Robot Cat

A new open-source AI called CarDiag uses audio to detect potential car problems. It’s still early days. Right now, the AI correctly distinguishes between healthy and faulty engines about 79% of the time. But here’s the impressive part: the entire trained model is only around 100 KB, making it incredibly lightweight and easy to run. Because the project is fully open source, the creator hopes developers and car enthusiasts will help train it into a much more accurate mechanic that fits in your pocket. Source. #OpenSource #Tool #Automotive

Someone showed me this on Telegram. It is very silly. It is clearly masquerading as "Free GPT and Claude". Anyone with half a brain knows this is malicious, but people will still fall for it. People asked what it is. I have some free time. I poked it with a stick, People discussing it said it is XMRig. That is not entirely accurate. This is not XMRig. This is flagged as XMRig from Triage and VirusTotal because it does indeed drop XMRig, but it is much more than that. This is a (maybe new) information stealer packaged with XMRig as a double whammy. This malware is interesting because of a few things: 1. It is position independent, they care enough to be evasive and strip out a majority of dependencies. This is usually indicative of more serious malware. 2. They .zip it delivers from the "Free GPT and Claude" is intentionally bloated (payload inflation). It is 97MB, which may evade a majority of anti-malware product (initially) due to it's large size. It packages itself with FFMpeg and various other audio codecs. 3. It accesses Microsoft Outlook e-mails, accesses Chrome stuff using the COM IElevationService, looks for any SFTP credentials It (currently) does not have any matching YARA rules from AV vendors. The closest approximation is LummaStealer. My knowledge base on the Information Stealer scene is out-of-date (it changes a lot). However, on first initial glance this appears like a new information stealer. Again, this should be taken with a grain of salt. It's also worth noting the domain it exfiltrates to does not appear in any malware reports. The domain is unique, and the payload does not match any existing YARA rules (it's behavioral characteristics do, but not a specific malware family), so this is actually a pretty interesting sample. A lookup though shows this is an emerging malware campaign. It first appeared around the end of May. This is (probably) a known Threat Actor who has switched it up a bit (or it's MaaS, whatever though). The malware appears online masquerading as various products. - ecore-sourceproject - LogiDA - GPT_Claude_Free - CortexSystems.v3.4.2.Stable - TikTokBot-v2.2 - CortexLauncher Funny enough, this malware would have been much, much, much, MUCH more evasive if they didn't package it with XMRig. VirusTotal and Triage immediately flagged it because after it establishes persistence, and steals any credentials on the machine, it pulls XMRig to turn into a cryptocurrency miner. If they did not pull the XMRig binary this stealer would be much more quiet. I have no idea why they decided to burn their OPSEC with XMRig. C2: dfwioeiofwr-dot-info Payload (and associated families from the C2) 027d576c6b5512d661081aaeeeb8e611f95a469ccf5ba35e0a390e8814334d05 5dcc599cf48227e65ea49d2708d08704fd1cb7e3b89736718d0d8e557857c49c 5e8b40b0b7512e1a1355374fb0cf34bfdf1260ebdb80a353c8f9da2490beeed3 6a0c332296b017220fc2b522da653fce36a8a3c5c79de0200d61c5fc31eb89ce a2f8ebf65d54a4d9c8b720d01da77ad796683f1a5b8bd3d08738d7df4365f8a 9d4aaa9842c947756b7c128c432292732098fb71d247ef0bce60368563572da3 c4caca93e2291c018e701c217b7d232c534e4dd142042a59aa4d32754ef3022a

A journalist noticed that Gmail's AI suggested she "improve" the sentence "Israel's genocide in Gaza" by replacing "genocide" with the word "actions". An article from September 2025 revealed that the Israeli government paid a U.S. company, Clock Tower X LLC, to, among other things, create new websites to influence how AI GPT models such as ChatGPT, which are trained on vast amounts of data from every corner of the internet, frame topics and respond to them — all on behalf of Israel. As part of this work, the firm will also use search engine optimization software MarketBrew AI, a predictive AI platform that helps clients adapt to algorithms and promote their work on search engines like Google and Bing, to “improve the visibility and ranking of relevant narratives.”

Korean company ROBOTIS has unveiled the 'AI Sapiens', a direct challenge to Unitree The creepy 4'3" (130 cm) headless gangnam-style-dancing humanoid boats fully open-source hardware with 23 DOF, powered by the brand-new Dynamixel-Q actuators. Price point expected in the mid-to-low 10 million KRW range (~$7,000 – $8,700 USD). #OpenSource #Robot #Robotis

DeepSeek released a 1.6T AI model built for coding agents DeepSeek unveiled its V4 generation with two new foundation models: DeepSeek-V4-Pro (1.6 trillion parameters, activating only 49B per token) and DeepSeek-V4-Flash (284B parameters focused on efficiency). The release pushes open-source AI closer to the frontier dominated by closed labs. What makes V4 different: • Introduces DeepSeek Sparse Attention (DSA), a new architecture designed to cut memory and compute costs. • Uses token-wise compression, allowing far longer context windows without the usual infrastructure burden. • Result: 1 million token context becomes standard across official DeepSeek services. Built for coding and reasoning: • V4-Pro is positioned as a new open-source leader in agentic coding and advanced reasoning. • Designed to compete directly with top proprietary models from OpenAI, Anthropic, and Google. • Supports Thinking mode for deeper step-by-step reasoning and Non-Thinking mode for faster responses. DeepSeek’s strategy is clear: don’t just match frontier labs, make frontier AI cheaper, open, and easier to deploy. Source. #OpenWeight #MoE #Deepseek #PressRelease

