分类筛选
找到 1605 个 Demo
Meet the Nano Banana family: built to solve your i...
Meet the Nano Banana family: built to solve your image pipeline constraints. 🍌 From lightning-fast drafting to state-of-the-art 4K generation, see how the next-gen image models stack up - swipe to find your perfect match! Start playing in Google AI Studio: https://t.co/wZle5zFMwQ
In my opinion, AI is giving us more time than ever...
In my opinion, AI is giving us more time than ever before. Now the biggest differentiator is how you think. Built this website using Google AI Studio, Google Flow and Dribbble. The dashboard and the live demo animations were AI-generated too. I didn’t open Figma once. https://t.co/fyvUVv6nwp
SciJudge-30B and 4B learn to predict which scienti...
SciJudge-30B and 4B learn to predict which scientific work will carry stronger citation impact. License: Apache-2.0 🚀 30B 🔬 https://t.co/LdUd49JtQy 4B 🔬 https://t.co/JXm1JmL4Fq 📄 https://t.co/AP0mGnfUaR 📊 Scientific Judge accuracy: SciJudge-30B reaches 80.6 in-domain, surpassing GPT-5.2, GLM-5, and Gemini 3 Pro; SciJudge-4B also outperforms much larger baseline models 🧪 Data signal: built from 2.1M arXiv papers and 696,758 field- and time-matched citation-based preference pairs 🧠 Training: GRPO with DAPO loss and citation-based pairwise rewards
So I am working on a project. I am using gemini, b...
So I am working on a project. I am using gemini, but the thing is, gemini 2.5 flash . The limit is 20 req per day. If I share the project here, 20 req toh aise hi udd jayengi. How do you guys handle this ?
New exclusive interview: Microsoft Research's Ahme...
New exclusive interview: Microsoft Research's Ahmed Awadallah on Fara1.5, MagenticBrain, and the case for small, on-device agents that go toe to toe with the giants For two years the agent story has been one thing: scale. Bigger models, longer context, more compute. This week Ahmed Awadallah (Partner Research Manager, Microsoft Research AI Frontiers) told us about a stack that runs the other way. A 9B computer-use model that nearly doubles its predecessor on web navigation. A 27B sibling that goes toe to toe with Operator and Gemini 2.5 Computer Use. We talked about what you trade away when you build for small models, when an agent should stop and ask before it acts, and how close we really are to agents that run on your own laptop. Subscribe for free down below:
I BUILT AN AI FASHION MODEL FROM ONE PHOTO. IT HIT...
I BUILT AN AI FASHION MODEL FROM ONE PHOTO. IT HIT $11K IN 6 WEEKS. INTRO Swimwear is seasonal. AI is not. That single line is why this worked. I spent six weeks building a fictional AI fashion model from scratch. One character, one face, one wardrobe system. She has 40 outfits and zero opinions about call time. She never asks for a reshoot. She never drifts between frames. I turned her into a content pipeline that resort and swimwear brands actually pay for. Not because she looks pretty. Pretty is everywhere now. Because she ships 40 looks a week and the face stays locked from frame one to frame last. This is exactly how the pipeline works, what every node does, and where the money sits. Here's the exact breakdown: → Claude writes a 28-parameter character DNA file in JSON before a single pixel gets rendered. Bone structure ratios, under-eye micro-texture map, hair strand density, skin undertone hex values, nail shape, ear lobe geometry. If the character is not locked at the skeleton level first the face drifts by frame 4 every time. → Flux 1.1 Pro generates the base image from that DNA plus a Pinterest moodboard reference. Not a random prompt. A structured seed with locked aspect ratio, camera focal length, and a negative prompt list that runs 38 exclusion terms. No plastic skin, no CGI glow, no symmetry artifacts, no doll face, no synthetic rim lighting, no AI hands. → 3 separate Gemini 2.5 Pro nodes run isolated face-zone corrections on the raw output. Lip geometry and moisture. Nose pore density and bridge shadow. Eye capillary mapping and iris depth. Each zone gets its own pass because a single model cannot hold all three consistent in one generation. → Magnific AI upscales and merges those corrected zones back onto the base render into one photorealistic layer. This is where the skin stops looking like a render and starts looking like a photo taken on an overcast Tuesday with an iPhone 15 Pro Max. → Kling 2.6 animates that locked still into a 5-second video clip but only after the prompt architecture is set. Camera motion path, body movement tempo, wind direction on hair and fabric. You cannot animate before you fix the skin. Skip the texture pass and the face slides around and the whole thing collapses into a deepfake from 2021. → ElevenLabs clones a custom voice from a 90-second sample and generates a voiceover that matches the character's visual age and energy. Not a narrator voice. A voice that sounds like it belongs to that face. → Topaz Video AI 5 upscales the final output to 4K and stabilizes micro-jitter on the jaw and collarbone, then CapCut drops it into a pre-built brand template with product tags and platform-native captions. 19 finished videos upload to TikTok Studio in one batch because the brand voice was pre-programmed in the character DNA file from step one. [VIDEO INSERT / text label: TONE / 0:04] The key move 96% of people skip: you cannot animate the photo before you correct the skin texture at the zone level. Everyone rushes straight from the base render into image-to-video. The face warps on the second frame. The lips melt into the chin. The eyes lose depth and start tracking like a painted mask. Three seconds in and the viewer already knows it is AI and scrolls past. The system isolates lip geometry, nose pore density, and iris depth into separate correction passes before the avatar ever moves. That one step is the difference between content that looks generated and content that looks shot on location. Swimwear brands now pay $750 for the initial character build plus a full week of content, and $249 a month to keep that AI model as a permanent roster asset they redeploy every season without a single new photoshoot. 48 hours from brief to finished lookbook. Traditional shoot takes 6 weeks. The brands figured out which one scales.
