Pinned Tweet
We live in bubble on this app.
A lot of us are running agents, building workflows, creating skills, comparing harnesses, connecting tools, and etc..
Go ask one of your friends what Kimi or GLM is. They will probably look at you like a madman.
OpenAI’s consumer research found that 49% of ChatGPT messages are still people asking for advice or information.
Another 40% are requests to complete a task.
Meanwhile, 70.2% of sampled Codex users had assigned at least one task estimated to take a person more than an hour.
I’m sure your timeline makes it feel like everyone is running agents and orchestrating workflows.
They are not.
Do not confuse being surrounded by early adopters with being late.
You are early. Keep learning and keep building.
A lot of us are running agents, building workflows, creating skills, comparing harnesses, connecting tools, and etc..
Go ask one of your friends what Kimi or GLM is. They will probably look at you like a madman.
OpenAI’s consumer research found that 49% of ChatGPT messages are still people asking for advice or information.
Another 40% are requests to complete a task.
Meanwhile, 70.2% of sampled Codex users had assigned at least one task estimated to take a person more than an hour.
I’m sure your timeline makes it feel like everyone is running agents and orchestrating workflows.
They are not.
Do not confuse being surrounded by early adopters with being late.
You are early. Keep learning and keep building.
176
245
3,126
106,758
Oh brother
NEW: Jensen Huang from Nvidia admits to infidelity with his wife with a Chinese spy saying in a memo to staff:
“I was unfaithful to my wife. My actions hurt her and our children. There is no excuse for what I did. I am deeply sorry. I love our country and our company, and I am Nvidia and supporting America’s national security."
He should have just gone with the "honeydicking" angle.
1
0
2
240
I own my house, and I still don’t think owning one automatically means you’re doing better than someone who rents.
Some people want roots. Some people want the freedom to leave.
Why do we only treat one of those as success?
Some people want roots. Some people want the freedom to leave.
Why do we only treat one of those as success?
13
3
26
833
So since y’all are saying AI winter is upon us, does that mean GPUs and other hardware are finally getting cheaper? 😭
Asking for a friend.
Asking for a friend.
13
0
23
766
I think @arcee_ai deserves more credit in the open model conversation.
Yes, I know we have NVIDIA’s Nemotron, Google’s Gemma, Ai2’s OLMo and others. But Arcee should come up more often too.
They’re an American lab training their own foundation models from scratch.
Trinity Large is roughly 400B parameters, with weights available for people to run and build on. They’re also behind MergeKit, which lets developers combine models.
Qwen, Kimi and DeepSeek get plenty of attention, deservedly. I just think more people should know what Arcee is building.
Anyone here actually using their models?
Curious what your experience has been.
Yes, I know we have NVIDIA’s Nemotron, Google’s Gemma, Ai2’s OLMo and others. But Arcee should come up more often too.
They’re an American lab training their own foundation models from scratch.
Trinity Large is roughly 400B parameters, with weights available for people to run and build on. They’re also behind MergeKit, which lets developers combine models.
Qwen, Kimi and DeepSeek get plenty of attention, deservedly. I just think more people should know what Arcee is building.
Anyone here actually using their models?
Curious what your experience has been.
3
1
10
508
Saturday AI roundup 👇
A mystery coding model, local AI on your Mac, new agent tools and a few creative builds.
Here’s what caught my attention today, November 7:
1. Qoduck — @qoder_ai_ide
Qoder is promoting a stealth model with its origin still undisclosed. Individual users get a limited daily free allowance, including free accounts. Available to try, but we don’t know which model is underneath yet.
2. Result for Mac — @tryresult
Introduced a Mac app with computer use through local models, connections to private inference hardware and support for open-weight cloud models. You can also access models running on your Mac from your iPhone.
3. DECIDE — @namanjn1998 / @bolna_dev
Tazo Labs and Bolna introduced a benchmark for the decisions voice agents make during calls: when to speak, wait, switch languages or use a tool.
4,867 questions from 1,081 real call clips. Getting the transcript right is one thing. Knowing what to do next is another.
