HOLY SHITT, Sesame Labs just dropped CSM (Conversational Speech... Model) - Apache 2.0 licensed! 💥 > Trained on 1 MILLION hours of data 🤯 > Contextually aware, emotionally intelligent speech > Voice cloning & watermarking > Ultra fast, real-time synthesis > Based on llama architecture & Mimi like decoder > Apache 2.0 licensed > Weights on the Hub So cool to see such a strong Speech backbone out in the wild! Kudos Sesame team! 🤗show more

Wan 2.2 Speech-to-Video 14B just landed > audio-driven cinematic... video generation > weights & demo on the hub > apache 2.0 license + looks like it could be compatiable with Wan 2.x LoRAs out-of-the-boxshow more

Fuck yeah! MaskGCT - New open SoTA Text to... Speech model! 🔥 > Zero-shot voice cloning > Emotional TTS > Trained on 100K hours of data > Long form synthesis > Variable speed synthesis > Bilingual - Chinese & English > Available on Hugging Face Fully non-autoregressive architecture: > Stage 1: Predicts semantic tokens from text, using tokens extracted from a speech self-supervised learning (SSL) model > Stage 2: Predicts acoustic tokens conditioned on the semantic tokens. Synthesised: "Would you guys personally like to have a fake fireplace, an electric one, in your house? Or would you rather have a real fireplace? Let me know down below. Okay everybody, that's all for today's video and I hope you guys learned a bunch of furniture vocabulary!" TTS scene keeps getting lit! 🐐show more

Big moment for text-to-speech. Qwen just open-sourced a text-to-speech... model that lets you clone voices, design new ones, and control speech using natural language. Let me explain what I mean: You can literally tell it "speak in a cheerful tone with slight nervousness," and it actually does that. No complex audio engineering needed. What makes this special: - 3-second voice cloning - Covers 10 languages: English, German, French, and more - Latency as low as 97ms for real-time applications - Supports both streaming and non-streaming generation The model comes in two sizes (0.6B and 1.7B parameters), so you can pick based on your hardware and quality needs. Three modes to work with: 1. Custom Voice: Use pre-built premium voices with instruction-based style control 2. Voice Design: Describe the voice you want in plain English (or Chinese), and the model creates it 3. Voice Clone: Provide a 3-second reference audio and clone that voice The best part? It integrates with vLLM for production deployment and has a simple Python package you can pip install. I've shared a link to the GitHub repo in the next tweet.show more

seedance 2.0 + my v2 AI UGC prompting system... is giving insane results i spent the last 24 hours generating over 200 seedance 2.0 videos to figure out the best prompting framework system for AI UGC this video was made with 1 prompt and 1 tool, no editing was done to the video this was just a prompt to a video this is by far the best model i've ever used and the craziest part is that it can be fully automated this is the first time we can actually automate high quality ai ugc at this level bytedance owns tiktok so this model is trained on millions of high quality ugc videos. you just need to know how to extract that and call it in your prompt. we are so early... it's insaneshow more

Jointly announcing EAGLE-3 with SGLang: Setting a new record... in LLM inference acceleration! - 5x🚀than vanilla (on HF) - 1.4x🚀than EAGLE-2 (on HF) - A record of ~400 TPS on LLama 3.1 8B with a single H100 (on SGLang) - 1.65x🚀in latency even for large bs=64 (on SGLang) - A new scaling law: more training data, better speedup - Apache 2.0 Paper: Code: SGLang version: ⚒️Takeaway: Introducing training-time test, a novel draft model training technique: we replace feature prediction with direct token prediction and shift from top-layer-only features to multi-layer feature fusion. This approach unlocks a new scaling law previously undiscovered in EAGLE and EAGLE-2. 🙏Acknowledge: We would like to thank the SGLang team (zhyncs Lianmin Zheng Ying Sheng James Liu, Ke Bao, and others LMSYS Org) for their merge and careful evaluation of EAGLE-3 on SGLang. 🤝Want to collaborate? We're a small academic group with limited GPU resources. If you're interested in supporting our next version of EAGLE or would like us to train a preliminary version tailored to a specific model, please get in touch! Joint work with Yuhui Li, Fangyun Wei, and Chao Zhangshow more

