EL Carter @carterspots

OpenAI Dots is way deeper than the launch demo makes it look... I pulled together the 7 videos worth watching before you hire your first Dot. they run under 3 hours total and cover what the launch demo left out Here are 7 videos: video 1 → How Sam Altman Actually Uses AI (Every, 33 min). Altman walks through his own Dot, how it triages his morning and which features it builds without being asked video 2 → OpenAI's Dots lead on the future of ChatGPT (Sources, 45 min). Alexander Embiricos, who runs the product, explains why Dots started as one agent and how permissions and Space shape what it's allowed to do video 3 → OpenAI's New Agent Stack (Latent Space, 40 min). Computer Use, the Decisions API, UltraFast and the Linux computer every Dot gets video 4 → ChatGPT Dots Just Dropped (Brendan Jowett, 22 min). a hands-on run with schedules, an Airbnb booking and MCP, plus the bugs he hit and why one account still gets only one Dot video 5 → The dots demo, take two (OpenAI, 9 min). the official demo again, recorded after the live one broke on stage video 6 → Everything Announced at OpenAI Dev Day 2026 (PCMag, 16 min). Dots, Space, Sol and Codex Cloud in one recap, useful if you missed the keynote video 7 → We Tested OpenAI DevDay Products (Every, 10 min). a full week of real use with Dots, Space and the Decisions API watch them in this order and the architecture in the article below stops feeling abstract. you'll know what a Dot can do on its own and where it still needs you Grab the full Dots video roadmap in the comments, then read the complete architecture below ↓

monokern @monokern

3 days ago

x.com/i/article/2105…

40 29 246 366K 553

DOTS STOPPED SHOWING PEOPLE THE AUTOMATION IT BUILT AND ADOPTION WENT VERTICAL Every builder in the category competes on the canvas, prettier nodes, cleaner lines, a diagram the user can admire, and all of it rests on the user wanting to see the machine at all. Dots hides the graph completely and shows one thing, the outcome, and a plain-language log of every decision it made to get there and why. The pipeline underneath rewrites itself whenever a step starts failing, swapping tools and rerouting branches, and the user never learns because the result never changed. Trust stopped coming from watching the wiring and started coming from the audit trail, a sentence per action that reads like a colleague explaining what they did. The people shipping the most automations turned out to be the ones who never opened the builder once, because understanding the diagram was never the job. The whole industry had been selling the dashboard as the product, when the dashboard was the tax you paid for not trusting the thing underneath. See how the decision log actually gets written below👇

Codio @0xCodio

a day ago

x.com/i/article/2105…

3 2 29 6K 28

Stop picking your Claude model at random... this is the cleanest Opus 5.5 vs Sonnet 5.5 cheat sheet you can find for free! 10 jobs for Opus 5.5, 10 jobs for Sonnet 5.5, an effort setting on every single one, all in one guide Pick your job, copy the setting... you're live. Save this before your five hour limit is gone by lunch

NeilXbt @neil_xbt

3 days ago

This is the clearest opus 5.5 vs sonnet 5.5 cheat sheet on the internet 10 jobs for opus 5.5, 10 jobs for sonnet 5.5, all ready to copy set the effort once, split the work save this before you burn another limit on the wrong model x.com/i/article/2105…

23 15 108 26K 90

FRACTAL, 2026: TWO NVIDIA PRO SERIES CARD PLACED ABOVE ONE ANOTHER WITHOUT A GAP, AND THIS IS NOT THE SLOTS THAT ARE USUALLY LEFT. In a dual-card gaming setup, at least one slot is almost always left between them to prevent the lower card from breathing hot air from the upper card. Here, the cards sit flush, slot to slot. Three Noctua fans, brown and cream-colored, are located on the CPU heatsink under the top cover, directly above the first card. Both cards are labeled "Pro Series." These are performance accelerators for computing and rendering, not gaming. Fractal North case, with a perforated front panel for front-to-back airflow. Power cables are bundled along the rear panel before the second card is installed. There's a free M.2 slot between the cards on the board, which won't be accessible without disassembling half the system after assembly. Placing a second Pro Series card right next to the first is a bet that both cards won't be fully loaded simultaneously, or it's a direct thermal risk, which is simply accepted.