4. Layro — @Ksparth12
A free React component library that coding agents can search and install through MCP or a CLI. Preview components, publish your own and get the source files directly in your project to edit.
5. Cortad — @moalsufi
Introduced a way to give coding agents context about how real people use your AI app. Through its MCP integration, the agent can consult that behavior while changing a feature instead of relying entirely on its assumptions.
6. Kathamo — @devshowcse
An early open-source workspace for running coding agents across multiple projects. Tracks which agents are working, blocked or finished, with notifications when they need your attention.
7. terminal-cowork — @leakorsawe
A free, open-source Windows/WSL workspace combining Claude Code sessions, an agent Kanban board, a shared browser and live previews. You and Claude work in the same browser, with a control to pause the agent and take over.
8. Reissue — @winsznx
A hackathon build for catching mismatches between invoices and packing lists.
Three agents check documents in Slack, request corrections from the supplier through Telegram, then recheck the files before preparing a package for the customs broker.
Early project, but a very specific problem people actually deal with.
9. Clueprint — @mariojankovic
Announced free public access to a Mac app that shows what coding agents leave behind: worktrees, dev servers, Docker stacks and caches.
Helps identify finished work for cleanup while leaving uncertain items for you to review. Running multiple agents can leave a lot more than code behind.
10. Tool Directory AI Skills — @VamsiBuilds
Launched a section with over 1,000 agent skills for Claude Code and Cursor. Browse by design, testing, databases, DevOps and more, sort by installation counts and copy the install commands.
11. An iPhone turned into a Mac mouse — @coolerwave
Shared a Claude-built project that turns an iPhone into a motion-controlled pointer or a trackpad when placed on your desk.
The builder says it runs over Bluetooth without servers and is free and open source. Point your phone at the screen, tap to click and hold to drag.
Which of these would you actually use?
A mystery coding model, local AI on your Mac, new agent tools and a few creative builds.
Here’s what caught my attention today, November 7:
1. Qoduck — @qoder_ai_ide
Qoder is promoting a stealth model with its origin still undisclosed. Individual users get a limited daily free allowance, including free accounts. Available to try, but we don’t know which model is underneath yet.
2. Result for Mac — @tryresult
Introduced a Mac app with computer use through local models, connections to private inference hardware and support for open-weight cloud models. You can also access models running on your Mac from your iPhone.
3. DECIDE — @namanjn1998 / @bolna_dev
Tazo Labs and Bolna introduced a benchmark for the decisions voice agents make during calls: when to speak, wait, switch languages or use a tool.
4,867 questions from 1,081 real call clips. Getting the transcript right is one thing. Knowing what to do next is another.
4. Layro — @Ksparth12
A free React component library that coding agents can search and install through MCP or a CLI. Preview components, publish your own and get the source files directly in your project to edit.
5. Cortad — @moalsufi
Introduced a way to give coding agents context about how real people use your AI app. Through its MCP integration, the agent can consult that behavior while changing a feature instead of relying entirely on its assumptions.
6. Kathamo — @devshowcse
An early open-source workspace for running coding agents across multiple projects. Tracks which agents are working, blocked or finished, with notifications when they need your attention.
7. terminal-cowork — @leakorsawe
A free, open-source Windows/WSL workspace combining Claude Code sessions, an agent Kanban board, a shared browser and live previews. You and Claude work in the same browser, with a control to pause the agent and take over.
8. Reissue — @winsznx
A hackathon build for catching mismatches between invoices and packing lists.
Three agents check documents in Slack, request corrections from the supplier through Telegram, then recheck the files before preparing a package for the customs broker.
Early project, but a very specific problem people actually deal with.
9. Clueprint — @mariojankovic
Announced free public access to a Mac app that shows what coding agents leave behind: worktrees, dev servers, Docker stacks and caches.
Helps identify finished work for cleanup while leaving uncertain items for you to review. Running multiple agents can leave a lot more than code behind.
10. Tool Directory AI Skills — @VamsiBuilds
Launched a section with over 1,000 agent skills for Claude Code and Cursor. Browse by design, testing, databases, DevOps and more, sort by installation counts and copy the install commands.
