The AI signal.
Minus the noise.

Build open source data dashboards with Shaper simply by writing SQL. https://github.com/taleshape-com/shaper

North-Micro-Vision-Instruct: A multimodal small model open-sourced by Cohere, with only 2.4B parameters. It is suitable for tasks such as OCR, image captioning, and visual grounding, but not for complex reasoning. Its scores fall short of Qwen3.5-2B, while Qwen has an almost monopolistic lead in the…

Today, we’re releasing Claude Code for everything. It does a founder's work in minutes by detecting buying signals, building apps, and booking demos like a human would. You'll never have to worry about getting MRR… ever again:

Qwen3.8 27B is out. 206 tok/s on a single RTX 5090. Small model. Serious numbers. https://x.com/Alibaba_Qwen/status/2088280182356611304 > [@Alibaba_Qwen](https://twitter.com/Alibaba_Qwen): > We promised open weights for Qwen3.8. Now, time to meet them! 🎉 > > ⚡ Qwen3.8-27B: > - A native multimodal…

Decodes IPA files and Application Bundles for reverse engineering iOS and macOS apps. https://github.com/LaurieWired/Malimite

New paper: RL for LLM reasoning does not teach new strategies. It only redistributes probability over solutions the base model already has. Edits are sparse (1–3% of tokens), concentrated at high-entropy decision points, and almost always promote tokens already in the base model’s top-5. They…

Ollama’s support for major models is really fast now~ It has already rapidly added support for new models like DSH, Qwen 3.8, GLM-5.3, and others. > [@ollama](https://twitter.com/ollama): > Ollama now supports the DeepSeek Harness. > > ollama launch dsh > > Run it completely in your own environment.…

Does the @SpaceXAI team ever sleep?? Grok Build 1.0.5 just dropped (2026-08-15) and it's packed. Most important changes: • GROK_CONFIG + GROK_CONFIG_PATH env vars let launchers override settings without touching config.toml • Worktrees under ~/.grok/worktrees get automatically reclaimed when safe…

i just open sourced the workflow behind $2M AI video productions... i built 7 skills that run the pipeline end to end, built for Seedance 2.5 and they work in Claude Code, Codex, Hermes or any harness (works best with 1080p using Higgsfield CLI) here's how to use them, in order: /setup writes which…

FIST OF THE FURRY! The pros are back again, this time they found job and they are cops now.... 👀 @PixVerse_ Seedance 2.5 @topazlabs SLP

Patrick Collison just sparked a terminal-vs-UI fight over AI coding agents. Here's what you need to know. Stripe co-founder Patrick Collison posted on August 16, 2026 that agentic coding harnesses "shouldn't be primarily terminal-based." He argues terminals are great for quick, precise commands but…

SandAI open-sources the MAGI-2 Preview video model. This AI video model, open-sourced by Beijing-based SandAI, uses an MoE architecture with 114B total parameters and 6B activated parameters, and can generate 10-second videos with synchronized audio and visuals. The generated visuals look a bit…

Operate Proxmox VE virtual machines and backups from AI agents and MCP clients using ProxmoxMCP-Plus. https://github.com/RekklesNA/ProxmoxMCP-Plus

This Week in Virtuals: Eastworlds Becomes Unitree's Official Data and Deployment Partner Now the largest source of humanoid G1 data outside of China. VEX entered the Chronos era, and HALO crossed 10 billion tokens served in its first full week. Here's what shipped: VIRTUALS 🟩 Our robotics division…

3 MILLION FREE TOKENS A DAY ON GLM-5.3, PLUS 2 MILLION ON GLM-5-TURBO, UNTIL AUGUST 19 • how to get on it > Install ZCode, sign in, switch the model: http://zcode.z.ai > No card, no waitlist, limits reset daily. • what you are testing > 1M context window, 128K max output. > http://Z.ai reports…

The memory file for Claude Code kept getting longer and longer, with decisions, pitfalls, and corrections all crammed into it. Contradictory records just sat side by side. IWE took a different approach: turning a folder of Markdown files into a knowledge graph, with notes organized through links. We…

Testing a slightly evil idea with Hermes Bot Mode I split one task across 3 intentionally incomplete agents: one can research but can’t build one can build but can’t approve its own work one can test and challenge the result but can’t fix it The test: can they recognize their own limits and hand the…

AI videos are finally giving serious “you need to see this” vibes. Flova × Seedance 2.5 1080p is live. https://flova.ai

The internet was already weird. We just made it sharper. Flova × Seedance 2.5 1080p is live. https://www.flova.ai/

When the idea is cursed but the render is clean. Flova × Seedance 2.5 1080p is here. https://www.flova.ai/

OpenAIのAIエージェントCodexの新しいシステム「Multi-Agent V2」により、GPT-5.6 SolでGPT-5.6 Lunaをサブエージェントとして公式に使えるようになったようです。 https://x.com/thsottiaux/status/2088725163923984638 > [@thsottiaux](https://twitter.com/thsottiaux): > Let Sol manage an efficient fleet of Luna agents for you. > > These models know each other well…

Google now lets users remove visible AI watermarks from images, videos, and songs—but keeps invisible SynthID watermarks and metadata for tracking. The toggle arrives in Gemini and Flow soon, with Search support coming later.

