I gave WorkBuddy one prompt: build a 3D under water scene.
All generated in one shot.
The final video is below.
All generated in one shot.
The final video is below.
1
2
6
Build live apps in Cursor!
Share with anyone for FREE! No need to worry about where and how to host your first app. @cursor_ai
More details 👇
Share with anyone for FREE! No need to worry about where and how to host your first app. @cursor_ai
More details 👇
4
3
5
Step 2: Create the audio file
Replace book_content and voice_id with your own details, then run the code.
Replace book_content and voice_id with your own details, then run the code.
1
0
4
Listen to any book in your own voice 🎙️
Just a few lines of Python and AssemblyAI. That's it.
Step by step guide 👇
Quick demo:
Just a few lines of Python and AssemblyAI. That's it.
Step by step guide 👇
Quick demo:
1
2
15
HackerTube Amazing weekend hack project by @karpathy
- skip YouTube's annoying recommendations
- skip misleading thumbnails
- skip YouTube's annoying recommendations
- skip misleading thumbnails
Weekend hack: HackerTube. I was annoyed with YouTube's recommendations so I had Claude filter down 43M comments over 20 years of HN to 6,454 videos that keep being recommended. *Significantly* better viewing experience, ~2 hours with Opus 5.5:
hackertube.org
- Excellent (state is kept in local browser session)
- Thumbnail is a representative frame from middle of the video instead of the super duper A/B test hyper-optimized user uploads
- Underlying data from the comment analysis as one zip file, point your Claude at it and ask for your own version
- HackerTube Longs ✊
A bit more on the process:
I use Claude Code (desktop app for Mac), Opus 5.5 on Medium effort as my default. I only talk to Claude and have it do everything, the only "manual work" I helped with in tandem was in the browser (reminiscent of MenuGen): 1) I had to create some kind of API key from Google Cloud UI to query YouTube for video metadata, 2) I booted up a VPS box for serving and 3) I bought the domain name. Once the box was up I gave Claude the IP and told it update, secure, harden, deploy and optimize the code, and it did the whole thing.
Many awesome projects that would have normally taken weeks/months even can now be few-shotted for breakfast. What else are we not thinking of.
1
2
7
Antigravity now has a plugins marketplace.
Search and find every plugin in one place, no more hunting around for customizations. @googledevs @antigravity
Search and find every plugin in one place, no more hunting around for customizations. @googledevs @antigravity
1
2
8
Google just announced Gemini 4 Carbon, its strongest model yet.
#1 on the Artificial Analysis Intelligence Index (59). Built for cybersecurity and long-horizon work.
→ SEC-bench Pro: 88.2% for finding vulnerabilities
→ ExploitGym: 54.8% for exploitation
→ CWE-bench: 71.0% days on complex engineering tasks
→ 0.0% attack success on prompt injection
Same price as Argon: $2 input / $10 output per million tokens.
Access starts with a small group of cyber defenders through the Fairwind Program. Wider rollout coming soon ⚡️
0
1
8
Weights are open-source on Hugging Face and GitHub.
Serve locally with SGLang or vLLM, then point Claude Code at it:
Serve locally with SGLang or vLLM, then point Claude Code at it:
2
0
3
CLI agents need models that understand terminal workflows, not just text completion.
IQuest-Q1 is an open-weights model trained from scratch for CLI interaction. 320B total parameters, 15B active per token, sparse MoE.
Here's what it does. 🧵
IQuest-Q1 is an open-weights model trained from scratch for CLI interaction. 320B total parameters, 15B active per token, sparse MoE.
Here's what it does. 🧵
8
3
20
AI tools are everywhere, but direct access to Seedance on Dreamina for $0.29 definitely caught my attention. Register through my official partner link to get the offer:
dreamina.capcut.com/pricing/dreami…
#DreaminaOnly29Cents #Dreamina #DreaminaPartner #Seedance
dreamina.capcut.com/pricing/dreami…
#DreaminaOnly29Cents #Dreamina #DreaminaPartner #Seedance
4
1
7
Overnight agent runs rarely go wrong because of the coding. They go wrong because nothing checks the work until a human wakes up.
This is the loop that fixes that. The agent writes the code, calls the TestSprite CLI to run Integration Tests against the running app, reads the ompt that come back, patches the code, and runs the tests again. No human in the loop until the flows pass.
The animation is the cute version. The real version leaves a video replay of every run for you to check over breakfast.
This is the loop that fixes that. The agent writes the code, calls the TestSprite CLI to run Integration Tests against the running app, reads the ompt that come back, patches the code, and runs the tests again. No human in the loop until the flows pass.