11. An iPhone turned into a Mac mouse — @coolerwave
Shared a Claude-built project that turns an iPhone into a motion-controlled pointer or a trackpad when placed on your desk.
The builder says it runs over Bluetooth without servers and is free and open source. Point your phone at the screen, tap to click and hold to drag.
Which of these would you actually use?
22
3
35
1,141
Getting the browser set up and signed into everything is an annoying part of using agents.
Hermes’ in-app browser will be able to inherit your cookies and passwords if you opt in. Planned for the next patch.
Fewer login screens between opening Hermes and getting something done.
Hermes’ in-app browser will be able to inherit your cookies and passwords if you opt in. Planned for the next patch.
Fewer login screens between opening Hermes and getting something done.
The in-app browser will also be able to inherit your cookies and passwords if you opt in like the agent's own browser tool does on next patch release
4
1
14
433
AWS is making it easier to run several smaller models on the same GPU.
SageMaker Inference now supports NVIDIA’s GPU partitioning, giving each model a dedicated slice of compute and memory on supported hardware.
Think of a document search app: one model creates embeddings, es, and another writes the answer. Each can have its own GPU partition instead of reserving an entire accelerator.
AWS also tested an H100 partition against a full A100 running Llama 3.1 8B. At sufficient concurrency, the partition delivered roughly 18–34% lower cost per million output tokens. At lighter loads, the A100 could be cheaper.
Useful for teams paying for GPU capacity their smaller models barely touch.
SageMaker Inference now supports NVIDIA’s GPU partitioning, giving each model a dedicated slice of compute and memory on supported hardware.
Think of a document search app: one model creates embeddings, es, and another writes the answer. Each can have its own GPU partition instead of reserving an entire accelerator.
AWS also tested an H100 partition against a full A100 running Llama 3.1 8B. At sufficient concurrency, the partition delivered roughly 18–34% lower cost per million output tokens. At lighter loads, the A100 could be cheaper.
Useful for teams paying for GPU capacity their smaller models barely touch.
3
0
10
431
Let me give you some AI statistics real quick.
• 29% more breast cancers detected. Sweden’s MASAI trial involved roughly 106,000 women. AI supported screening increased cancer detection while reducing radiologists’ screen reading workload by 44% compared with standard screening.
• Over 200 million protein structure predictions. AlphaFold made these freely available to researchers, helping them investigate how proteins work and their role in disease.
• Flood forecasts covering areas home to 2 billion people across more than 150 countries. Google’s AI-powered Flood Hub provides river-flood forecasts up to 7 days ahead, giving communities more time to prepare.
• 20% better data-center water efficiency in one year. AWS reports that its water withdrawals per unit of IT electricity fell to 0.12 liters per kilowatt-hour in 2025—a 20% improvement over 2024.
I’m barely scratching the surface of what AI is already helping people do, let alone what’s coming next.
Now, with the current political climate and changes in leadership, some politicians are pushing for moratoriums on AI data centers and pauses in AI development.
I understand the concerns about water use, pollution, electricity costs and the impact on surrounding communities. I get it...
But cooling technology and operating practices can already reduce water and energy use. Require companies to adopt effective improvements, publish their numbers and address pollution. Make them prove what’s improving at each facility.
I don’t support a blanket pause or broad slowdown. Politicians need to understand the benefits, the limitations and the engineering options before deciding how to restrict this tech.
AI will help us cure more cancers. It will accelerate the discovery of new treatments. It will help communities prepare for disasters and help engineers build more efficient infrastructure.
It will give more people the ability to turn an idea into something useful (It's already doing this).
Hold these companies accountable, for sure. But I want us to continue moving forward.
• 29% more breast cancers detected. Sweden’s MASAI trial involved roughly 106,000 women. AI supported screening increased cancer detection while reducing radiologists’ screen reading workload by 44% compared with standard screening.
• Over 200 million protein structure predictions. AlphaFold made these freely available to researchers, helping them investigate how proteins work and their role in disease.
• Flood forecasts covering areas home to 2 billion people across more than 150 countries. Google’s AI-powered Flood Hub provides river-flood forecasts up to 7 days ahead, giving communities more time to prepare.
