与「language」相关的搜索结果

QVAC SDK 0.14.0 is live. This release makes the on-device stack faster on mobile, ships the developer-agent path, and takes local text-to-speech to 31 languages. Main highlights: - OpenCode and OpenClaw. The first official OpenCode plugin, plus a maintained OpenClaw compatibility path, both built on managed mode and qvac serve. Point a coding agent at a local model with far less setup and far fewer surprises. - Brain-computer interface transcription, on the SDK. Take recorded neural signal data and decode it into text, fully on-device, no cloud. Stream it in chunks through a simple API. In 0.14 it runs GPU-accelerated on iOS. - Text to Speech in 31 languages with our Supertonic3 upgrade. VOICE AND SPEECH - Supertonic3 multilingual TTS, 5 languages to 31. - Chatterbox and Supertonic now run on the Android GPU, with lower memory use (especially on iOS), quantized s3gen Chatterbox support, and a fix for Chatterbox occasionally emitting random speech. - Whisper transcription now runs on the iOS GPU. Parakeet runs on the Android GPU, with steadier real-time streaming. VISION AND OCR - VLM multi-tile batching: high-resolution Pan and Scan images are encoded in one pass instead of tile by tile, for faster vision throughput. - OCR on ggml (EasyOCR and DocTR) reaches full speed parity with the onnx path, across Metal, OpenCL, and Vulkan. PLATFORM AND RELIABILITY - Dynamic compute backends on Linux: one build picks the right backend at runtime, and opens the door to ROCm and CUDA support without per-backend builds. - Thinking tokens are kept out of the model context, so reasoning no longer fills the KV cache. SDK 0.14.0 is now leaner and faster to start. Let’s build.

A new week is a good reminder that the best time to build is before everyone else notices the opportunity. That is one reason I've been following @CNPYNetwork. Instead of asking developers to learn entirely new systems, the platform is making blockchain development accessible through familiar programming languages while giving builders greater control over their own appchains. The goal is not just to simplify development. It is to remove barriers so more people can turn ideas into working applications. If you're interested in where AI assisted development and appchain infrastructure are heading, this feels like a good week to explore the testnet and see what is already possible. The next generation of builders is already getting started.

That slow head turn + “explain yourself” stare is every wife’s universal language. Bro is just smiling like “yes dear, you got me” Marriage in one clip. Writer: Oliver

Andrew Carnegie spent a fortune building 2,500 free libraries so a poor kid could read what a rich one read. We are finishing the job he started. The only difference is the library now talks back, in every language, and it never closes.

Binance getting squeezed in Europe is bigger than one exchange headline. This is MiCA turning from paperwork into market structure. The world’s biggest crypto venue failed to get its EU license in time, pulled its Greek application, and now has to restrict services for EU users from July 1 That matters because regulation does not just change legal language. It changes where liquidity lives. If users move, volume moves. If volume moves, spreads move. If spreads move, execution changes. That is the part traders should actually watch. The funny part is there is no full panic yet. Binance saw $400M + in weekly net outflows, but that is still small compared to the size of the platform. So this is not a bank run story. MiCA is forcing crypto into a cleaner, more licensed, more supervised market. More rails, more rules, more winners and losers.