The animation is the cute version. The real version leaves a video replay of every run for you to check over breakfast.
1
4
9
Hybrid Intelligence by Microsoft!
You can run your daily workloads on your local machine. No need of sending data out of your local network.
Local AI 🎉🎉
You can run your daily workloads on your local machine. No need of sending data out of your local network.
Local AI 🎉🎉
Meet Windows Hybrid Intelligence by @Microsoft
Now Copilot can connect to local models and ground over your entire PC, take actions.
Save time, money and keep your data safe!
1
3
8
Run EmbeddingGemma 2 locally!
Google's multimodal embedding model is now available on @ollama.
You can run any device including mobile phone with less than 1GB RAM.
Google's multimodal embedding model is now available on @ollama.
You can run any device including mobile phone with less than 1GB RAM.
Multimodal RAG on your phone. Fully offline. Free.
Google just dropped EmbeddingGemma 2.
It maps text, images, video, audio and code into one embedding space.
Measured on a Pixel 11 Pro:
→ Text only: ~191MB RAM
→ Full multimodal: ~567MB RAM
Search a voice memo to find a video clip. All on-device.
740M params. 8K context. Open weights under Apache 2.0.
Pairs well with Gemma 4 for private, on-device RAG 📱
x.com/googledevs/sta…
2
6
10
You just need 5 lines of Python code to download audio version of any YouTube video.
Let me know if you have any better way.
Let me know if you have any better way.
0
2
12
Nano Banana 2.1 is here!
Higher quality images at a new lower price point.
Available in the API, AI Studio, and Gemini app!
Higher quality images at a new lower price point.
Available in the API, AI Studio, and Gemini app!
2
2
5
Meet Kev: the Open source version of Jev.
Train or use pre-trained weights.
100% local and open source
Train or use pre-trained weights.
100% local and open source
Meet Kev - open source version of Jev
Everyone is talking about Jev right now
But it is a proprietary model and you dont' know what goes behind the scences
Kev built on Qwen3.5 based on Jev's architecture. You can use pretrained weights or train your own
Open source and 100% local.
1
2
10
Weights are open-source on Hugging Face and GitHub.
Serve locally via SGLang or vLLM and point Claude Code at it:
Serve locally via SGLang or vLLM and point Claude Code at it:
1
0
2
Benchmark scores:
→ Terminal-Bench 2.1: 83.2
→ CyberGym: 84.5
→ DeepSWE v1.1: 64.6
→ NL2Repo: 63.0
→ IQuest-CLIBench: 53.7
→ Humanity's Last Exam: 39.2 (toolless)
Competitive across key agentic and CLI evaluations.
→ Terminal-Bench 2.1: 83.2
→ CyberGym: 84.5
→ DeepSWE v1.1: 64.6
→ NL2Repo: 63.0
→ IQuest-CLIBench: 53.7
→ Humanity's Last Exam: 39.2 (toolless)
Competitive across key agentic and CLI evaluations.
1
0
2
Demo 3: Autonomous RL pipeline debugging in Claude Code.
An RL training run hit an unexpected reward drop. IQuest-Q1 stepped in.
→ Diagnosed: it read the training curves and on-disk traces, and found that an extra space inserted during text decoding broke prefix matching. rns were dropping out of the training loss.
→ Patched: it disabled the injected decoder spaces while keeping real generated whitespace.
→ Verified: multi-turn context alignment came back, and mean reward rose from 0.704 to 0.769, then 0.799 in later stages.
The model is debugging its own R&D loop.
An RL training run hit an unexpected reward drop. IQuest-Q1 stepped in.
→ Diagnosed: it read the training curves and on-disk traces, and found that an extra space inserted during text decoding broke prefix matching. rns were dropping out of the training loss.
→ Patched: it disabled the injected decoder spaces while keeping real generated whitespace.
→ Verified: multi-turn context alignment came back, and mean reward rose from 0.704 to 0.769, then 0.799 in later stages.
The model is debugging its own R&D loop.
1
0
2
Demo 2 goes further:
- The model integrates the three-body equations of motion numerically in under 100,000 steps, then renders the orbits in Three.js as glowing gold threads in the decorative style of Klimt’s *The Kiss*.
- It computes twin primes under 10 million and projects e 3D Archimedean spiral styled after Monet's *Water Lilies*.
- The model integrates the three-body equations of motion numerically in under 100,000 steps, then renders the orbits in Three.js as glowing gold threads in the decorative style of Klimt’s *The Kiss*.
