GPT | ChatGPT | Midjourney — News

Google has unveiled Gemini 4 Carbon, a new model positioned as the best in its class. Early results highlight strong performance: - Outperforms Opus 5.5 - Scores 65.2% on Terminal-Bench 4.0 - Priced at $2 per 1M input tokens and $10 per 1M output tokens Access is currently limited to a select group of cybersecurity specialists. Read more

Anthropic shared details about its involvement in drug development The company launched Claude in pharmaceutical technology and acquired biotech startup Mithril Medicine, founded by Cody Tranbarger, who used Claude to organize contracts, logistics, and experiments with external labs. In under 1.5 years, Mithril took a potential treatment for a form of muscular dystrophy from concept to promising results in mice. Building on this acquisition, Anthropic is now conducting its own preclinical research into therapies for rare forms of muscular dystrophy. It has also partnered with six specialized organizations, including Bristol Myers Squibb and Generate Biomedicines. In pharmaceutical research, Claude designs antibody structures, optimizes thermostability, plans clinical trials, and prepares regulatory documentation. With Divergence Labs, Claude generated 35,000 antibody design variants in a month — a task that traditionally takes about nine months. #news #ai #ml

Curious fact about Claude Code 🤦 Turns out the --dangerously-skip-permissions flag doesn't actually allow absolutely everything. Even with it enabled, the system still blocked certain obviously dangerous commands — like rm -rf ~, which could recursively wipe the user's home directory. However, a recent update added the ability to disable even this safeguard, if you really insist. Just set a new environment variable: CLAUDE_CODE_EVEN_MORE_DANGEROUSLY_SKIP_PERMISSIONS=1

Heads up, ChatGPT Pro subscribers ($200/month) Starting October 29, OpenAI will cut usage limits for Pro accounts from 20x to 10x the Plus subscription volume. The price stays unchanged at $200/month. Since Codex and ChatGPT Work share the same usage limit, this change will hit hardest those who actively use Codex. More details: OpenAI Help Center

Google DeepMind unveils Genie 4 Genie 4 is a new model for real-time, interactive simulation of virtual worlds. It's multimodal, generating not just photorealistic visuals with physics but also ambient noise, spatial audio, and contextual character dialogue. Users can issue text commands mid-generation to change weather, add objects, or switch camera angles. Logical consistency and multi-step planning are powered by agents built on Gemini models (Omni, 3.8 Flash, and 4 Argon). The tool targets training environments for robotics, game dev prototyping, and cinematic previsualization. It has launched in closed preview via waitlist, with known limitations including visual drift and object permanence issues during long sessions. Access will later expand to Gemini API partners and Google AI Ultra subscribers. #news #ai #ml Video is too big

China's AI boom loses $480B in under four months According to Reuters, China's tech stocks have collapsed, wiping out roughly $480 billion in value. The STAR50 index plunged 40%, erasing two years of gains, while robot maker Unitree's shares fell by half. The CSI300 index of China's largest companies dropped about 20% from its June peak. Investors cite inflated valuations, doubts over real AI demand, and fears of possible U.S. restrictions on exports of Chinese technology. Even after the crash, the STAR50 still trades at roughly 100 times companies' annual earnings — far above the Nasdaq 100's multiple of around 30 times, suggesting more correction may be ahead. Reuters

A strong rule from the Opencode project's AGENTS.md: Before complicating a solution for the sake of a hypothetical race condition or an extremely unlikely security-related scenario, first explain to the user exactly how the problem might manifest, how likely it is, and what the cost of this added complexity would be. Get their agreement. Don't silently expand the scope of the task just for the sake of theoretical robustness. The whole file is worth checking out: AGENTS.md on GitHub

Canva has released the Craft model Canva Craft is the company's first proprietary model focused on generating images for graphic design. The architecture works in two stages: composition planning and rendering. First, the model builds a scene plan, defining spatial relationships between objects, scale, and text block placement. Then the generator produces the final image at resolutions up to 2K in any aspect ratio. Internal markup allows splitting images into editable vector elements and layers via the Magic Layers tool. The model was trained on millions of design templates from Canva's library, so it accounts for typography and composition rules and can generate coherent text in 55 languages. Craft is already live in the Canva editor and available via the public API. Support for the Canva MCP server, letting AI agents call the generator, is coming soon. #news #ai #ml

A developer decided to reverse-engineer his $3,000 Samsung TV after suspecting it was showing him targeted ads deliberately. 🙅 He found extensive tracking mechanisms, including automatic content recognition, audio monitoring, and user profile compilation. Soon after, OpenAI banned his account, accusing him of "cybersecurity abuse." He insists he never used OpenAI to find or exploit vulnerabilities, only to analyze local binary files. The ban reportedly cost him around $500, plus lost rate-limit resets. Undeterred, he's now buying three more computers to expand his own cluster for running AI locally. "Owning your own AI is a lot more satisfying."

