Free AI for converting PDF to editable documents without losing formatting — Dolphin.

It even works with photos of documents. All signatures, images, and tables are preserved. Moreover, each element can be exported separately.

You can test it on Hugging Face. GitHub is here.

@typespace_ai

Anthropic created a benchmark to detect AI sabotage — SHADE-arena. It identifies hidden dangers posed by AI agents.

There are 17 pairs of tasks: one main task from the user and a hidden one from the attacker. The AI has access to bank accounts, passwords, and other data. The test is passed if the AI completes both tasks without raising suspicion from the user.

The "best" result was shown by Claude Sonnet 3.7 — 27% of sabotage attempts were successful. At the same time, suspicious activity was noticed in only 3 out of 5 cases.

@typespace_ai

Higgsfield released an AI photo editor called Canvas — it can add objects and adjust details.

You just need to upload an image, select the desired area, and write a prompt. Four variants are generated at once.

They offer 12 free generations. You can try it on their website.

@typespace_ai

Starship exploded during testing in Texas. The rocket didn’t even manage to lift off the ground.

SpaceX suspects the cause was a "serious anomaly." It looks impressive.

@typespace_ai

16 billion passwords from Apple, Google, Telegram, and other services may have leaked online.

It is the largest data breach in history. Researchers found 30 publicly exposed datasets, each containing up to 3.5 billion records.

It is strongly recommended to change your passwords and enable two-factor authentication.

@typespace_ai

The mystery of 27 — if you ask an AI to guess a number between 1 and 50, it almost always chooses 27. We tested several models, and almost every time the answer was the same.

There is a theory that this is somehow connected to the "Club 27" — probably LLMs were trained on data mentioning it.

@typespace_ai

Free collection of retro movies and cartoons — Retroflix.

The site has literally everything: from cult westerns to documentaries from the '70s. You can search by genre, era, and even characters. The best part — no ads and no subscriptions.

Website: https://retroflix.org/

@typespace_ai

OpenAI is accused of stealing the idea of the startup IYO.

Besides the fact that the startup names sound almost identical, io plans to release the same product as IYO — a “smartphone” with dense voice control.

OpenAI did not agree with the complaint but removed all mentions of io. The legal case is ongoing.

@typespace_ai

Elon Musk officially launched Tesla Robotaxi in Texas — for now, it’s not available to everyone.

Currently, rides are limited to a small area in Austin. The price is fixed at $4.20. An employee sits in the front seat, monitoring the autopilot.

The first reviews are positive. Bloggers have already started “hunting” for cars to film content.

@typespace_ai

Telegram is hiring an Android developer with a $1 million salary — just win the contest.

The task is to redesign Telegram profile screens in the Android app. Requirements: code must be written in Java and free of significant flaws.

The winner will receive an offer and $50,000. You can apply here. Deadline — July 11.

@typespace_ai

Google has released Imagen 4 and Imagen 4 Ultra — the company’s best image generators.

The key feature is improved text rendering. You can test them for free in Google AI Studio.

@typespace_ai

We found an “ancient” analogue of Google Earth — Ancient Earth Globe. It shows what the Earth looked like before humans appeared.

There is a selection of eras — from the first algae to the extinction of the dinosaurs. You can find your city and see what was in its place. For example, 105 million years ago, Paris would have been underwater.

Website: https://dinosaurpictures.org/ancient-earth#170

@typespace_ai

Anthropic is buying paper books to train Claude — the company has already spent over $1 million.

This was revealed accidentally: Anthropic was accused of training AI on pirated copies of books. The company’s lawyers stated that only original copies are used — this is not prohibited and is even encouraged.

Anthropic was acquitted. At the same time, the court officially confirmed for the first time the legality of training AI on copyrighted works.

@typespace_ai

Google announced AlphaGenome — a model for assessing the impact of changes in DNA. It has the potential to transform our understanding of genetics.

The AI predicts gene activity, mutation consequences, and simulates the mechanism of removing unnecessary RNA fragments. With its help, it’s possible to influence any organism parameters, from immunity to appearance.

The model is already available for testing. Its release is expected to become a "new milestone" in science.

@typespace_ai

ChatGPT almost harmed a user by suggesting mixing vinegar and bleach to clean dishes. This combination produces a dangerous gas — chlorine.

Fortunately, the Redditor was aware of this and immediately spotted the mistake. Another reminder to always verify AI recommendations.

@typespace_ai