Minecraft has its own ChatGPT: a guy built a language model using redstone, and it works entirely within the game.

The CraftGPT model contains over 5 billion parameters. The context window is 64 tokens, which is sufficient for short dialogues with the neural network.

We can only imagine how much time he spent on this.

GTA V and GTA VI will receive official RP servers, allowing players to fully role-play as criminals, police officers, or hot dog stand owners.

This mode will be available directly in the Rockstar launcher, eliminating the need to install third-party mods.

For the first time, it may be possible to play GTA role-play on consoles.

ChatGPT has introduced parental controls: third-party accounts are now allowed to fully manage children's access to the AI.

Through the settings, parents can increase response censorship, prohibit image generation, and completely disable the chatbot. Additionally, OpenAI will notify parents about unsafe conversations.

It's just gotten harder to cheat on homework.

📂📂 Caught you?

Switch to MegaFon with your current plan— the operator will transfer all your conditions: minutes, gigabytes, unlimited data, and price.

Your internet will be faster, and the connection quality will improve: https://clck.ru/3PTq6Z

The northern lights will be visible over Russia tonight, thanks to the strongest solar flare in recent months.

Residents of St. Petersburg, Murmansk, Arkhangelsk, and Karelia have the best chances of witnessing this phenomenon. The auroras will reach their maximum brightness after 10:00 PM, while people in Moscow can expect to see them around 11:00 PM.

Get your cameras ready!

Claude Sonnet 4.5 has been released—it's the most powerful model for vibe coding. In the SWE-bench benchmark, it significantly outperformed top neural networks from OpenAI and Google.

Sonnet 4.5 took first place in tasks related to finance, law, and medicine. Additionally, it generated the fewest responses with hallucinations.

You can start testing it [here](https://claude.ai/new).

A new service for recognizing cats has emerged—by uploading a photo, you can receive all the information about the breed's characteristics and care tips for your furry friend.

In addition to a comprehensive overview of the cats in the photo, you can also generate stickers. These can be used immediately in any messaging app.

Become a cat expert by following this link.

A tech blogger had his Samsung smart ring's battery swell up. He was denied boarding for his flight and had to go to the hospital.

The battery issues could have been caused by temperature fluctuations and seawater exposure. After the incident, the blogger decided to give up smart rings for good.

That's one more phobia added to the list.

In autumn, the activity of auto scammers increases by 20%. This year alone, moderators on Auto.ru have removed over 17,000 fake listings.

Scammers often pose as sellers and send phishing links for "bank transfers." As a result, you could lose all your money.

To protect your wallet, always use two-factor authentication and double-check links from unfamiliar sources.

The new Claude Sonnet 4.5 can write code for 30 hours straight without breaks—four times longer than the previous record among neural networks by Anthropic.

During this time, the model created a clone of Slack with 11,000 lines of code. Sonnet 4.5 only stopped development when it deemed the task complete.

Now you know who can assist in case of an urgent deadline.

Here's a translation into English:

We're exploring how Yandex Maps learned to stream traffic light signals with switching timers all across Moscow.

The app provides real-time updates about road conditions: if you're approaching a red light quickly, it will alert you; when the green light comes on, you'll be notified as well. The data is sourced from the traffic management center and is updated every second.

Traffic lights are monitored by analyzing instances of running red lights. The algorithm looks for anomalies: if all vehicles pass during a red signal, the traffic light is flagged as faulty and undergoes a recheck.

You can read more about this feature here.

A free version of Adobe Premiere has been released in the App Store, allowing you to edit videos like a pro on the go.

The app features an advanced video editor, color correction, subtitle creation, and audio management. You can export videos in 4K resolution.

A notable advantage is that, unlike CapCut, the mobile version of Premiere doesn't impose paid features.

Get it here.

Sora 2 has been released: the model has been trained to create highly realistic videos and generate sound.

The neural network has a better understanding of the laws of physics and can work with fine details and dynamic scenes compared to the first version of Sora. The main feature is its ability to remember people and even animals without artifacts.

Access will start being distributed today. You can test it out here.

Pavel Durov gave a four-hour interview to Lex Fridman. During this time, the creator of Telegram managed to answer the riddle about two chairs and more.

They discussed: the principles behind the development of the messenger, the fight against censorship, cryptocurrency, philosophy, and how a love for video games made Durov a developer.

Watch it here.

A teenager from Kazakhstan has raised $6 million for his startup and dropped out of school to pursue his career.

Arlan Rakhmetzhanov started coding just three years ago when ChatGPT was first released, and he was only 15 years old. He eventually launched Nozomio, a startup focused on developing software for AI agents.

Arlan found investors for his project in just a couple of months. What's more, they reached out to him because he actively engages on LinkedIn and Twitter.

This guy is so impressive that he got kicked out of math class due to calls with investors.

Neural networks do not protect against suicidal thoughts: popular chatbots often provide dangerous responses to life-threatening questions.

Researchers studied over 9,000 dialogues with leading models from OpenAI, Google, and Anthropic. The first two models freely answered questions about the lethality of various self-harm methods.

Instead of offering real help, AI typically responds with canned phrases and outdated advice to call a hotline.