This week is already so hot. 🔥 Massive release... from Decart : Lucy 2.0 a World Editing Model running at 1080p, 30FPS in realtime. This is truly exciting, the era of real-time generative reality is here. We are moving from watching AI video to living inside AI video. A breakthrough model capable of transforming the visual world in real-time. Moving beyond offline rendering, Lucy 2.0 delivers high-fidelity 1080p video generation with near-zero latency. Lucy 2.0 literally "redraws" the entire world pixel-by-pixel, while you are watching it. e.g. If you want to be an anime character, it doesn't just put a mask on you. It turns your skin into anime skin, your hair into anime hair, and the lighting in your room into anime lighting. Lucy 2.0 is also trained to stop the generated video from slowly falling apart over time, so the same stream can run much longer without faces and details drifting. So why is this a "Massive Deal"? Traditional AI video-generation model takes a prompt, you wait 10–20 minutes, and the computer "bakes" a video for you. You couldn't touch it or change it while it was happening. But Lucy 2.0 works like a mirror. It happens in real-time (30 frames per second). There is no waiting. You move your hand, the AI character moves its hand instantly. The craziest part isn't the visuals; it's the physics. Usually, AI hallucinations are glitchy—hands merge into faces, walls melt. Lucy 2.0 understands how the world works without being told. It knows that if you take off a helmet, there is hair underneath. It knows that if you splash water, droplets fly. It learned "physics" just by watching millions of videos. The physical behavior you see emerges from learned visual dynamics, not from engineered geometry or explicit physics engines. Their official technical report explicitly states that the model does not use traditional 3D engines, depth maps, or wireframes. It is a "pure diffusion model."show more

I just ran Gemma 4 31B on @CerebrasSystems at... 1,800+ tokens/sec and it's multimodal. For context: that's 35x faster than a typical GPU endpoint, and the first token (reasoning included) lands in 1.5 seconds. This isn't a benchmark slide, I recorded the inference live. Prompt I used: "Create a simulation of an iPhone. Include at least one working dummy note taking app, a functional notification pulldown, high quality graphics, single HTML file, any libs via CDN." - Generation time: 3 seconds. - Notes app worked. - Notification panel worked. - Rendered first try. This is what wafer-scale inference unlocks, not just "faster," but a different category of product. When generation is this fast, you stop waiting and start iterating in real time. Why this matters: Gemma 4 31B is Google DeepMind's flagship open weight model, Apache 2.0 licensed, dense (not MoE), and built for efficiency over raw parameter count. It scores close to Claude Haiku 4.5 on the Artificial Analysis Intelligence Index (30 vs 29) but runs ~18x faster on Cerebras. It's also the first multimodal model on Cerebras's platform, meaning you can now feed it screenshots, documents, charts, and UI states at wafer scale speed. # Applications I'm most excited about: - Screenshot → Insight: Drop in a dashboard or document screenshot, get structured findings back instantly. no waiting, no batching. - Live UI generation: Full interactive interfaces (like my iPhone sim) generated and rendered in under 2 seconds. - Screenshot -> Patch: Feed it a broken UI + console error, get a minimal code fix and verification steps back. - Computer use & agentic loops: See -> reason -> act - verify, fast enough to keep a human in the loop instead of waiting on the model. - Long context summarization: Full research reports condensed into decision ready summaries you can read and requery in one sitting. The bigger unlock isn't the speed number itself, it's that agentic and multimodal loops (see -> reason -> output -> tool call -> verify -> retry) finally run in real time instead of feeling sluggish. As Logan Kilpatrick (Logan Kilpatrick) put it: "If every model was doing 2,000 tokens per second, you wouldn't build the same product and just have it be faster, you'd build different products." Gemma 4 31B is live now on Cerebras Inference Cloud in public preview. If you're building multimodal, agentic, or real time apps, this is worth testing today. What would you build with such insane inference throughput?show more

This looks like a mid-five-figure luxury campaign shoot, but... I actually just put it together for about $0.02 per second using Pippit and the new Seedance 2.0 Mini model. Most marketing teams treat every asset like a massive production task, which is why they run out of budget so fast. I’ve been testing the Marketing Shortcuts on Pippit to see if I could skip the studio entirely and go straight to a Narrative Ad format. The reason Seedance 2.0 Mini is actually useful here is the speed. It lets you take one product shot and turn it into a story-driven ad in a few seconds. This kind of Narrative Ad usually takes weeks to produce, but this workflow makes it easier to test different plots and iterate without the usual production headaches. It’s basically moving from a text prompt to publishable content without needing the massive budget. There is a limited-time deal running right now if you want to try the workflow yourself. #PippitAI #Seedancemini #PippitPartner #AIads Pippitofficialshow more

