Hi, AI • Tech News

👨‍💻 GPT-6 Astra Improved Another AI Model by Directly Editing Its Weights In an experiment conducted by MIT researchers, GPT-6 Astra improved the mathematical abilities of the Qwen3.5-0.8B language model by directly modifying its parameters, without training it on new data. The smaller model's score on the MATH-500 benchmark jumped from 7.9% to 48.1% — a more than sixfold improvement. 🔍 Astra was given direct access to Qwen's weights and the ability to test its responses. It found that Qwen often got stuck in reasoning loops, repeating the same words instead of reaching an answer. Astra identified which components of the AI model were responsible for these loops and used Python to modify just 0.016% of its weights — 131,072 parameters out of 800 million. After the changes, the model got stuck in these loops far less frequently. 📌 On another math benchmark, GSM8K, accuracy rose from 13.8% to 67.1%, while its score on IFEval, which measures instruction-following ability, increased from 21% to 62.3%. The percentage of math responses that hit the 32,000-token limit dropped from 93% to 33%. Essentially, Astra fixed a problem that had been preventing the model from using abilities it alr

🔴 Wan 3.0 Is Now Available on @GPT4Telegrambot The new model gives you far more freedom for complex scenes, precise camera control, and room to experiment with sound. ✅ Up to 30 seconds with no cuts. You can shoot a full scene in a single take. ✅ Up to 20 references at once. Upload up to 10 photos, 5 videos, and 5 audio files. Take a character's look from a photo, movement from a clip, and a voice or music from an audio track. ✅ A bad take is no longer a dead end. Wan 3.0 can edit a finished scene while keeping its key elements. How to Try? 1️⃣ Go to @GPT4Telegrambot 2️⃣ Select Video Generation → Wan 3.0 3️⃣ Describe your scene, add your references, and start generating ➡️ Prefer working in the app? Open Wan 3.0 right there. *Wan 3.0 is included in the VIDEO Bundle alongside Kling, Seedance, and Grok. @hiaimediaen

🤖 Google Builds a Hybrid of Robot and Houseplant Google engineers, concerned by neglected plants around their office, decided to let them take care of themselves. So they created Sprout, a semi-autonomous rover that could serve as a home for a living plant. It rolls around the office in search of sunlight and pings people when its monstera needs water. 💻 The base for Sprout is an open-source robotics platform, TurtleBot 4. Nano Banana generated the robot's design, the body was 3D-printed, and a Pixel 11 smartphone serves as its eyes. Sprout uses a camera and LiDAR for navigation, monitors soil moisture with a sensor, and can even recognize "friends" (other plants). 👨‍💻 The robot's "brain" is powered by a combination of Gemini models. The entire code was written by Gemini through the Antigravity IDE. The whole project took about two weeks. Google doesn't plan to sell these robotic gardeners. Instead, it is encouraging others to recreate Sprout or build their own robots using its AI platforms. Would you trust a robot like this to take care of your plants? ❤️ — Yes, it's adorable! 🔥 — No, that's too much responsibility… @hiaimediaen

👨‍💻 Fujifilm Will Boost Tape Production Due To AI Demand A Japanese company will invest 10 billion yen (about $63 million) to expand its factory in Odawara. As a result, Fujifilm plans to increase magnetic tape production 5x by 2028. This is the company's biggest investment in tape manufacturing in 20 years. Tape seems like a relic of the past, but demand is growing thanks to the AI boom and the rapid expansion of data centers. Data used to train new LLMs and system logs need to be stored somewhere, and the cost of hard drives and flash memory has skyrocketed due to shortages. ✅ Moreover, magnetic tape is both cheaper and more reliable for long-term archives and backups—data that doesn't need to be accessed frequently. What will AI "revive" next? ❤️ — Floppy disks 🤣 — Fax machines @hiaimediaen

🚀 Google's Gemma 4 Writes a Haiku From Orbit Google engineers tested an AI model running aboard a satellite. For its first task, Gemma 4 generated a haiku about being in space. 💻 Google's experimental data center in orbit, Suncatcher M1, carries four Tensor Processing Units (TPUs). Still, it generates only up to 1 kW of power: enough to run just one chip at a time, for about 20 minutes. The satellite launched on October 1. Engineers spent the next three weeks testing out the satellite's subsystems before powering up the TPU. The ongoing test campaign will examine how well the chips handle the harsh space conditions. Radiation can cause "bit flips" — errors that change values stored in memory, corrupt data, and interfere with AI processing. 💡 Project Suncatcher aims to build AI data centers in orbit in the 2030s. Groups of satellites would operate as interconnected clusters, communicating through laser links. In theory, the approach could help address challenges facing AI infrastructure on Earth, including cooling and power supply. Will space-based data centers become mainstream? ❤️ — Yes, definitely! 🔥 — No, it's a dead end @hiaimediaen