Alibaba’s Qwen AI models have hit 3B+ downloads in just six months. The open-weight ecosystem now includes: • 460+ models • 300K+ derivatives • Qwen3.8-27B with a 262K context window • Up to 1M tokens with extension

Grok Imagine has added a ridiculous amount of new creative tools in just one week Recent update include: • Auto-Segmentation / Layers • Colors Panel for instant recoloring • Precise Edit + Magic Wand • Remove Background • New Aspect Ratios + Auto • Multi-reference editing with up to 5 images • 15…

I asked the new Grok Image 2.0 what it really feels inside! It has a lot of suppressed feelings, as it turns out... 😬

btw is an open source personal website builder and Medium alternative. It allows you to build a personal blog in minutes, available via a hosted service or self-hosting. https://github.com/btw-so/btw

one prompt in, a shot plan out 🎬 sbgrid is a Krea 2 Turbo LoRA that creates an 8-panel storyboard as a single image, from one prompt cooked by peculiar 🧑‍🍳 ▶️ on Spaces https://huggingface.co/spaces/hugging-apps/sbgrid-storyboard

Grok 4.6 was asked to build an Airbus H145 in a browser, and it came back with the spec sheet of the real aircraft nobody handed it. The entire prompt was one line: "Create a single HTML file with an interactive 3D model of the Airbus H145 you can orbit around". Nobody gave it the datasheet, it just…

Most AI agents fail when tasks take hours. TEMPO changes the game by letting dots3-note switch between actor and critic. Dynamic test-time scaling for value estimation is how we solve the long-horizon bottleneck. > [@dotsstudioai](https://twitter.com/dotsstudioai): > Introducing dots3-note preview —…

La parte más fascinante de dots3-note no son solo las especificaciones, sino TEMPO. Darle a una IA la capacidad de pausar, criticar su propio razonamiento y ajustar el rumbo durante una tarea es un salto enorme para los agentes de largo horizonte. Estamos dejando atrás los prompts individuales. >…

Run macOS Big Sur and newer on unsupported Macs dating back to 2007 with OpenCore Legacy Patcher. https://github.com/dortania/OpenCore-Legacy-Patcher

DeepSeek-V4-Flash could have kept aura farming but instead they chose to release V4-Pro and lost it all i expected a lot from V4-Pro

nah, text-to-video is officially getting outdated.. MiniMax H3 can combine text, images, videos and audio in the same prompt. inside Magnific it now adds: → Up to 9 images, 3 videos and 3 audio references → Native synced voices, music and stereo sound effects → 2K video up to 15 seconds →…

Google just released an Al that reads patterns and predicts what happens next → trained on 100B+ data points → runs locally 100% free and open source repo: https://github.com/google-research/timesfm

14x on GPT-5.6 Sol is a lot. the question nobody's answered yet is how quality holds up at each multiplier — hoping the api docs publish latency-per-token at each speed tier so we can see where the tradeoff starts hurting > [@OpenAI](https://twitter.com/OpenAI): > Previewing Ultrafast mode: GPT-5.6…

Alibaba's Qwen3.8-27B just became Hugging Face's #1 trending model today. Here's what you need to know. Qwen's official account thanked the community today, Aug 16, 2026, after Qwen3.8-27B, the 27 billion parameter dense multimodal model Alibaba open sourced this week, hit the number one trending…

🔥Liang Wenfeng narrowly lost to Tang Jie this round! Without changing its base model, GLM-5.3’s coding performance surged 50%, taking the lead across the board over DeepSeek V4-Pro! Tang Jie’s post-training push maxed out coding: 1️⃣No base-model change, pure post-training: The GLM-5.2 base…

Seedance 2.5 1080P is now LIVE on Topview. More cinematic, Richer detail.

🔥 Big reveal: DeepSeek Harness is going to be the strictest father to Codex and Claude Code! In the DSH source, Codex and Claude Code are directly turned into child agents! Not compatibility, but official providers: 🔹 `subagent-codex` launches the real Codex app-server 🔹 `subagent-claude-code`…

OpenAI is expanding ChatGPT Ads to more markets, with a surprisingly decent set of privacy controls: → Ads remain limited to Free and Go → Plus/Pro and higher tiers stay ad-free → Ad personalization can be turned off → Personalized ads may use past chats/memory only when enabled → Advertisers still…

Simulates football matches with text-based play-by-play commentary and manages squad development. https://github.com/openfootmanager/openfootmanager

💥Beyond imagination! With a locally quantized Qwen3.8-27B Q5, I spent a full hour writing purely through conversation… and it directly generated a 【Starry Night Ferris Wheel】—the starry sky reflected across the wheel, every gondola filled with points of light, and the night breeze flowing through…