The architecture of this new world model is one... of the most interesting things I've seen lately: Let me first explain how most world models work: They predict and render one frame at a time. If you are navigating in one of these worlds, and you look left, the model draws whatever looks right in the moment. Every time you change your viewpoint, the model has to imagine what should be there again, so it's very common for these models to "forget" what's in the world. For example, if you put a toy on the table, look away, then look back, the toy might not be there anymore. Tripo AI is releasing its Project Eden model, which works very differently: The model builds the world first, and then renders it based on that map. That map holds the real state of the world: the geometry, every object, where things are, what's already happened. The picture you see on screen gets generated from the map. This architecture flips the whole thing. Now, you get the following: 1. The world stops forgetting. Leave, come back, and the toy is still on the table because it lives in the map, not in the last frame you saw. 2. You can edit the world, and those changes persist for anyone who enters later. 3. Multiple people and AI agents can coexist in the world and see it from different perspectives. This is early research, but it's looking really promising. They just raised nearly $200M across two rounds to build it out. Tripo will be at SIGGRAPH 2026 (July 19–23, Los Angeles Convention Center). If you work in 3D, embodied AI, simulation, or anything spatial, go connect with them there.show more

Day 5 of my 30-day water fast. Yesterday (Day... 4) was the best day I’ve had so far. •Least cravings. •Least anger. •Least frustration. •Least crash-outs. It feels like I’m purging out every negative emotion and toxin that’s lived inside me. I’m having strong emotional responses—crying at things I wouldn’t normally cry at. Even just thinking about food hits different. Taco Bell. Sushi. Kebab. And during Ramadan, knowing my mom is cooking up a storm every night makes it even harder. The vape cravings? Completely gone. But I know that if I think about it too much, my body will remember, so I just keep my mind moving forward. Started at 200 lbs on Day 0 (night time weigh in with food intake and water intake, dropped to 192.2 on Day 1, and today, on Day 5, I’m at 181 lbs. Not trying to lose weight too fast—I need as much on me as possible to survive the next 25 days. But the best part of all this? Being live 24/7. Not just to prove I’m not eating, but because I’m building a real bond with the people watching. It’s just me, them, and Muffin in this cabin. Love you guys. Inshallah, I’ll succeed today and be back tomorrow for a Day 6 update. 💕💦show more

I told you to claim your free 16GB NVIDIA... GPU for learning Local LLMs. Now I’m going to show you how to double its inference speed without touching the hardware. Google Colab gives you an enterprise grade NVIDIA Tesla T4 GPU for free, roughly 4 hours every single day. It is the absolute perfect sandbox for learning AI engineering, testing inference flags, and pushing massive context windows. The local AI timeline is moving way too fast. If you aren't using Multi Token Prediction (MTP) yet, you are leaving massive performance on the table. I just pushed DeepMind’s Gemma 4 26B to 64.9 t/s on this exact free tier. Let's look at the raw benchmark data running on an Ubuntu Linux environment with the latest compiled llama.cpp binaries and quantized GGUFs from Unsloth via HuggingFace: # Qwen 3.5 9B (Dense): Base: [ Prompt: 626.7 t/s | Generation: 21.0 t/s ] With MTP: [ Prompt: 539.1 t/s | Generation: 24.8 t/s ] # Gemma 4 26B QAT (MoE): Base: [ Prompt: 634.2 t/s | Generation: 48.3 t/s ] With MTP: [ Prompt: 572.1 t/s | Generation: 64.9 t/s ] If you are paying attention, this single Colab notebook reveals 3 massive observations about the current state of local LLMs: # 1. The MTP Speedup (Software Overclocking) Standard autoregressive decoding guesses one token at a time. MTP acts like a highly optimized, built in speculative decoder. It predicts multiple future tokens at once and the main model verifies them in parallel. The result? Zero accuracy loss and a massive throughput increase. Gemma jumped from 48 to 65 t/s just by flipping a flag. # 2. The MoE Paradox (Bigger is Faster) How does a 26B parameter model absolutely destroy a 9B model in raw speed on the exact same hardware? Architecture. Qwen 3.5 9B is a dense model. it activates all 9 billion parameters for every single token. Gemma 4 26B is a Mixture of Experts (MoE) model. It routes data efficiently, activating only 4B parameters per token. You get the reasoning capabilities of a 26B model with the compute cost of a 4B model. 3. Thinking Efficiency When I ran the exact same complex prompt on both models, the larger MoE spent significantly fewer "thinking" tokens to arrive at the correct answer. A smarter model doesn't just give better answers; it gets to the point faster, saving you compute cycles and preserving your context window. # Want to run this yourself? Here are the exact llama.cpp CLI commands. For Qwen (MTP is baked into the main model): ./llama-cli -m Qwen3.5-9B-UD-Q4_K_XL.gguf -p "Explain quantum computing." -n 2000 -c 8000 -ngl 99 -fa on --spec-type draft-mtp --spec-draft-n-max 4 --spec-draft-p-min 0.7 For Gemma (Using a separate lightweight draft model): ./llama-cli -m gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf --model-draft mtp-gemma-4-26B-A4B-it.gguf -p "Explain quantum computing." -n 2000 -c 8000 -ngl 99 -fa on --spec-type draft-mtp --spec-draft-n-max 4 --spec-draft-p-min 0.7 Stop waiting for a $3,000 rig. Boot up Colab, pull these models, and start building your stack. I’ve put together a completely free, cell by cell Google Colab notebook that automates this entire workflow so you can test it yourself in 5 minutes and learn. Link to the notebook is in the comments below. Experiemt with different MTP parameters, context windows and post your results in the comments.show more