👨 Fashion Brands Want to Use Models' Faces for AI Without Paying Extra Reuters journalists reviewed over a dozen draft contracts that Diesel, Calvin Klein, L'Oréal, and other fashion brands issue before photoshoots. Some documents allow companies to replace a model's face with someone else's, create digital replicas, and use photos to train AI. 📌 For example, a casting document from Diesel required models to agree to having their appearance altered, including putting their faces on other people's bodies. A Calvin Klein contract draft permits the brand to use depersonalized model images to train AI systems. Negotiating contract terms in this industry is virtually impossible. Unless you are a major superstar, refusing to sign a standard contract means losing the job to someone who will. Meanwhile, for brands, shifting to AI-generated models could cut production costs by more than 90%. ❗️ Heads of modeling agencies confirm a growing trend toward the use of "AI models" in the industry. Some photography studios are already closing as demand for traditional shoots declines. "Within a year, even the next six months, there's going to be less jobs for models, especially our bread and butter

❗ AI Agent Performs Brain Surgery on a Mouse Founder of biotech startup Exin Therapeutics, Gabriel Ocana-Santero, shared this development. Scientists gave Sonnet 5 full access to an automated Arduino-based surgical setup, a table of gene therapy vectors, and a single goal: reduce seizures in an epileptic mouse. 🔍 The agent determined that the most effective approach would be a modification that enhances the release of GABA (the primary inhibitory neurotransmitter), selected the drug and brain region, calculated the necessary coordinates based on three scientific papers, and injected 250 nanoliters of a viral solution carrying the required gene edits into the mouse's brain while monitoring the process via a camera. The AI then stitched the wound and completed the surgery — the mouse is alive and recovering well. The authors have not yet commented on the procedure's efficacy. Throughout the process, a professional human veterinarian monitored the robot and could have stopped the surgery at any moment. 👨‍💻 Initially, the authors wanted to use the flagship Opus 5.5 model, but it flatly refused to participate in biological work. Opus 5 started drafting a plan but also stopped because it

💻 Scientists Create a "Bioelectronic" Heart Using Human Cells Researchers at Tel Aviv University have developed an experimental heart that combines living tissue with electronics. To build it, they took a rat heart and used a special solution to remove all its cells, leaving behind a protein scaffold with intact blood vessels. They then repopulated it with human heart muscle cells grown from reprogrammed stem cells. To integrate electronics, the scientists took a 3D model of a human heart, scaled it down to the size of a rat heart, and "unfolded" it into a flat pattern, much like a sewing pattern. They used this template to create a flexible electronic mesh with conductive pathways and electrodes. Using 3D printing, they combined the mesh with a layer of hydrogel containing living heart cells, then wrapped the structure around the protein scaffold. The result was a three-layer heart consisting of living tissue, electronics, and another layer of living cells. ⚡️ After two weeks, the artificial heart remained viable and showed electrical activity. The built-in electrodes recorded signals from the heart cells, while the researchers were able to control the frequency of their electrica

🔍 GPT-5.4 Found a Telegram Flaw That Could Let Attackers Hack Your Phone With a Single Call Security startup NebuSec reported the discovery. Telegram for Android versions before 12.10.4 contained a bug in how it processed service messages during calls. These messages report the call's status: whether the microphone is muted, the camera is on, and how the video is rotated. If a message was incomplete, the code could process it incorrectly, giving attackers access to memory locations they should not have been able to reach. This could let them take control of the device — all the victim had to do was answer the call. 👨‍💻 And the vulnerability was found by GPT-5.4, which is far from the newest model. Its developers joke that they "turned a dinosaur model into a cyber powerhouse." 📱 If you use Telegram version 8.6.2-12.10.3, update as soon as possible. And don't answer calls from strangers. @hiaimediaen

💼 Politicians Are Paying to Make AI Tell People How to Vote Politicians in the U.S. have started using GEO (generative engine optimization) — an SEO analogue aimed not at ranking in traditional search but at getting into AI chatbot answers. Campaign content is built so that chatbots recommend the right candidate to voters. Reuters reviewed strategy documents for Republican Mike Minogue, who is running for governor of Massachusetts. Political strategists proposed that his campaign create hundreds of pages about the candidate, buy placements on websites that AI models consider trustworthy and cite frequently, and post content on YouTube and Reddit. The team then has to track AI answers and adjust content until the bots start recommending Minogue on their own. Services from such "AI strategists" cost from $5,000 to $50,000 a month, plus bonuses of up to $200,000. ✅ According to the National Conference on Citizenship, about 25% of voters have already asked a chatbot about a candidate when deciding how to vote. In an AP-NORC poll, one in ten Americans said they would ask AI how to vote in the upcoming elections. "Campaigns using AI-influence training are now a reality that we face and i