POV: It's 2026, and someone captured an ordinary day on a 35mm camera. No dramatic story. No perfect poses. Just warm sunlight, candid smiles, quiet streets, and little moments worth remembering. 📸 Prompt Seedance 2.0 ↓ Prompt: Main subject from image1 lock her completely. Face, hairstyle, body proportions, skin tone, outfit, accessories, and overall appearance must remain perfectly identical throughout the entire video. She is a sweet and naturally beautiful Indonesian girl, 18 years old, soft warm tan skin, expressive dark brown eyes, long black hair tied into a casual messy bun with loose bangs, youthful smile, minimal makeup, wearing a simple white oversized T-shirt, light blue denim jeans, white sneakers, and carrying a vintage brown canvas camera bag. Around her neck hangs a classic 35mm analog film camera. Theme: "A Day With My Camera" A relaxing weekend in 2026, following a girl who loves documenting ordinary moments with her analog camera. The video should feel authentic, nostalgic, and effortlessly beautiful. Everything takes place around modern Indonesia in 2026: quiet neighborhood streets, small coffee shops, city parks, bookstores, pedestrian sidewalks, and cozy residential areas. The entire video must look like it was filmed naturally by her best friend using a modern mirrorless camera. Absolutely no posing, no influencer behavior, no scripted acting. Every smile, glance, and movement feels spontaneous and real. Camera language is the highest priority. Natural handheld movement. Gentle camera sway. Occasional focus breathing. Real autofocus adjustments. Slight framing imperfections. Natural motion blur. Documentary-style tracking shots. Soft depth of field. Authentic available sunlight. Premium lifestyle vlog aesthetic. Sequence: She walks out of her house while adjusting the analog camera hanging around her neck, smiling softly as morning sunlight lights up her face. She stops beside a quiet street and photographs flowers, bicycles, passing people, and little everyday details with genuine curiosity. She enters a cozy camera store, happily choosing a fresh roll of 35mm film before carefully loading it into her camera. Walking through the city, she photographs random candid moments: children playing, street cats, trees moving in the breeze, reflections on café windows, and interesting architecture. She sits alone inside a cozy café reviewing her newly taken photos, smiling while enjoying an iced latte. Golden hour arrives. She slowly walks through a city park while continuing to shoot photographs, warm sunlight creating beautiful natural lens flares. She notices the person filming her, laughs shyly, lifts her analog camera toward the lens as if taking one final photo of the viewer. The camera gently lowers while she keeps smiling naturally. Recording suddenly cuts off mid-motion like a real weekend vlog ending. Lighting: Bright natural morning sunlight transitioning into warm golden hour. Color: Natural realistic skin tones. Soft warm colors. Subtle filmic contrast. No exaggerated grading. Audio is strictly natural. Birds. Footsteps. Camera shutter clicks. Film advance lever. Coffee shop ambience. Light wind. Distant traffic. People talking naturally. Leaves moving. No music. No narration. No subtitles. No logos. No watermark. Ultra-realistic. Looks exactly like a real lifestyle vlog filmed in 2026.

Fine-tuning massive LLMs used to be painfully slow, but not anymore! 4 open source libraries that accelerate fine-tuning of Large Language Models 1. Unsloth AI • Fine-tune models like Qwen3, Llama 4, and Gemma 3 up to 2× faster with 70% less VRAM • Uses optimized Triton kernels and manual backprop for exact accuracy • Supports low-resource setups and runs on consumer GPUs or even Colab/Kaggle with ~3 GB VRAM GitHub repo → 2. LLaMA Factory • Fine-tune over 100 models (LLaMA, Mistral, Gemma, etc.) using a simple CLI or WebUI • Supports LoRA, QLoRA, full or frozen fine-tuning across 2–8‑bit precision • Includes built-in dataset templates, training monitors, and model export options GitHub repo → 3. DeepSpeed • Built for large-scale distributed fine-tuning with ZeRO and FSDP • Optimized for multi-GPU and multi-node training with advanced memory management • Trusted in production environments for scalable LLM training GitHub repo → 4. Axolotl • Yaml-based setup for fine-tuning, LoRA/QLoRA, DPO, GRPO, and multimodal workflows • Includes kernel optimizations for memory-efficient training • Actively maintained with support for Hugging Face, model export, and inference GitHub repo →