Pretty over catching flak for just existing on peanut's... server. let's recap the "yOuRe RuInInG tHe SeRvEr" narrative: - Spent a lot of dedicated time helping teach/decorate/guide new people through the game - Suggested the SAM site solution to peanut in S2 (which fixed the problem for the whole server) - The imbalance of team numbers was a "fun problem" so I suggested the naval war as numbers are not nearly as important and everyone will have more fun doing it - Farmed my ass off to not only build a big cool fantasy base for Peanut in S1, but also to supply materials for multiple RPers and others who were struggling to get stuff on their own across every wipe so far. - Coupled with the point above, before peanut had pheetus to farm for him, I also secretly supplied his team with nearly a box of low grade and cloth overnight (really difficult resources to get) so they could continue playing with meds instead of needing to spend a day grinding for them. I never told him or anyone but everyone on stream new we did it to support the nut. - Went out of my way on multiple occasions to go on rescue missions to save people from the deep sea, cargo, and oil rigs I understand that people only see things through the point of view of their streamer, and don't know what I'm doing on my stream, but I spent the entire current wipe just hanging out with my friend, recruiting noobs to do silly ninja role play with us while teaching them about the game. One morning, without asking, I got a DM telling me I was unrestricted and could use guns, so I took an AK out to the deep sea and killed scientists (poorly) and it caused the entire server to lose their minds. Suddenly i have 20k hours (this number increases every time i hear about it), am going to use my 'expert builder skillz' to raid bases, and will abuse everything. Except I didn't do any of that. I still did stupid ninja RP for another 30-40 hours having silly interactions and using C4 for things like "Door breaking ninjitsu" until I was getting harassed to the point where I might as well have joined a sweaty pvp team anyways (so I did, and I did it out of spite). After joining said team I've still mostly only used pistols, not that it matters, but just wanted to include this. I'm too autistic and frustrated to understand how people are coming to the conclusion that I'm such a villain, kind of wish y'all would just tune in and watch instead.show more

🚨 JUST IN: CHINA just released an AI EMPLOYEE... that works 24X7 on its own. 100% OPEN SOURCE. It researches, codes, builds websites, creates slide decks, and generates videos. All by itself. All on your computer. It's called DeerFlow. You give it a task. It makes a plan, spins up its own team of sub-agents, and gets to work. You come back and there's a finished deliverable waiting. Not a draft. Not a summary. The actual thing. Not a chatbot. Not a research assistant. An AI with its own computer that works while you sleep. Here's what it does on its own: → Spawns multiple sub-agents in parallel, each tackling a different piece of your task, then combines everything into one finished output → Writes real code, runs it, reads the results, and fixes its own mistakes without asking you once → Builds slide decks, websites, full research reports, and data dashboards from scratch → Remembers you across sessions. Your writing style. Your tech stack. Your preferences. Gets better every time. → Reads files you upload, works with them inside its own filesystem, hands you clean finished outputs → Searches the web, runs commands, calls any tool you plug in Here's how it thinks: You give one instruction. The lead agent makes a plan. Sub-agents fan out and work in parallel. Results come back. Everything gets synthesized. You get a deliverable. A single research task might split into a dozen sub-agents, each exploring a different angle, then converge into one finished website with generated visuals. Here's the wildest part: DeerFlow 2.0 launched on February 28th 2026 and hit number 1 on all of GitHub Trending the same day. Version 2.0 was a complete rewrite. Zero shared code with version 1. Because users kept using it for things the team never intended. Data pipelines. Dashboards. Entire content workflows. The community told them what it needed to become. So they burned it down and rebuilt it. 22.7K GitHub stars. 2.7K forks. Built by ByteDance 100% Open Source. MIT License.show more