Researchers from the Korean institute KAIST AI have released a cool model called EgoX based on Wan2.1-I2V-14B — you upload your video and specify from whose perspective you want to view it. The output is something similar to a body camera recording. #VideoGen #OpenWeight #Wan #PoV

Me and Claude vibe coding an app at 3AM

Google DeepMind launched Gemini 3.1 Flash TTS, a text-to-speech model for precise audio control. Its "Audio Tags" feature lets users set vocal style, emotion, and pacing via text. Supporting over 70 languages with native SynthID watermarking, the model is available in preview on the Gemini API and Google AI Studio, rolling out to enterprises on Vertex AI, and integrating into Google Vids for consumers. Source. #AudioGen #TTS #PressRelease

A poll of people in the US, and how they feel about AI. While critical, I am curious, excited, and optimistic. I'm building shit, having fun, advancing my skills, and meeting others like me. I think I'll do alright. #USA

C-level executives are increasingly advocating for a reduction in working hours. Sam Altman, with his policy paper, suggested that a 4-day week with 32 hours is practically indispensable and that we need a new social contract. There are growing voices suggesting that, in the wake of AI and robotics, people should work less for higher pay. Source. #Jobs #Labor

This one's legit this time I'm pretty sure. A phone number-stealing backdoor was found in Nekogram The malicious code is concealed within Extra.java (notably, the public repo version of this file is clean). It operates by using obfuscated code to send data through an inline query to @nekonotificationbot, leaving almost zero footprint. Additionally, if several accounts are connected within the app, the developer can easily see that they belong to the same person. The Google Play version also has this malicious code Sources: 1. Nekogram 12.5.2 2. Backdoor investigation post 3. Screenshot by @andreyduhen #Hack #Privacy #PSA #Telegram

Anthropic has deleted an official blog post announcing Claude Mythos "By far the most powerful AI model we've ever developed." "A new tier that crushes Claude Opus 4.6 in software coding, academic reasoning… and especially cybersecurity." Their own words: Mythos is so far ahead in discovering vulnerabilities and enabling exploits that it foreshadows an “impending wave” of AI-driven cyberattacks defenders can’t match.

Andromeda Robotics brings Abi, an emotional companion robot, to the US after 2 years in Australian care homes. • Talks, plays music, runs games, leads group sessions • Remembers residents and builds ongoing relationships • Helps offset care home staffing shortages • Supports 90 languages • 3,500+ companionship hours logged Their mission is to replace 1 billion hours of loneliness with meaningful companionship.

A Kentucky family has declined a $26 million offer to sell their farmland for an AI data center. #Datacenter #Agriculture #FightBack #USA

Bernie Sanders spoke to Anthropic’s AI agent Claude about AI collecting massive amounts of personal data and how that information is being used to violate our privacy rights. What an AI agent says about the dangers of AI is shocking and should wake us up.

Kimi AI introduces Attention Residuals, a new way to rethink how neural networks use past layers Researchers at Moonshot AI just proposed a new architecture tweak that could make large AI models more efficient and smarter about how they use information from earlier layers. Instead of the traditional residual connections used in deep networks, they introduce Attention Residuals, a system where each layer can selectively attend to representations from previous layers. Here’s what’s new: Attention over past layers: • Traditional residuals simply add outputs from earlier layers in a fixed way. • Attention Residuals let the model dynamically choose which earlier layers matter for a given input. Solves depth dilution: • In very deep models, useful information from earlier layers can get diluted. • Attention-based retrieval allows the network to pull specific past representations when needed. Block AttnRes for scale: • Layers are grouped into compressed blocks so cross-layer attention remains computationally practical. Efficient in practice: • Reported 1.25× compute advantage • <2% extra inference latency, meaning almost no slowdown. Tested on the Kimi Linear model: • Evaluated on 48B parameter architecture (3B activated parameters). • Shows consistent downstream performance improvements. Source. #ML #ANN #Paper #Moonshot #Kimi

Try not to scream if humanoid robots sprint toward you through the morning fog on the streets of Beijing. They are training for their half-marathon! Over 20 teams joined the first trial run. The official race will be held on April 19. Officially suggested practice times are midnight through til morning. #Robotics #Sports #China

Andrej Karpathy built a map of every job in America and how likely AI is to replace it The dataset: • Scraped 342 occupations from the U.S. Bureau of Labor Statistics • Each job was evaluated using an LLM scoring rubric from 0–10 • Built an interactive treemap visualization • Rectangle size = number of workers in that job • Color = how vulnerable the role is to AI The key rule behind the scoring If the work product is digital and the job can be done entirely from a home office, the exposure score rises dramatically. Some example scores: • 0–1: Roofers, janitors • 4–5: Nurses, retail workers, physicians • 8–9: Software developers, paralegals, data analysts • 10: Medical transcriptionists Overall result: Average exposure across all occupations: 5.3 / 10 Fully open source, Karpathy also released the entire pipeline: • BLS data scraping • LLM scoring methodology • The visualization system https://github.com/karpathy/jobs #Jobs #OpenSource #Tool