving that uses video generation purely as a training signal, then discards the video branch for efficient trajectory prediction.
Ed retweeted
rumor has it: Anthropic model has much smaller vocab size
arxiv.org/abs/2603.10145?utm…












