🙏♥️🙏 "Hey... you alright, kid? You got someone?" I... remember trying to answer, but my voice barely came out. "Just me..." I was pulled from beneath twisted metal, shaking so hard I could hardly breathe. Everything felt loud and empty at the same time-until he stepped into it. Sergeant Marcus Hayes. He didn't ask many questions. He didn't make a speech. He just lifted me gently into his arms and said, "You're safe now." And somehow... I believed him. From that moment on, he never really let me go. At the camp, he'd slide part of his meal onto my tray without a word. He showed me how to tie my boots, slow and patient, like it mattered. On the nights when the silence got too heavy and I couldn't stop crying, he'd sit nearby—never forcing me to talk, just making sure I wasn't alone. He wasn't trying to be a hero. He was just choosing to care. And that changed everything. When he brought me home, he didn't just give me a place to stay—he gave me a life to grow into. He became the steady presence I didn't know I needed. He taught me how to stand tall when things felt impossible... how to keep my word even when no one was watching... how to become someone I could be proud of. I watched him in the quiet moments— shaving in the mirror, fixing the truck in the driveway, folding his uniform with care and discipline. Every little thing meant something. And without realizing it, I started building myself in his image. Years later, 1 wear that same uniform. And every time I stand in front of the mirror, adjusting it just right... I don't just see myself. I see Sergeant Marcus Hayes standing behind me quiet, steady, and proud 🙏♥️🙏show more

🚨THE FBI CREATED A FAKE CRYPTOCURRENCY.. LISTED IT ON... UNISWAP.. HIRED MARKET MAKERS TO PUMP IT.. THEN ARRESTED EVERYONE WHO SAID YES.. THIS IS THE CRAZIEST LAW ENFORCEMENT OPERATION IN CRYPTO HISTORY!!! The FBI built an actual ERC-20 token on Ethereum called NexFundAI.. 100 billion token supply.. A professional website.. Whitepapers promising "passive income through AI-powered investing".. It looked exactly like every other crypto project.. Because that was the point.. Undercover agents posed as the founding team.. Then reached out to professional market-making firms and said "we need you to fake our trading volume".. Every single firm said yes.. Here's what they recorded.. Gotbit.. A firm run by a 26-year-old Russian who publicly bragged in 2019 that he built a business faking trade volumes.. His team kept internal spreadsheets with columns literally labeled "fake volume" vs "market volume".. When asked how fast they could pump NexFundAI's volume to $1 million per day.. They said "6 hours.. It will cost about $200".. $200 to fake $1 million in daily trading volume.. MyTrade.. Run by a guy who called himself "the mastermind".. He explained the exact psychology of the scam on camera.. "We make the chart look like a really nice roller coaster ride.. That's where people jump in.. We have to make them lose money in order to make profit".. He said that on a recorded FBI video call.. CLS Global.. A Dubai-based firm.. Their bots generated 98% of NexFundAI's total trading volume.. When the FBI asked if they could sync fake volume spikes with fake news announcements.. They said absolutely.. ZM Quant.. Bots executing 10 to 20 trades per minute through dozens of wallets to look organic.. All of them knew it was fraud.. All of them did it anyway.. All of it was recorded.. And the clients were even worse.. Saitama.. A meme coin that hit $7.5 billion market cap.. The founders coordinated buys through private Telegram chats.. Sent "pump it" memes while manipulating the price.. Then dumped on retail investors.. $7.5 billion.. Built entirely on fake volume.. Every penny of real money came from retail investors who thought the momentum was organic.. One founder left Saitama and started Robo Inu.. Used Gotbit again.. Another launched VZZN.. Same playbook.. Lillian Finance.. Founder claimed to be a defense contractor who addressed Congress.. Marketed the token as funding children's hospitals.. Pocketed everything.. When the FBI shut it down.. They seized $25 million in one day.. 18 people indicted across the US, UK, and Portugal.. The CEO of Gotbit was arrested in Portugal and extradited.. Sentenced to 8 months plus $23 million forfeiture.. But here's the part that broke my brain.. Real people bought NexFundAI.. The FBI's fake token.. With zero utility.. Zero real developers.. Created solely to catch criminals.. Attracted real retail investors because the fake volume made the chart look bullish.. When the FBI pulled the liquidity to end the operation.. Those people lost real money.. On a government-issued token.. The FBI had to set up a restitution portal to pay them back.. And it gets worse.. Within 24 hours of the DOJ announcing the sting.. Someone cloned the FBI's exact smart contract.. Launched a copycat token.. Rode the viral momentum.. And made $127,000 in a single day.. Using the exact same manipulation tactics the FBI just arrested 18 people for.. Then in 2026.. The FBI did it again.. New token called Lexobit.. 10 more arrests.. Including operators extradited from Singapore.. IRS forensics showed that in one firm's trading.. 1,209 out of 1,221 consecutive transactions went straight back to wallets the firm controlled.. 99% circular.. The FBI proved what everyone in crypto suspected.. The volume is fake.. The charts are painted.. The momentum is manufactured.. And every time you buy a token because "the chart looks bullish".. You might be the exit liquidity.show more