• 20% better data-center water efficiency in one year. AWS reports that its water withdrawals per unit of IT electricity fell to 0.12 liters per kilowatt-hour in 2025—a 20% improvement over 2024.
I’m barely scratching the surface of what AI is already helping people do, let alone what’s coming next.
Now, with the current political climate and changes in leadership, some politicians are pushing for moratoriums on AI data centers and pauses in AI development.
I understand the concerns about water use, pollution, electricity costs and the impact on surrounding communities. I get it...
But cooling technology and operating practices can already reduce water and energy use. Require companies to adopt effective improvements, publish their numbers and address pollution. Make them prove what’s improving at each facility.
I don’t support a blanket pause or broad slowdown. Politicians need to understand the benefits, the limitations and the engineering options before deciding how to restrict this tech.
AI will help us cure more cancers. It will accelerate the discovery of new treatments. It will help communities prepare for disasters and help engineers build more efficient infrastructure.
It will give more people the ability to turn an idea into something useful (It's already doing this).
Hold these companies accountable, for sure. But I want us to continue moving forward.
4
0
16
488
Hybrid 🏃🏾 retweeted
This Echo Dot sat in a closet for at least 4 years.
Now she’s Cheryl, powered by Hermes, with her own voice, memory, etc.
Something cool/fun for the kiddos.
Feels good to “unfuck your old hardware,” as @NousResearch would say 🤣
Now she’s Cheryl, powered by Hermes, with her own voice, memory, etc.
Something cool/fun for the kiddos.
Feels good to “unfuck your old hardware,” as @NousResearch would say 🤣
My old Echo Dot run Hermes now, instead of Alexa.
I built a Home Assistant integration that connects it to a Hermes profile, with its memory and tools.
Code, setup skill and the hardware recovery notes are public.
github.com/majorminorlabs… @NousResearch
22
9
150
6,984
If you change your mind, AI needs to keep up.
Meta is rolling out updated models for Muse and Meta AI that better distinguish lasting preferences from temporary information and stale context.
On Meta’s own memory evaluation, Muse correctly used critical context on 90% of tasks, up from 76%.
Meta is rolling out updated models for Muse and Meta AI that better distinguish lasting preferences from temporary information and stale context.
On Meta’s own memory evaluation, Muse correctly used critical context on 90% of tasks, up from 76%.
Today we're bringing specialized models trained using Muse Wonder to @Muse and Meta AI, giving users access to some of Muse Wonder's frontier capabilities for complex tasks at work and in everyday life.
These specialized models include Muse Wonder's Active Memory and Active y aligned with your goals as tasks unfold:
→ Active Memory helps recognize when earlier information matters and distinguish what's still relevant from what's become stale
→ Active Focus helps recover from setbacks or find alternatives when constraints change
In Muse, the improved models show a 2.4x reduction in forgetting information critical to your goals and a 67% improvement in getting unstuck as constraints evolve. That means less repeating yourself and fewer dead ends when plans change.
In Meta AI, these improvements help it break complex problems into steps, ask sharper follow-up questions and adapt when you change direction.
Read more on our research blog: go.meta.me/4kCPcHW
1
0
9
397
That time of the day!!
AI raises, new models, robotics, video tools and more today.
Here’s what caught my attention👇
Raises
Darwin Robotics — @darwin_robotics
Raised $3.6M pre-seed to build robot technicians that assemble, commission and maintain infrastructure, starting with AI data centers. Led by Link Ventures, with YC, Liquid 2 and Rebel Fund participating.
Vensum Power
Raised €9M Series A for software-controlled power converters designed to reduce electricity losses and heat. AI data centers are among its target markets.
Physicl
$10M seed for synthetic training data for robotics and physical AI. Teams can change lighting, clutter and camera positions to generate examples of situations their models struggle with. Getting coverage today, though its own announcement is dated November 4.
Models and products
Cactus Mirage — @cactuscompute
A 22MB vision-language model for scene descriptions, object detection, OCR, segmentation and spatial analysis. Cactus reports 74ms execution on a Galaxy S26 NPU.