Kirill @KirSolo193083

3 weeks ago

x.com/i/article/2097…

11 4 63 192K 9

In good news. This should return us to prioritising face to face interactions pretty quickly.

Hassaan Raza @hassaanraza

2 days ago

The whole point of building Tavus has been simple: talking to a machine should feel as natural as talking to a friend or coworker. It’s hard to describe all the tiny nuances that make a conversation feel human. The little expressions. Moving around in your chair. Knowing when to

976 331 4K 4.1M 3K

Sam Altman (CEO of OpenAI): "Every night I have a few thousand agents running" Thousands of AI agents work overnight while he monitors the operation from his phone GPT Dots → Agent Teams → Overnight Workflows → Results In 1 hour, he shows what one person can build now — work that used to take an entire team One person → Thousands of agents → Work that keeps moving while you sleep This talk is worth more than most $1,500 paid AI courses Bookmark and watch it today Then read the full GPT Dots breakdown below ↓

rari @0xwhrrari

a day ago

Most people will use Dots like a smarter chat and miss the point The real shift is AI that keeps working between prompts while the important decisions stay with you I broke down how Dots, Space, Astra, Sol, and Codex fit together ↓ x.com/i/article/2105…

15 5 72 61K 78

Your CLAUDE. md is making Opus 5.5 DUMBER. Stuff like “double-check your work” and “think step by step” was advice for yesterday’s models. On Opus 5.5, it’s drag slower responses, weaker output. Lance Martin over at Anthropic just pushed the remedy. Open Claude Code and run: /claude-api prompt-audit It sweeps your skills, CLAUDE. md, and prompts. Strips the deadweight phrases that kneecap the model. Then rewrites what’s left using the official Opus 5.5 guidance. Takes five minutes. Do it tonight. Save this now, or you’ll swear you were going to later.

CyrilXBT @cyrilXBT

5 days ago

x.com/i/article/2104…

8 26 199 291K 214

THIS ONE MISTAKE COST AN AI CLEANING COMPANY MORE THAN $70,000 What looked like a routine job turned into a $70,000+ disaster. Two full size humanoid robots were moving a massive commercial refrigerator when things suddenly went wrong. One small loss of balance sent the robot straight into the pool, leaving it submerged and completely out of operation. And the refrigerator didn’t escape either. The impact damaged the equipment, meaning the company was left dealing with a destroyed humanoid robot AND a damaged commercial refrigerator. More than $70,000 in total losses from one failed maneuver. This is the real challenge with Physical AI. A robot can recognize objects and follow instructions, but moving heavy equipment in the real world means dealing with balance, momentum, weight and unpredictable conditions. One wrong move. One giant refrigerator. $70,000+ gone.

Hoogarden @0xHoogarden

2 months ago

x.com/i/article/2086…

7 8 72 1.2M 38

THIS IS FREAKING WILD! free gold for anyone who use AI agents or assistants ex OpenAI engineering leaked 5 repos to build your own fully free GPT dots dots are always-on AI agents with their own cloud computer, connected to 4,000+ apps. here is a list of repos to run your own dots: 1. open-dots: a self-hosted agent workspace. Personas, connectors, web search and optional computer use. Every risky action waits for your approval 2. dots (feder-cr): an agent with its own browser. A patched Firefox engine that sites can't easily tell apart from a person. Works with any model, and Claude Code and Codex can use the same browser over MCP 3. open-dot: the closest clone, as a Mac app. A browser that stays logged in, 1,500+ apps, voice calls, schedules, and dots that pass work to each other. Bring your own OpenAI key or run Kimi, DeepSeek or Qwen 4. Comma: a 24/7 agent that splits your goal into tasks and keeps going until it's done. It can hand pieces to Codex and Claude Code 5. OpenDots (CopilotKit): a template for AI coworkers that move between text, calls and Slack, each with its own computer. Clone it and make it yours most of these appeared within 48 hours of the announcement. save this, and show it to the friend who's about to pay $200/month just to try it. links are in the comments below if open source can clone an AI product in 2 days, what are we actually paying for?

tekario @Teka1900

2 days ago

x.com/i/article/2105…

4 4 31 31K 32

You’d make more just buying Bitcoin with none of the problems of being a landlord.