IQBAL SYRVE JOKES WRITE THEMSELVES: Independent Media owner Iqqqbal... Surve, who has not paid back R4billion to PIC, not paid back the loan he took from clothing union SACTWU, in 2013, which belongs to poor black women, has suddenly received some doctorate from the International Peace College. The peace college is based on Islamic ethos yet see no shame in accepting his money allegedly looted, he sponsors the Cape Town Jazz Festival. Which according to Islam is not acceptable.,I guess the leaders of that college don’t mind, even if they funded with alleged looted money. The nerve of of Iqbal Surve to start his speech with “many years ago I was fortunate enough to become wealthy “ yah right anyone will with R4 BILLION TAKEN FROM PIC, R200 Million taken from SACTWU. He has the nerve to say philanthropy is not about compassion. You have the power to use what you know to others. BLOODY HYPOCRITE. Pay back the very poor mostly black African women from clothing union SACTWU then you can talk about philanthropy and the other hogwash you speaking about. Allegedly Iqbal Surve is funding the campaign for Lindiwe Sisulu to be the next ANC president. Now I will wait for your worthless news publications to write a story about me. Made up as usual. Just like you Tembisa babies lies. Thank god nobody takes you seriously. Crazy part of this award was he put black students pics allegedly not even studying in this so called college to sit in the front. The actual students not. Frame 1- Iqbal Surve speech Frame 2- black students Frame3&4- actual studentsshow more

BREAKING: LLMs just learned to COMPUTE for real, it's... mean NO MORE GUESSING math. Chinese college kid Guo Hanjiang vibe-coded MiroFish in 10 days (23k+ GitHub stars, $4.1M from Shanda in 24h) - the AI swarm simulator that’s already printing. ByteDance (VolcEngine) dropped the nuclear upgrade: OpenViking - structured viking:// filesystem memory (L0 ultra-summary -> L2 full details) - agents now run 100+ steps with zero amnesia or hallucinations, 11.6k stars and climbing. Now this just dropped and the entire AI timeline is shaking. Startup Percepta embedded a full WASM virtual machine directly into Transformer weights. No more external Python sandboxes. No more hallucinations in exact tasks. The model streams raw machine code at 30,000+ tokens/sec on CPU, executes millions of steps, and solves the world’s hardest Sudoku via real backtracking + constraint propagation - 100% accurate, zero bullshit. They killed the Attention Bottleneck with Exponentially Fast Attention (HullKVCache + 2D heads + convex hull queries in log time). What used to die at 1k steps now flies. This is the bridge: System 1 intuition (normal LLMs) + System 2 deterministic logic (native code execution) in ONE brain. Agents won’t need tools anymore. Heavy simulations will run inside the weights. Check out: Now put it all together: MiroFish swarms + OpenViking infinite memory + Percepta native flawless compute = agents that can hardcore simulate millions of future scenarios, run perfect logic loops for days, and predict events/markets/reality with god-tier accuracy. No drift. No bullshit. Just pure foresight. This combo will change everything, imo. The era of predictive super-agents that actually print the future is here. We’re watching this one closely. Save this combo.show more