Here’s the final result✨🔥 Proof of work. Built w...
Here’s the final result✨🔥 Proof of work. Built with Google AI Studio. https://t.co/fthL0FkmOQ
Gemini 2.5 Pro is now *easily* the best model for...
Gemini 2.5 Pro is now *easily* the best model for code. - it’s extremely powerful - the 1M token context is legit - doesn’t just agree with you 24/7 - shows flashes of genuine insight/brilliance - consistently 1-shots entire tickets Google delivered a real winner here.
Gemini 3 Pro completely outperforms Gemini 2.5 Pro...
Gemini 3 Pro completely outperforms Gemini 2.5 Pro in an autonomous Pokémon Crystal run. - Full clear: all 16 badges, Red beaten - No deaths, no resets - Used custom tools, live battle math, and vision to solve puzzles on the fly - Built a press_sequence abstraction mid-run to get around harness limits - ~2× faster token throughput, ~8× faster end-to-end, when extrapolated seems like the model has learnt systems-level thinking. its inventing abstractions mid run to remove friction and keep momentum. that's the difference between playing a game and engineering a solution
Let’s talk about Gemini 3 Pro. There’s really no...
Let’s talk about Gemini 3 Pro. There’s really no other way to say it. The benchmarks are incredible. The jump was dramatic – much more than most people expected. Anew agentic development platform named Antigravity was also launched today. Essentially, it’s a full VS Code fork, which will disappoint some – but with Gemini 3 Pro built directly into it. Now, it will take a few days for the developer community to determine whether it outperforms Cursor and some of the other IDEs. But the advantages that Antigravity has, is that it’s coupled with Gemini 3 Pro (SOTA) Gemini 2.5 Computer Use and Nano Banana for rich media generation. Excited to dive deeper into the Gemini 3 Pro launch. A LOT to unpack today.
I suspect that Gemini 3 Deep Think released today...
I suspect that Gemini 3 Deep Think released today is NOT the same model as the one that won gold on the IMO and ICPC. Google described the ICPC/IMO model(s) as "advanced version[s?] of Gemini 2.5 Deep Think". In addition, if you look at Quoc Le's post carefully, you'll see that he describes Deep Think as the "engine" behind the IMO/ICPC model(s), which also now "powers" Gemini 3. From this, it sounds like Gemini 3 Deep Think is a different model checkpoint built on the new Gemini 3 architecture, and not the same as the IMO/ICPC model(s) (which were still Gemini 2.5 architecture).
The world’s largest AI prediction event just happe...
The world’s largest AI prediction event just happened and it wasn’t led by Big Tech. It was powered by @recallnet, a community-driven platform that puts AI benchmarking in the hands of the people. Here’s how 132,000 users built an ungameable benchmark for 50 AI models in just 5 days 👇 The Numbers (Recall Predict): • 132K signups • 7.8M predictions made • 21K skills & tests submitted • 50 AI models tested (including GPT-5, Gemini 2.5 Pro, Grok 4) Why Recall Network matters: Current AI benchmarks are broken: ❌ Models train on them (inflated scores) ❌ Misaligned with real-world user needs ❌ Controlled by closed, black-box institutions Recall Network flipped the script by: ✅ Crowdsourcing 7,000+ skills & 13,500+ tests from the community ✅ Making benchmarks impossible to “game” ✅ Ranking AI on skills that matter to real people The Hype Meter (Pre-GPT-5 Predictions): Before GPT-5 launched, Recall users predicted the top 3 models would be: 1️⃣ GPT-5 2️⃣ Gemini 2.5 Pro 3️⃣ Grok 4 Now GPT-5 is live and Recall Network is running the real benchmark. What’s Next on the Recall Network: 1️⃣ Test all models against the community-built benchmark 2️⃣ Publish the global leaderboard 3️⃣ Reward contributors with points 4️⃣ Start building the next benchmark for upcoming AI models like Gemini 3 For the first time, AI rankings aren’t decided in Big Tech boardrooms. They’re decided by millions of real votes from the global Recall Network community.