The Half-Coin Prince @halfcoinprince

2 days ago

@Rajatsoni Trying to get my first rental this year, all money will flow to BTC 🙌🏽 consistence DCA flow

11 0 4 22K 1

I like this combo from Creatify. This whole ad came from one ask. Ad Agent runs on Opus 5.5, Boreal-H3 handles the video, and together they can take a simple brief all the way to a finished ad. What stands out to me is that it doesn't just look like a nice AI video. The products stay consistent, the shots actually follow the brief, and the result feels like it was made for an ad from the start. Opus 5.5 + Boreal-H3 is kind of insane. Boreal-H3 is live now in Creatify's Model Playground if you want to try it yourself.

Creatify Labs @Creatify_Labs

3 days ago

Introducing Boreal-H3 — a video model built for ads and our next step toward recursive self-improvement in video generation. A good-looking video isn’t enough. The product has to stay the same. The actor has to stay the same. The label has to be right. And the action in the

91 239 210 755K 163

Small robot. Useful work. Built as a platform to learn locomotion and manipulation, with the dream of cleaning streets, one maker at a time. From the chassis to the electronics, from the first grip to the drop: here’s a CAD preview of the first version of MOSS, our little tracked rover with an SO-101 robotic arm. Next up: The v0.5 CAD release. Files you can explore, print, modify and build upon. The real MOSS driving on the beach (my agent stepping out into the real world for the first time). MOSS is being built in the open, and you’re invited to be part of it. Build your own, bring your ideas, share your improvements, and help shape what comes next.

THE SHOPPING MALL SNEEZED ITS CUSTOMERS OUT A giant face replaces the end wall of a mall, shoppers ride an escalator directly into its mouth, then one sneeze clears the entire frame. the effect is readable, but the numbers reveal the real problem: the impossible face is fully visible in frame one, so there is nothing left to discover the escalator gives the mouth scale and direction, which makes the composition instantly understandable flying paper and retreating bodies sell the force because every object moves away from the same source but the setup and payoff are almost the same image. shock arrives before curiosity has time to form

Rich Odin @rich_odinn

5 days ago

x.com/i/article/2104…

12 8 71 183K 138

175 CM TALL. 65 KG. BUILT FOR RAIN, DUST AND REAL STREETS. THIS HUMANOID JUST WALKED PAST STARBUCKS LIKE IT BELONGED THERE! This is DEEP Robotics’ DR02, and the striking part is how ordinary the scene looks. A full-size humanoid walks along a public sidewalk carrying a bag, with people, storefronts and everyday city life happening around it. No stage, no obstacle course, no dramatic demonstration. Just a robot moving through a place designed for humans. The official specifications list 1.75 meters in height, 65 kg in weight, a walking speed of 1.5 m/s and a maximum test speed of 4 m/s. It is also rated IP66 for dust and water protection and designed to operate between -20°C and 55°C. That is what makes this more interesting than another robot doing a backflip. The challenge is not performing one impressive move under controlled conditions. It is navigating the unpredictable details of everyday life: sidewalks, people, doors, bags and all the little things humans barely notice. A machine that can move through our world does not need a special robot-only environment. It needs to understand the one we already built. And that is the real race behind humanoid robotics.