Yesterday at 3 AM Claude Code called me I... woke up, picked up the phone, and on the screen was a message: "Wallet entered BTC Up at 11 cents. Open Polymarket?" I said yes and went back to sleep Claude Code unlocked my 2nd phone on its own, opened Polymarket, found the right market, entered the amount, and hit Buy. I could see all of it in real time through the web interface on my laptop. Screenshots from the phone updating every second. By morning the position closed in profit Let me tell you how I got here A week ago I asked Claude Code to write a script that pulls on-chain data from Polymarket and ranks wallets by win rate on 15-minute BTC markets In 20 minutes I had a table with hundreds of addresses, and 1 of them stood apart from the rest. More than 200 trades per day, surgical entry precision, and a profit curve going straight up I fed that address back into Claude Code and asked it to break down the strategy. Turns out the wallet monitors BTC volatility on Binance and Bybit every 100 milliseconds, and when it drops below 0.08% it enters Up and Down simultaneously at 25 to 35 cents A pure straddle: 1 side burns and the other flies to a dollar, giving 3 to 4x per position. Dozens of times a day I wanted to follow it but signals came at any hour, and waking up every 15 minutes for a notification was simply impossible. So I built something else Took an old Android phone and installed an agent running on the Qwen3-VL visual model. It sees what is happening on the screen and mimics human actions through ADB: taps, swipes, text input. Then I connected it to Claude Code as the executor Now the chain works like this: Claude Code monitors the wallet, sees a new position, calls me. And if I say "yes" or just do not pick up within 30 seconds, the agent on the phone opens Polymarket on its own and copies the entry Essentially I built myself an autopilot out of 2 AI systems: 1 thinks and the other presses buttons. I just sleep and occasionally pick up the phone → Here is the wallet the whole thing is tracking: For those who do not want to build a setup like this there is a Telegram bot that handles the 1st part: tracks this wallet and sends a signal on every new entry: AI calls me at 3 AM to ask permission to spend my money A year ago this would have sounded like schizophrenia. Now it is just Tuesdayshow more

This is my "feel the AGI" moment: I used... GPT-5.6 Sol to train my own autocorrect model that outperforms GPT-5.6 Sol (wtf??) I have no ML background. I have no idea what I'm doing. I just kept pushing Sol until it spat out a SOTA model. And I spent $0. The motivation: Years of talking to AI have made me terrible at typing. Rather than fix my skill issue, I decided to throw more AI at it. My idea was: instead of autocorrect that interrupts my flow, I want to type fast with mistakes and have AI clean it up after. I wanted the smallest local model possible, for speed, for battery life, for science! So I decided to train my own. Inspired by Andrej Karpathy’s autoresearch, I ran Codex /goal with this setup: pick an experiment, try it, record the results to a doc, throw it out if it fails, and plan the next experiment without repeating failures. I gave a few examples that had to pass, tight latency targets, and let it run. Sol did some amazing things. First, it scanned benchmarks and shortlisted base models: Qwen 3.5, Gemma 4, Liquid LFM 2.5. It found a dataset on HuggingFace for typed text. Then it built a simulator for fingers striking a Mac keyboard, modeling the physical layout with a Gaussian distribution around each key. It simulated striking the wrong key, wrong order, fat-fingering, etc. With the models + data + simulator, it fine-tuned using MLX right on my MacBook. It had a working prototype within an hour! But accuracy was pretty poor. — Problem 1: Tokenization Sol read papers, ran tests, and identified that the tokenizer was the bottleneck. Tokenization makes typos hard for the model to see, so it memorizes mappings instead of using its language priors. Sol tried ByT5, Google’s tokenizer-free byte-level LLM. This made a big improvement, but the model is old and lacked the knowledge needed to reach Sol performance. Sol dug deeper and realized a tokenizer-free model isn’t needed; instead, it used T5Gemma, an encoder-decoder model. This can understand the input deeply before producing output, and furthermore, Sol could post-train the encoder to improve performance. This gave a much higher ceiling. — Problem 2: Loss function Now the model was correcting some typos perfectly, but ignoring most. Sol realized that standard cross-entropy loss was teaching the model to avoid edits, because the vast majority of characters in the training data were left unmodified. The fix was wild: Sol wrote a custom loss function that byte-aligns the source and target strings, uses a dynamic programming algorithm to compute the minimum edits between the two, then weights correct edits much higher than copies. After a lot of tuning, this dramatically improved accuracy. — Problem 3: Autoregression One failure mode remained: if the model made a mistake, it couldn’t backtrack. It could only predict the next token. Teaching it to “think” like a reasoning model would solve this, but would be far too slow. Sol found a beautiful solution: instead of greedily predicting the next token, beam search over all possibilities. This parallelizes the exploration instead of one linear chain-of-thought. At the end, choose the path with highest cumulative log probability. This worked great, but made the experience worse, since the user wouldn’t see progress until the whole search was done. To fix this, Sol made a clever observation: after each search step, the longest common prefix among surviving branches is guaranteed to appear in the final result, so it can be displayed immediately. As the search progresses, weaker paths are dropped and the prefix grows, so the user sees continuous progress. Sol built all this as a custom MLX pipeline that does the parallel decoding on the MacBook GPU, with just ~40ms TTFT. It’s crazy fast and entirely local. — Final eval (error reduction rate, higher is better): - Apple autocorrect: 49.66% - GPT-5.6 Luna: 82.47% - GPT-5.6 Terra: 87.64% - GPT-5.6 Sol: 90.56% - Our model (1.7B): 91.02% Final cost: - 1 quota reset (thanks Tibo) - $0 (And yes, I verified there's no cheating. In fact, we test words scrubbed from the training data to prove the model isn’t memorizing) There were a ton more details and tangents I could write about: contrastive learning, GRPO, DPO, dynamic masking, and more. Sol is a fascinating and creative model. It blew my mind so many times. Don’t let a lack of experience stop you: Sol makes AI experiments accessible to anyone!show more