Manta — @EmperoAI
Announced a model claiming a 100-million-token context window, alongside published recall tests. Early access is coming; it isn’t publicly available yet.
ToyBlocks — @allen_ai
An open simulated tabletop for teaching and evaluating robots. Includes 85K demonstrations, two trained models, simulation assets and training/evaluation code.
Higgsfield Miracle — @higgsfield
Announced video post-production technology for separating footage into editable layers and making localized changes. Currently in private beta, with MCP access planned.
Alexandria SEC data — @firecrawl
Agent access to SEC information covering 8,000+ public companies. Research institutional holdings, insider transactions and company disclosures through tools like ChatGPT and Claude.
Conduit AI — shared by @asynchronous_x
Introduced a natural-language interface to its industrial automation platform, connecting work across factory data, systems and robots.
New model access and offers
GPT Image 2.6 in Runway — @runwayml
Now available in Runway, with GIF generation, SVG exports and precise edits that preserve character and product details. This is new platform availability.
LongCat-2.5 in Nous Portal — @NousResearch
One week of free access to Meituan’s multimodal model. The model launched earlier this month; the free access is today’s update.
Free FLUX 3 Max generations — @fal
Free generations through November 9, with text/image-to-video output up to 15 seconds at 720p. An access promotion, not a new model release.
Research, benchmarks and training
Cua Driver Bench — @trycua
A new benchmark with Scale AI testing whether desktop agents can finish tasks without interrupting the person using the same computer. That includes stealing keyboard focus or moving your pointer.
Frontier3DBench — @EinsiaAI
145 tasks across nine categories testing coding agents on 3D objects, scenes and interactive worlds in Blender and Unreal Engine 5. Evaluates appearance, structure and behavior.
Expanded speech evaluation — @huggingface
Egyptian and Nigerian English conversations join the Open ASR Leaderboard’s private evaluation data, expanding its accent coverage.
Harness adapter training — shared by @huggingface
Live reinforcement-learning runs training adapters for Liquid AI models and webAI’s Twil-9B to improve agent-task performance. Training is underway; these aren’t finished model releases.
AI-assisted cryptanalysis — @GoogleDeepMind
New research on human-guided AI agents uncovering weaknesses in legacy 1024-bit RSA systems. The team says these results do not affect modern RSA-2048 security.
AI raises, new models, robotics, video tools and more today.
Here’s what caught my attention👇
Raises
Darwin Robotics — @darwin_robotics
Raised $3.6M pre-seed to build robot technicians that assemble, commission and maintain infrastructure, starting with AI data centers. Led by Link Ventures, with YC, Liquid 2 and Rebel Fund participating.
Vensum Power
Raised €9M Series A for software-controlled power converters designed to reduce electricity losses and heat. AI data centers are among its target markets.
Physicl
$10M seed for synthetic training data for robotics and physical AI. Teams can change lighting, clutter and camera positions to generate examples of situations their models struggle with. Getting coverage today, though its own announcement is dated November 4.
Models and products
Cactus Mirage — @cactuscompute
A 22MB vision-language model for scene descriptions, object detection, OCR, segmentation and spatial analysis. Cactus reports 74ms execution on a Galaxy S26 NPU.
Manta — @EmperoAI
Announced a model claiming a 100-million-token context window, alongside published recall tests. Early access is coming; it isn’t publicly available yet.
ToyBlocks — @allen_ai
An open simulated tabletop for teaching and evaluating robots. Includes 85K demonstrations, two trained models, simulation assets and training/evaluation code.
Higgsfield Miracle — @higgsfield
Announced video post-production technology for separating footage into editable layers and making localized changes. Currently in private beta, with MCP access planned.
Alexandria SEC data — @firecrawl
Agent access to SEC information covering 8,000+ public companies. Research institutional holdings, insider transactions and company disclosures through tools like ChatGPT and Claude.
Conduit AI — shared by @asynchronous_x
Introduced a natural-language interface to its industrial automation platform, connecting work across factory data, systems and robots.