Japan's Prime Minister briefed her country with wet hair last night. Beijing would have detained the citizen who filmed it. At 10:29 p.m. on Friday, a magnitude 5.6 earthquake struck Japan's Yamanashi Prefecture, registering a maximum seismic intensity of lower 6 in the town of Fujikawaguchiko at the foot of Mount Fuji. By 11:15 p.m. — forty-six minutes later — Prime Minister Sanae Takaichi was standing at the Prime Minister's Office briefing the nation. Crisis management center activated. Director-general-level emergency gathering team convened. Human life first. Information to the public, promptly and accurately. She was also visibly straight out of the bath. Hair still wet. No makeup. Takaichi posted on her own X account a short time later, in plain language: she had come directly from the bath without time to dry her hair or apply makeup, and apologized for her appearance. She did not have to volunteer that detail. She chose to. That choice is the story. Because somewhere about 1,700 miles to the west, operating under the same physics but a very different political philosophy, the first hour after a magnitude 5.6 earthquake would have looked nothing like this. It would not have been spent activating a crisis center, dispatching emergency teams, and putting the head of government in front of cameras to admit she had rushed straight out of the shower. It would have been spent deciding what to tell the public, what to delete, and which citizen with a camera to detain. We know because we have watched it happen. In Wuhan in early 2020, the doctors who tried to warn the world about a novel coronavirus were summoned by police and forced to sign confessions for "spreading rumors." The citizen journalists who filmed the morgues and the sealed apartment doors — Chen Qiushi, Fang Bin, Li Zehua — were disappeared by the state. Fang Bin would later be sentenced to three years in prison; he was held for the duration. In Zhengzhou in July 2021, passengers drowned trapped in a flooded subway tunnel while state propaganda ran headlines about heroic rescue. When BBC correspondent Robin Brant asked the local government how a metro system less than a decade old could leave passengers to die on a platform, the Henan branch of the Communist Youth League posted his whereabouts to its 1.6 million followers and called for people to track him down. Death threats followed within hours. In Hebei in August 2023, when the floodwaters from Typhoon Doksuri had to go somewhere, authorities diverted them away from Beijing and into Zhuozhou — and the Hebei provincial Party Secretary, Ni Yuefeng, publicly declared the province would "serve as a moat for the capital." Videos of the submerged villages disappeared from Chinese social media within hours. And in Sichuan in 2008, after a magnitude 8.0 earthquake killed at least 5,335 schoolchildren in school buildings that collapsed while government offices nearby remained standing — what citizens named "tofu-dreg schoolhouses" — the writer Tan Zuoren tried to compile a list of the dead. He was sentenced to five years in prison. Huang Qi, the activist who tried to help the parents, got three years; in 2019, the Party gave him twelve more on state-secrets charges. He is still inside. The pattern is not a series of accidents. It is a system. In the People's Republic of China, the function of the state in a disaster is not to serve the public. It is to protect the Party from the public. Compare and contrast. In Tokyo on Friday night, the head of government decided that telling the country what she knew, forty-six minutes after the ground stopped shaking, mattered more than how her hair looked. In Beijing under any equivalent scenario, the head of government would not be at a podium for hours, or days. The citizens with cameras would already be on a list. Wet hair is not the real headline. Wet hair is the headline because of what it accidentally exposes: a democracy is a system that runs toward its citizens in the dark. A dictatorship is a system that hides from them. Sanae Takaichi did not need to apologize for her hair. The Chinese Communist Party owes apologies it will never make, to families whose dead it never named. ACI — Aric Chen | Insights

NVIDIA just made AI detect objects 10x faster by deleting one step. It's called LocateAnything, and it removes the biggest bottleneck no one else was fixing in vision-language models. Normally a model builds each bounding box one coordinate token at a time. 100 objects means thousands of tokens before an answer. NVIDIA scrapped that: their Parallel Box Decoding predicts the whole box in a single forward pass, as one atomic unit. → 12.7 boxes/sec on one H100 → 10x faster than Qwen3-VL → +3.8% F1 on LVIS, accuracy up, not down → 3B params, runs on one consumer GPU Treating the box as one unit keeps its coordinates tied together, which is why accuracy climbed instead of falling. One model handles detection, GUI grounding, OCR, and document understanding, ready for computer-use agents, robotics, and document pipelines. 100% open source, weights, code, demo, and paper all live.

Today, we’re excited to announce our $50M Series B, led by @GreenfieldVC, with participation from @lightspeedvp and @notablecap. 🚀 At Patronus AI, we develop simulations and evals to train and improve AI. The first phase of AI was built on static benchmarks, but that era is over. As agents are used to solve longer and longer tasks, they need to practice in dynamic, living worlds to get better. Simulations are the critical infrastructure powering this next phase. As a company, we’re behind the most influential research and products in AI evaluation, like FinanceBench, Lynx, and Percival. And things have moved at the speed of light since.⚡ We partner with the world's leading frontier AI labs and enterprises, and our revenue has grown more than 15x over the past year. Additionally, today, we’re introducing a preview of the first Digital World Model for AI agent training and simulation: Patronus-DWM. Digital World Models are language diffusion world models that predict realistic environment behaviors and steer agent actions across digital workflows. Just as physical world models predict how objects move through space, we’re developing the equivalent for the digital world: predicting how agents act in digital workflows, then using that to scale the creation of high-quality training data for LLMs. Digital World Models help us push the frontier of ultra long horizon workflows, and unlock a new class of self-improving RL environments. This is our scalable approach to simulating all of the world’s intelligence. The round was also joined by @datadoghq, @SamsungVentures, @gokulr, @factorialcap, and a large cohort of amazing AI leaders across @AnthropicAI, @OpenAI, @GoogleDeepMind, @nvidia, @Recursive_SI, and more.✨ It has been the ride of a lifetime. But we’re just getting started. The best is yet to come. "Do not go gentle into that good night, Rage, rage against the dying of the light" - Dylan Thomas (1954)