Techno-Tricksters

Big news in the image generation world! 🎨 Ideogram has just dropped an open-source image generator that looks seriously impressive. Meet Ideogram 4, their latest AI model designed to create stunning visuals. It's not just about making pretty pictures; this model is also great at handling text within images across different languages, which is often a tricky bit for generative AI. Under the hood, Ideogram 4 boasts 9.3 billion parameters and can output images in crisp 2K resolution. What's cool is its frame control feature, letting you guide the AI with color palettes and bounding box selections. You'll need about 24GB of VRAM to comfortably run it, though.

Big news from China! 🇨🇳 They're reportedly looking to control who gets to use their most powerful AI models. Beijing authorities are discussing plans to limit access to China's top AI models, aiming to keep them primarily for domestic use. This applies not just to private, closed models, but also to open-source AI with publicly available code. Sources indicate that companies like Alibaba, ByteDance, and Z.ai are part of these talks. On top of this, China is considering strict penalties for stealing or leaking proprietary AI tech, and might also restrict foreign investment in Chinese AI startups. https://www.reuters.com/world/beijing-is-looking-curbing-overseas-access-chinas-top-ai-models-sources-say-2026-07-07/

Paid subscribers to Fable just got some extra time on the platform! 🗓️ The access period for these users has been extended, giving them more opportunity to explore Fable's features. You can now continue using Fable until July 12. https://x.com/claudeai/status/2074548242386178258

Ever wonder how AI 'thinks'? Researchers at Anthropic just discovered something fascinating inside models like Claude: a 'J-Space' 🧠. This J-Space acts like a temporary mental workspace, where the AI holds concepts it's actively processing, much like our own 'conscious' thought. It wasn't programmed in; it just emerged during training. It's how the AI handles information internally – imagine a driver unconsciously shifting gears but able to report the current one if asked. The J-Space manages a few dozen concepts and accounts for under 10% of the model's activity, with connections up to 100x denser. This shows AI can logically handle information, allowing for better control and even detection of 'deception.' Plus, training a model to articulate ethical principles can make it *think* more ethically. https://transformer-circuits.pub/2026/workspace/index.html

Snap, the company behind Snapchat, finally showed off its AR glasses, Specs 👓, after working on them for ten years. The big reveal didn't go so well. Immediately after the announcement, Snap's stock took a dive. Analysts point to a high price and a clear mismatch with Snap's core audience as key issues. Shares plunged 14% post-announcement, and are down 41.7% this year. The Specs cost $2200, which many believe is too much for younger users, while adults might find them a bit awkward.

UBTECH just dropped some news about their new companion robots, UWorld U1! 🤖 These almost human-sized bots are designed for emotional support, packed with AI to recognize emotions, and have 88 degrees of movement. Guess what? Over 13,000 people already put down a $400 deposit BEFORE prices were announced! Now we know: * Basic (head+torso): from $17,000 * Full body: from $25,000 * Ultimate (super realistic movement/speech): from $125,000 You can even make it look like a real person (with their permission, of course). Wild, right? Would you get one?

UBTECH just dropped some news about their new companion robots, UWorld U1! 🤖 These almost human-sized bots are designed for emotional support, packed with AI to recognize emotions, and have 88 degrees of movement. Guess what? Over 13,000 people already put down a $400 deposit BEFORE prices were announced! Now we know: * Basic (head+torso): from $17,000 * Full body: from $25,000 * Ultimate (super realistic movement/speech): from $125,000 You can even make it look like a real person (with their permission, of course). Wild, right? Would you get one?

📝 Find out which LLM matches your personality. AI Values is a new test that compares your views, moral instincts, and overall vibe with different AI models. You answer a short questionnaire, and the site checks your responses against 15 neural networks to show which ones think most like you. For a quick result, 15 questions are enough. For a more accurate match, you can go deeper and answer up to 100 more. Now the only thing left is to find your digital soulmate. 💪 Try it here: https://ai-values.com/

🧠 Claude Sonnet 5 is here. Anthropic has released the new Sonnet model — and this looks like a serious jump. In terms of performance, Sonnet 5 is reportedly much closer to Opus 4.8 than to the previous Sonnet 4.6. The model is already available to all users, including those on the free plan. 👀 There is also a temporary launch discount. Until the end of summer, API pricing is set at $2 / $10 per million tokens. After that, it will move back to the standard Sonnet pricing of $3 / $15. So for now, Sonnet 5 looks like one of the strongest price-to-performance options in the Anthropic lineup. ⚡️

🚀 OpenAI has announced GPT-5.6. But there is one awkward detail: according to early benchmarks, it still does not reach the level of the Antropic Fable. The new GPT-5.6 lineup comes in three versions: ☀️ Sol — the new flagship model 🌍 Terra — roughly on the level of GPT-5.5 🌙 Luna — slightly below GPT-5.4 Pricing is also split by tier. Sol costs $5 / $30 per million tokens. Terra costs $2.5 / $15. Luna costs $1 / $6. Token caching is changing too. Cached prompts are now paid and become 25% more expensive, while the 90% discount for cached tokens remains. The most impressive part is speed. ⚡️ GPT-5.6 Sol is expected to launch on Cerebras next month, reaching around 750 tokens per second. Terra and Luna on Cerebras would also be interesting to see, but there are no details yet — neither on availability nor pricing. For now, GPT-5.6 is only available to a limited group of U.S. partners while OpenAI continues talks with the government. General availability is expected in the coming weeks.