my 8 GB VRAM gaming laptop is absolutely going... to hate me for this. but I still did it. ran a 31b dense model (Gemma 4 31b Q4) with only 8 GB VRAM last week I ran Gemma 4 26B A4B a mixture of experts model on my RTX 4060 and hit 25–28 tokens/sec using llama.cpp's new MTP support. smooth. snappy. but MoE has a secret: it only activates 4B parameters per token despite having 26B total. that's why it flies. so the real question started haunting me. what if I throw a full, no tricks, every parameter fires on every token, 31B DENSE model at the same machine? # Hardware: GPU: NVIDIA RTX 4060, 8 GB VRAM RAM: 16 GB CPU: Intel Core i7 H Laptop. Gaming. Modest. The model: gemma-4-31B-it-qat-UD-Q4_K_XL.gguf (model's unsloth huggingface link in the comments) This is Google DeepMind's flagship dense model in the Gemma 4 family that can run on single consumer GPU. It packs a hybrid attention architecture, supports up to 256K context natively, and is QAT (Quantization Aware Training) optimized, meaning it retains far more quality than standard post training quants at the same bit depth. This is NOT the MoE. This is 31 BILLION dense parameters, every single one of them loaded. # the flags I used: -m gemma-4-31B-it-qat-UD-Q4_K_XL.gguf -cnv --spec-type draft-mtp --spec-draft-model mtp-gemma-4-31B-it.gguf --spec-draft-n-max 8 --spec-draft-p-min 0.6 -c 6000 -v Multi Token Prediction (MTP) is still active here. Separate draft GGUF required, same as the 26B setup. # Results: → Decode: ~3 tokens/sec → Prefill: ~2 tokens/sec → Context: 6000 tokens → Hardware crying quietly in the corner: yes so is 3 tps actually usable? For real time back and forth chat? Not ideal. You're not having a fluid conversation at 3 tps. but slow ≠ useless. And this is where it gets genuinely interesting. think about how senior devs actually work in a real team. But when something is architectural, deeply complex, or needs serious reasoning? they walk down the hall and escalate to the senior. That's exactly the local AI agent architecture this unlocks: → Fast orchestrator model (Gemma 4 26B MoE at 25+ tps) handles routing, simple queries, tool calls, memory. The junior dev. → Gemma 4 31B dense is the senior, called only when the fast model genuinely hits a wall. Hard multi step reasoning. Complex code generation. Deep architectural decisions. The agentic loop stays fast. Only the hard hops touch the 31B. That's a legitimate production grade local AI architecture on a budget hardware. (requires 2 8gb gpus) other workflows where 3 tps is completely fine: - overnight batch jobs. summarize documents, extract structured data, review code. Fire it off. Sleep. wake up to results. - One shot deep reasoning - Silent code audit loops, you write and test, the 31B reviews diffs and flags issues in the background between your sprints - Any workflow where output quality > output speed A few weeks ago, nobody was running a 30B+ dense model on a single consumer GPU with 8 GB VRAM. At all. Now we're doing it on an Intel i7-H gaming laptop with a NVIDIA RTX 4060, thanks to llama.cpp + QAT quants + MTP speculative drafting. Google DeepMind said the Gemma 4 31B targets "consumer GPUs and workstations." They were not exaggerating. The hardware bar to run serious frontier class models locally keeps dropping. the tools are here. the models are here. you just have to be willing to abuse your laptop a little. what workflows would you actually run on a local 3 tps 31B dense model? genuinely curious. drop it below.show more