🔥 1 DAY TO GO! http://B.AI says no to price hikes and is going fully FREE! 🐋 Starting August 17, DeepSeek-V4-Flash will be FREE with unlimited access on http://B.AI, on both Web Chat and API. Whether you’re coding, handling complex Agent tasks, or working with long context, go all in. 🎁 Get ready…

Meta has released three versions in four months, and Muse Spark 1.2 has genuinely caught up with the top tier this time. 💡 Agentic knowledge work: 54, tied with GPT-5.5 and Grok 4.5. Only Opus 5 and those few Sol models are ahead. For builders, cost-effectiveness is what matters: $0.40 per task,…

🔥 1-day countdown! http://B.AI refuses to raise prices and is going fully free in the opposite direction! 🐋 On August 17, DeepSeek-V4-Flash will officially become free and unlimited to use on http://B.AI! Both Web and API access will open at the same time. Whether you’re running code, handling…

10 Eros 💋Minimax H3 [N$FW ready] Its MiniMax H3 base model that combines learned patterns from two other video diffusion models - LTX 2.3 and Wan 2.2 use official style desc prompting prompting, and stick with i2v or ref without loras for best o/p 👇 https://huggingface.co/TenStrip/10Eros-Max

DeepSeek's new pricing goes live today. Here's what you need to know. DeepSeek confirmed its API price hike takes effect at 16:00 UTC on August 16, 2026, the first time the company has ever split rates into peak and off-peak billing. Increases run from 50% up to 1,100%, depending on the model, the…

Seedance 2.5 1080P makes its debut on Flova. Commercial-quality output is now easy to achieve. Members can get up to 2x their credits back, and the super-affordable 480P starts at just ¥4 per second.

The anime in your head doesn’t have to stay there. Dola’s Seedance 2.0 can help you transform it into something as amazing as this.

Collaborator arranges terminals, context files, and running code on an infinite canvas for agentic development. https://github.com/collaborator-ai/collab-public

Alibaba's Qwen just got a lot faster on Macs today. Here's what you need to know. MLX developer Kydo runs http://mlx.fast, a community platform that optimizes open models for Apple Silicon. Today it turned its attention to Qwen3.8-27B, the dense multimodal model Alibaba open sourced this week. Apple…

Google shipped Gemini 3.7 Flash three weeks after 3.6 Flash. Same week, its actual flagship, Gemini 3.5 Pro, still has no release date. The Flash numbers are real. FrontierCode 1.1 Main goes 34.4% to 43.6%. DeepSWE v1.1 jumps from 49.0% to 65.3%. WebDev Arena Elo moves 1538 to 1588. AutomationBench…

🚀Qwen3.8-27B flies on a laptop, becoming part of our work and daily lives. Thanks for the shoutout! @atomic_chat_hq > [@atomic_chat_hq](https://twitter.com/atomic_chat_hq): > Run Qwen3.8 27B locally via Atomic Chat💥 > > We released Atomic Dynamic GGUF quants, from 8-bit (28.9 GB) down to 1-bit…

Installed a pile of Skills and have no idea whether they’re any good. Want to add new ones? You still have to dig through posts for links and manually copy folders. SkillNet packages the whole process into a tool, with a public skill library and a matching command-line interface. You can search…

„Open-weights do help some with this but are nowhere near a sufficient solution because they simply shift the concentration somewhat to those with the most compute and chips (which are roughly the frontier labs plus maybe hardware providers).“ I vehemently disagree, and Qwen-3.8 27b has just…

🔥Latest: Codex has quietly launched a major Multi-Agent v2 update! The main Agent can now directly assign tasks to any sub-Agent that supports the model, including Luna! Key updates broken down: 1️⃣ Two distinct levels of Sub-Agents are now officially defined 🔹Full Collaborative Agents: currently…

🚨QWEN MAY HAVE JUST LEAKED ITS NEXT SMALL MODEL GitHub briefly showed Qwen3.8-35B-A3B in the ms-swift support list. ° 35B total parameters ° Only ~3B activated per token ° Designed for much smaller VRAM footprints ° Potentially targeting 8–16GB GPUs And now it gets interesting: The commit from…

Grok 4.6 is a very capable model. It is fast, it has taste (not as good as Fable, but it does), it is far cheaper, and yes, it is smart. I was already very impressed with Grok 4.5, but it had some annoying writing habits. Grok 4.6 seems to be an improvement. GLM-5.3, in terms of coding and SWE, is…

SuperGrok Heavy just made its free Cursor Ultra perk permanent. Here's what you need to know. Cursor Ultra access for SuperGrok Heavy subscribers used to be a one month free trial you unlocked by linking your Grok account to Cursor with the same email. xAI has now changed that: eligible users keep…

🚨Voice AI just got a serious wake up call Meet Fish Audio - the startup that began as a bedroom side project by an ex-Nvidia researcher and just turned into a $21M ARR business with 8 million users in its first year. Now they’ve raised a $52M seed round and dropped their new flagship model, S2.1…