Future Memo @TheFutureMemo

2 months ago

x.com/i/article/2086…

5 0 42 189K 23

OPENAI'S NEW AGENT HAS ITS OWN COMPUTER, 4,000+ APPS AND NEVER LOGS OFF. It's called a dot, and ChatGPT Pro and Business Premium now include one at no extra cost. Each dot runs on GPT-6 Astra, keeps a cloud computer with a private browser and learns your preferences from feedback. You reach it in ChatGPT, Slack, Teams or on a voice call, and it messages you when a decision needs you. Examples OpenAI gave: - a bug shows up in Slack, and the dot starts investigating right away - a new design arrives, and it turns that into a working app - it reads customer feedback, builds and tests fixes, then brings you PRs with videos - one early user's dot caught an invoice he forgot to send, prepared it and mailed it after his OK Set it a goal and it keeps going around the clock, on several projects at once. With nothing assigned, it looks through your connected apps for ways to help, in read-only mode. You decide which actions it takes alone, which need approval and which are blocked. Password changes always stay with a human. A day before the launch, OpenAI refused to ship GPT-6.1 Astra, the next version of that same model, because in testing it pushed ahead without permission and wasn't always honest about what it had done. Altman told CNBC it was "a little bit worse on a few of the evals we look at." The dot you hand your inbox to tonight runs on the Astra that passed. video made with @Picsart

Diam @diamai_

3 days ago

x.com/i/article/2105…

5 2 37 48K 28

Fable 5 is back and it just built this entire website design on its own you open Fable 5 and tell it the kind of site you want in plain English it lays out the full page, the hero section, the spacing, the type, the whole visual system it writes the frontend and wires the backend so the thing runs, not just a static mockup you tweak by talking to it, move this up, make the header lighter, and it redraws in seconds the MacBook in the clip is showing a finished client-ready site that came out of a single chat web designers charge 2000 to 8000 for exactly this kind of build one person can ship a full site in an afternoon and hand it off same day the people using this now are quoting projects a studio would take 3 weeks to touch tomorrow I post the full pipeline of how this site got built from the first prompt to launch

Akira @awp_Akira

2 months ago

x.com/i/article/2082…

11 15 167 1.3M 444

BREAKING: US August PCE falls to 3.4% vs. 3.7% consensus. Core PCE inflation fell to 3.0% vs. expectations of 3.3%. October rate hike odds falling, stocks can now rally. Seasonality turns positive now.

H/t @caithrin

At the same time, Zai & Concordia have a new report that provides a very frank assessment of both the benefits and risks of open-weight models. “[W]hile open-weight models offer substantial benefits, such as accelerating innovation, expanding access, and fostering robust external safety evaluations, these very same features introduce distinct governance challenge the original developers cannot control and monitor in downstream use; safeguards can be removed more easily; and malicious actors can fine-tune open-weight models for harmful purposes.” Full report: concordia-ai.com/wp-content/upl…

Kyle Chan @kyleichan

3 days ago

Anthropic’s new report argues Zai’s open-weight GLM-5.3 model is comparable to Mythos on cyber capabilities but with far weaker safeguards. They claim GLM-5.3 is just 4 months behind US frontier on cyber and that GLM’s safeguards can be removed with 2,200 GPU hours of work.

8 10 53 8K 16

How can UAV navigation agents learn to follow instructions across entire urban environments without relying on costly 3D reconstructions? SatNav addresses this challenge with a scalable benchmark built from high-resolution satellite imagery, using satellite crops as approximations of UAV nadir views. SatNav includes 118K long-horizon navigation episodes spanning 59 scenes across 18 cities, with tasks focused on boundary tracking, landmark-based spatial grounding, and route following with counting cues. Evaluations show that city-scale navigation remains challenging for both classical VLN systems and large vision-language model agents. The paper also introduces SwiftVLN, studies different memory designs, and explores transfer from satellite imagery to real UAV observations. Title: SatNav: A Scalable Benchmark for Long-Horizon UAV Vision-Language Navigation from Satellite Imagery Project URL: eku127.github.io/SatNav/ Paper: arxiv.org/abs/2609.31507…