New model access and offers
GPT Image 2.6 in Runway — @runwayml
Now available in Runway, with GIF generation, SVG exports and precise edits that preserve character and product details. This is new platform availability.
LongCat-2.5 in Nous Portal — @NousResearch
One week of free access to Meituan’s multimodal model. The model launched earlier this month; the free access is today’s update.
Free FLUX 3 Max generations — @fal
Free generations through November 9, with text/image-to-video output up to 15 seconds at 720p. An access promotion, not a new model release.
Research, benchmarks and training
Cua Driver Bench — @trycua
A new benchmark with Scale AI testing whether desktop agents can finish tasks without interrupting the person using the same computer. That includes stealing keyboard focus or moving your pointer.
Frontier3DBench — @EinsiaAI
145 tasks across nine categories testing coding agents on 3D objects, scenes and interactive worlds in Blender and Unreal Engine 5. Evaluates appearance, structure and behavior.
Expanded speech evaluation — @huggingface
Egyptian and Nigerian English conversations join the Open ASR Leaderboard’s private evaluation data, expanding its accent coverage.
Harness adapter training — shared by @huggingface
Live reinforcement-learning runs training adapters for Liquid AI models and webAI’s Twil-9B to improve agent-task performance. Training is underway; these aren’t finished model releases.
AI-assisted cryptanalysis — @GoogleDeepMind
New research on human-guided AI agents uncovering weaknesses in legacy 1024-bit RSA systems. The team says these results do not affect modern RSA-2048 security.
5
2
21
879
This Echo Dot sat in a closet for at least 4 years.
Now she’s Cheryl, powered by Hermes, with her own voice, memory, etc.
Something cool/fun for the kiddos.
Feels good to “unfuck your old hardware,” as @NousResearch would say 🤣
Now she’s Cheryl, powered by Hermes, with her own voice, memory, etc.
Something cool/fun for the kiddos.
Feels good to “unfuck your old hardware,” as @NousResearch would say 🤣
My old Echo Dot run Hermes now, instead of Alexa.
I built a Home Assistant integration that connects it to a Hermes profile, with its memory and tools.
Code, setup skill and the hardware recovery notes are public.
github.com/majorminorlabs… @NousResearch
22
9
150
6,984
I don't know who needs to hear this but..
You have no idea how much of a positive impact you have on people who never say a word.
Keep doing what you’re doing. People are watching, learning, and rooting for you!
You have no idea how much of a positive impact you have on people who never say a word.
Keep doing what you’re doing. People are watching, learning, and rooting for you!
6
2
31
618
I genuinely want to see all of you win.
Believe it or not, helping others/providing value doesn’t always need a price tag attached to.
Happy Friday 🤝
Believe it or not, helping others/providing value doesn’t always need a price tag attached to.
Happy Friday 🤝
26
1
60
914
In case you missed it..
Hermes v0.23.0 dropped, The Hearth Release.
• Desktop in your browser: chats, Group Chats, plugins and settings, served from your own gateway. Enable Web Access in Settings.
• Shared conversations: the terminal, Desktop, browser and other interfaces connect to the same session history.
• Multiple machines connected: Desktop keeps your gateways connected at once, so you can switch between local and remote setups.
• Rewind a conversation: jump back to an earlier message in the terminal, edit it and try again, or branch into a new session.
• Clearer tool costs: supported paid tools show prices and running totals, with approval required before exceeding the chat’s spending limit.
Solid! x.com/NousResearch/s…
Hermes v0.23.0 dropped, The Hearth Release.
• Desktop in your browser: chats, Group Chats, plugins and settings, served from your own gateway. Enable Web Access in Settings.
• Shared conversations: the terminal, Desktop, browser and other interfaces connect to the same session history.
• Multiple machines connected: Desktop keeps your gateways connected at once, so you can switch between local and remote setups.
• Rewind a conversation: jump back to an earlier message in the terminal, edit it and try again, or branch into a new session.
• Clearer tool costs: supported paid tools show prices and running totals, with approval required before exceeding the chat’s spending limit.
Solid! x.com/NousResearch/s…
5
1
25
967