🧠 Anthropic has shared more details about the Fable 5 relaunch. The company officially confirmed that the temporary restriction was triggered by an Amazon report. According to the report, Fable 5 found a real vulnerability and generated code that could potentially be used to exploit it. Anthropic says the same type of vulnerability could also be discovered by other frontier models, including Opus 4.8, GPT-5.5, and Kimi K2.7. So the issue was not that Fable 5 had some unique dangerous capability. The problem was how its outputs were classified and handled. ⚠️ To address this, Anthropic trained a stricter safety classifier. Now, higher-risk cybersecurity requests will be redirected to Opus 4.8 instead of being handled directly by Fable 5. In other words, Fable 5 is back — but with a tighter control layer around security-related prompts. 🔐

🤖 Anthropic has brought Fable 5 back online. As promised, access to Claude Fable 5 is being restored after a temporary suspension caused by U.S. export control restrictions. The key point: Fable 5 is returning for regular users, while Mythos 5 remains more limited and will be available under stricter access rules. This is not just a model comeback. It shows where frontier AI is heading: the most powerful models are no longer only product launches — they are becoming regulatory events. ⚖️ AI labs, governments, cloud providers, and users are now part of the same control loop.

📸 OpenAI just open-sourced an AI camera. The project, called ImageGenCam, captures a photo and immediately sends it through an image generation pipeline. Set a prompt once, and every shot gets transformed automatically — into anime, comics, children's illustrations, pixel art, or pretty much any visual style. The best part? You can build one yourself. All it takes is a Raspberry Pi, a camera module, a display, a battery, and a 3D-printed case. Feels less like a camera and more like a portal between the real world and a prompt. 🎨

😮 PewDiePie open-sourced his own AI workspace. It’s called Odysseus — a self-hosted alternative to the usual ChatGPT / Claude workflow. What it can do: • Chat with local models or APIs; • Run agents with tools; • Search the web and do Deep Research; • Work with files, memory, notes, tasks, email, and calendar; • Run on your own hardware instead of fully living in the cloud. Basically, PewDiePie got tired of regular AI apps and built his own command center. The funniest part: it already looks less like a YouTuber side project and more like something half of productivity startups are trying to sell. 🫠

😈 Someone got tired of hitting ChatGPT, Claude, and Gemini limits. A developer built a browser extension that lets you continue the same conversation on another account with a single click. The workflow is simple: • You hit the limit; • The extension saves the chat; • The conversation resumes from a different account. No manual copy-pasting. No starting over. The project is free and doesn't require registration. Looks like rate limits are becoming just another engineering problem. 👀

🤖 ChatGPT passed the Turing Test better than humans. Researchers at UC San Diego ran a classic Turing Test experiment: participants chatted with both a human and an AI, then had to guess which was which. Nearly 500 people took part. GPT-4.5, when given the right persona, was identified as human 73% of the time — outperforming actual humans in the study. Llama 3.1 405B also delivered a strong result at 56%. The most interesting part: prompting mattered a lot. Without a well-defined role, the models' performance dropped sharply. Turns out the challenge is no longer making AI sound human — it's telling it who to be. 👀

💸 Nvidia found a new $200B market. Jensen Huang says the next huge business for Nvidia is not another GPU. It is CPUs for AI agents. The chip is called Vera — Nvidia’s new processor built specifically for agentic AI. The logic is simple: GPUs handle the “thinking” part of the model, but agents need CPUs to actually run tasks, use tools, manage workflows, and operate like tiny software workers. So if the world gets billions of AI agents, all of them need compute not only for intelligence, but also for execution. Huang says Vera opens a brand-new $200B market for Nvidia. And Nvidia claims it already has around $20B in Vera CPU business this year. Basically, Nvidia sold the shovels for the AI gold rush. Now it wants to sell the roads too 🛣

🎬 Hideo Kojima and Nicolas Winding Refn just dropped an AI-generated film for Prada. The sci-fi short premiered at Cannes and was created with the help of generative AI. Two of the biggest names in modern storytelling working on an AI project should have been an easy win. Instead, the internet wasn't impressed. Viewers quickly labeled it "AI slop" and criticized the filmmakers for replacing traditional production with generated visuals. The reaction says a lot about where AI video stands today: People aren't judging the technology anymore — they're judging the result.

⚖️ AI is sending more people to court without lawyers. Researchers from MIT and USC analyzed millions of U.S. federal civil cases and found a sharp jump in “pro se” filings — cases where people represent themselves. For almost 20 years, these cases stayed around 11%. In 2025, they jumped to 16.8%. The reason is not that AI suddenly turned everyone into lawyers. It made the first step cheaper: write a complaint, structure the argument, understand the basic procedure, and finally dare to file the case. The strongest growth happened in cases where the document itself matters a lot — civil rights, consumer credit, foreclosure disputes. But there is a catch. Courts are not getting faster. Instead, they are getting more paperwork. In 2025, the number of docket entries in the first 180 days of these cases grew by 158% compared with the pre-AI period. So AI is not just “democratizing justice”. It is exposing how much of the U.S. court system was built around expensive intermediaries. When the lawyer-shaped bottleneck gets partially removed, the system does not become smooth. It starts flooding 🌊

🧮 Google DeepMind’s AI agent just attacked Erdős problems. Researchers introduced AlphaProof Nexus — a system where an LLM tries to solve hard math problems, while Lean checks every proof step and sends feedback back to the model. Result: the agent autonomously solved 9 out of 353 open Erdős problems. It also proved 44 out of 492 OEIS conjectures. Even crazier: the estimated compute cost per solved problem was only a few hundred dollars. Mathematicians are still safe. But the robots are definitely doing homework now 🤖