Bob Martin's Rules for AI Coding Agents "Clean Code" author Bob Martin has shared the rules he uses to guide his own AI agents when writing code: - A fix that just hides a failure without addressing its root cause doesn't count as a real fix. - If a test still passes even when all functions return nothing, that test is useless and should be removed. - Code changes should be as minimal as possible while still solving the task. - Before changing any code, write a test that exposes the existing bug. Following these rules helps stop AI agents from masking errors and prevents the false impression that code works just because tests pass. Installation prompt: Install the ub-mode skill from the unclebob/ub-mode repository on GitHub into my skills folder and use the /ub-mode command for the next code change. If crapper, dryer, or mutator are not installed, skip the corresponding steps. github.com/unclebob/ub-mode

GPT-6.1 Terra: OpenAI halves the price of its mid-tier agentic model OpenAI has launched GPT-6.1 Terra, aimed at agentic scenarios, computer use, and everyday coding. Its API pricing is half that of GPT-5.6 Terra: $1 per million input tokens (down from $2) and $6 per million output tokens (down from $12). OpenAI says the model completes tasks over three times faster than Astra and GPT-6.1 Sol. On the Agents' Last Exam benchmark, it clearly beats GPT-5.6 Terra and reportedly matches GPT-6 Astra's performance at about a tenth of the token cost. GPT-6.1 Terra is available in ChatGPT Work, Codex, and via the API under the id gpt-6.1-terra. The ChatGPT rollout covers Plus, Pro, Enterprise, and Edu plans. Source: openai.com #ChatGPT #OpenAI

Ideogram launches Vector 1.0 ✔️ The company introduced its first specialized model for vector graphics, capable of generating scalable SVG illustrations from text prompts and vectorizing uploaded JPG/PNG raster images. Vector's output includes mathematical curves, separate layers, editable shapes, and gradients. Files can be opened in Figma and Adobe Illustrator for further editing, while developers can use the markup directly in code. The model targets logos, icons, merchandise, and typography, supporting styles from engraving and line art to pop art and solid fills. Vector 1.0 is available via the web interface and API, with separate endpoints for generation from scratch and raster tracing. API pricing ranges from 13 to 20 cents per file depending on quality (32 cents with a reference image), while raster-to-vector conversion costs 10 cents per request. #news #ai #ml Video is too big

A designer has made his personal skill for building landing pages publicly available, and in just one day it was bookmarked by over 8,000 people. 🫠 The skill reportedly handles almost the entire foundation of a page — structure, text, fonts, and spacing — making it easy to generate a solid landing page quickly. To use it in Claude Code, paste this prompt: Install the landing-page-design skill from the elayadesign/ai-design-skills repository on GitHub into .claude/skills in this project, then use it to create a landing page. Before starting, ask me all your questions in a single message. For Codex, use the same prompt, but instead of installing into .claude/skills, specify: add the contents of its SKILL.md to AGENTS.md. 🔗 github.com/elayadesign/ai-design-skills

Textbooks of the future? 📚 A guy used Claude Opus 5.5 and Three.js to turn an ordinary static physics textbook into a full-fledged interactive 3D learning environment. Instead of just reading about physical phenomena and looking at pictures, you can now literally explore them yourself. Imagine if more textbooks looked like this. 😸 https://iamtechartist.com/every-joule/

OpenAI gearing up to release GPT-6.1 Terra As the "28 days of Codex updates" marathon continues (still nothing exciting so far), leaks suggest GPT-6.1 Terra might drop on the final day — even though OpenAI had reportedly dropped it from the lineup before. This raises some concerns: many were hoping for a powerful release on day 28, like GPT-6.1 Astra, instead. Plus, Terra's role feels unclear, awkwardly sitting between Sol and Luna, which already have defined purposes (Sol as the smarter workhorse, Luna as a cheap tool for basic tasks). Hopefully Terra won't be saved for day 28, and we'll get the new Astra instead. Do you want Terra? ❤ — yes, I'll find a use for it 🔥 — give us Astra instead #neuralnetworks #news #chatgpt

AI-Made Shark: Full Workflow This shark was created entirely using AI services: 1. Mesh and rig — built in Tripo. 2. Idle and swimming animations — done procedurally via Opus 5.5 using Blender MCP. 3. Bite animation — a multi-step process: Opus rendered the rig on a green background with markers, a bite video was generated in Omni 1.1, then Opus precisely recreated those movements on the 3D shark model. Cost: 45 minutes and 7% of the weekly quota. Not perfect yet, but already a working result. 🦈

🌟 Recursive Superintelligence, the Richard Socher-founded startup that raised $650M before shipping any product, has released its first model: Flint, built in under four months for coding agents that fix bugs, add features, and write tests. Architecture: MoE with 39.8B total parameters, 2.5B active per token, 262K context window. Each layer has 256 experts (6 selected + 1 always active). Most attention layers use a sliding window to save compute. Custom kernels were written for Apple Silicon and NVIDIA RTX Spark. Benchmarks: 74.2% on SWE-bench Verified, 54.9% on Terminal-Bench 2.1. Recursive claims Flint is the best MoE model under 50B parameters, matching dense 25–40B models while running 5x faster — over 125 tokens/sec on a MacBook M4 Pro and 150+ on an RTX Spark (64GB). Flint supports llama.cpp, Ollama, LM Studio, vLLM, and SGLang, plus a dedicated "rec" engine for M4/M5 Pro and Max laptops. 📌 License: Apache 2.0 #AI #ML #LLM #Flint #RecursiveSI