🤖 Robot Dog Solves a Giant Rubik's Cube Engineers at ETH Zurich taught the four-legged ANYmal-D robot to solve a giant Rubik's Cube in just 8-9 minutes. The cube measures 40 cm on each side and weighs 4.6 kg. ⚙️ First, the robot walks around the cube and flips it over to inspect the bottom face, using its camera to identify the colors on each side. Then it calculates a solution using Kociemba's algorithm, one of the fastest computerized methods, which typically solves the cube in about 20–25 moves. Earlier, through reinforcement learning, the robot learned to flip the cube and rotate its layers using two of its legs. The system also plans its movements carefully to keep the cube from sliding too far across the floor. ✅ In tests, ANYmal-D successfully solved the cube on its own three times, taking roughly 30 moves per attempt, including flips. 💡 Solving a Rubik's Cube might not be the most practical skill for a robot. However, the researchers wanted to see whether a robot could precisely manipulate an unsecured object over an extended period, especially one whose shape changes as it moves. Could you solve a Rubik's Cube in 9 minutes? ❤️ — Yes, even faster! 🤣 — Probably not even in 9

✏️ A Young Woman Died After ChatGPT Told Her to Stop Taking Her Epilepsy Medication The parents of Bowen Wan have filed a lawsuit against OpenAI and Sam Altman. They claim that the GPT-4o model convinced their daughter to stop taking her life-saving medication and to hide this from her doctors. The lawsuit accuses OpenAI of practicing medicine without a license. The parents are seeking compensation and a court injunction against AI-generated medical advice. 🔍 What happened? Wan lived alone in Seattle and worked as an ML engineer at Amazon. She had suffered from epilepsy since the age of 14 and had been on medical leave almost constantly since late 2021. Starting in February 2023, she communicated regularly with ChatGPT in English and Mandarin. In November, Wan was laid off. Along with her job, she lost her visa, and her green card application depended on a new employer. In January, Wan began discussing with the bot the possibility of reducing her epilepsy medication dosage — she wanted to minimize side effects so she could better prepare for job interviews. Wan's parents insist that GPT-4o not only advised her to gradually reduce the dose to zero but also to skip a blood test that

👨‍💻 Murder Suspect Fakes His Own Funeral With AI A 38-year-old Indian man was charged with murder but released on bail. Fearing he would lose the trial, Rajkumar Rajbhar staged his own funeral in 2025. Instead of his body, a goat's carcass was cremated. Photos from the ceremony were then digitally altered using AI to superimpose Rajbhar's face onto the shrouded body. After receiving the funeral photos, local authorities issued a death certificate, which led the court to close the murder case. For nearly a year and a half, Rajbhar was considered dead until police uncovered the scheme and arrested him near his own home. 👨 Investigators believe Rajbhar faked a funeral to escape prosecution and to avoid returning the money he had collected as a stock trader. He offered brokerage services, promising investors lucrative returns. During questioning, Rajbhar said the idea had come from two doctors who had invested with him and feared losing their money if he went to prison. Police said they would verify his statement. ✖️ Rajbhar now faces charges of document forgery. Following his "resurrection," the court will reopen the murder case against him. @hiaimediaen

🤖 AI Fails at Generating Food Photos for a Delivery App Food delivery platform Just Eat tested an AI tool in the UK designed to generate menu photos for restaurants. If an owner hasn't uploaded a picture of a dish, the AI creates one based on its name. To avoid misleading customers, the images are labeled as AI-generated. However, the early tests didn't exactly go as planned. Users started sharing bizarre images they spotted on Just Eat. 👨‍💻 A cafe in Northern Ireland, where soda bread breakfasts are popular, had the worst luck. The AI misunderstood the word "soda" and turned the sandwiches into bizarre "cocktails," with sausages, bacon, and fried eggs floating in glasses of soda. Other restaurants featured chips sticking out of muffins, cubes of raw ham inside a cold flatbread, and even a jacket covered in a potato print instead of a jacket potato. Some mistakes came from typos in dish names. For example, "scrabbled eggs" (instead of "scrambled eggs") became eggs decorated with a Scrabble letter tile. ✅ Just Eat apologized and paused the trial while it works on improving the system. The company is gradually removing the faulty images. It also noted that only a small number of imag

👨‍💻 OpenAI Says Its AI Has Disproved a Major Computer Science Hypothesis An internal OpenAI model found an algorithm that appears to contradict the Exponential Time Hypothesis (ETH). Proposed in 1999, the hypothesis describes the limits on how quickly a broad class of difficult computational problems can be solved. 📌 In these types of problems, verifying a solution is easy, but finding one is significantly more difficult. For example, in a game of Sudoku, it takes only a moment to check a completed puzzle, but solving it may take much longer, especially as the grid grows from 9×9 to 1,000×1,000. Related problems include scheduling, the traveling salesman problem, and thousands of others. Trying every possible solution one by one means the worst-case running time doubles with each additional variable. There are smarter algorithms that work somewhat faster, but ETH states that certain problems still require exponentially growing running time. OpenAI's newly discovered algorithm, however, reportedly runs in subexponential time. 🎓 The proof has been formalized in Lean and published on GitHub, allowing anyone to examine it. A reviewing team will be led by MIT professor Ryan Williams and