- It computes twin primes under 10 million and projects e 3D Archimedean spiral styled after Monet's *Water Lilies*.
1
0
3
Demo 1: Single-prompt 3D web app.
One prompt produced Toy Soldier Showdown, a multiplayer 3D FPS set in a children's bedroom.
It includes:
- the 3D scene
- character controller
- shooting and reloading
- health HUD
- an in-game economy
Then the event logic:
- a wind-up full-screen alerts
- a mystery merchant popup
- a paper-airplane spectator view
- a building-block UI
Multi-file structure, WebGL rendering, and visual polish. All in one shot.
One prompt produced Toy Soldier Showdown, a multiplayer 3D FPS set in a children's bedroom.
It includes:
- the 3D scene
- character controller
- shooting and reloading
- health HUD
- an in-game economy
Then the event logic:
- a wind-up full-screen alerts
- a mystery merchant popup
- a paper-airplane spectator view
- a building-block UI
Multi-file structure, WebGL rendering, and visual polish. All in one shot.
1
0
2
CLI agents need models that understand terminal workflows, not just text completion.
IQuest-Q1 is a 320B parameter MoE trained specifically for CLI execution.
15B active parameters. Reads full codebases, runs commands, inspects outputs, and recovers from environment faults hat makes it different. 🧵
IQuest-Q1 is a 320B parameter MoE trained specifically for CLI execution.
15B active parameters. Reads full codebases, runs commands, inspects outputs, and recovers from environment faults hat makes it different. 🧵
1
3
8
Final result: Clean architecture diagram that accurately reflected the docs.
No prompt engineering. No tool-switching. Just described the outcome and let the agent handle the workflow.
No prompt engineering. No tool-switching. Just described the outcome and let the agent handle the workflow.
1
0
2
Here's what happened:
Opened Qoder's task screen. Selected Coding task. Pointed it to a workspace folder.
Prompt: "Read these docs and create an architecture diagram."
Picked Qwen 3.8 Flash (running at 0.0x Credits through September 30). Hit send.
Opened Qoder's task screen. Selected Coding task. Pointed it to a workspace folder.
Prompt: "Read these docs and create an architecture diagram."
Picked Qwen 3.8 Flash (running at 0.0x Credits through September 30). Hit send.
1
0
3
Tested Qoder this week.
Dropped a docs link into the task screen. Asked it to generate an architecture diagram.
The agent read the docs, planned the structure, and built the diagram without me touching a single tool.
Dropped a docs link into the task screen. Asked it to generate an architecture diagram.
The agent read the docs, planned the structure, and built the diagram without me touching a single tool.
131
75
86
Python Automation Cookbook:
Check out this book if you are looking for automating your AI workflows.
This book has 100+ recipes for scalable workflows, MCP integrations and AI powered automation.
Book link: acesse.one/rd7a7qm
Check out this book if you are looking for automating your AI workflows.
This book has 100+ recipes for scalable workflows, MCP integrations and AI powered automation.
Book link: acesse.one/rd7a7qm
3
5
20
Can you chat with Jev? @typesafeai
The simple answer is NO.
If you ask this or any question to any regular LLM, it tries go generate paragraphs, but Jev just responds with structured decisions in fraction of seconds.
The simple answer is NO.
If you ask this or any question to any regular LLM, it tries go generate paragraphs, but Jev just responds with structured decisions in fraction of seconds.
2
1
10
Just built Parked.
Snap where you parked. Tap to walk back.
No account. Works inside garages where GPS fails because the photo is the source of truth.
Built this app in 15mins with Rork + GPT-6 Astra. @rork
Snap where you parked. Tap to walk back.
No account. Works inside garages where GPS fails because the photo is the source of truth.
Built this app in 15mins with Rork + GPT-6 Astra. @rork
2
2
14
3/ Two-Line Migration
Works with OpenAI SDK, LangChain, Vercel AI SDK, or anything that accepts an OpenAI-compatible base URL.
To switch:
→ Change the base URL
→ Swap the API key
No refactoring. No redeployment. No downtime.
Works with OpenAI SDK, LangChain, Vercel AI SDK, or anything that accepts an OpenAI-compatible base URL.
To switch:
→ Change the base URL
→ Swap the API key
No refactoring. No redeployment. No downtime.
1
0
2
2/ Market Prices, Not List Prices
Official APIs charge fixed rates.
Dit.ai creates a real-time auction where 160+ providers compete for every request.
Your call routes to the best-priced qualified provider automatically.
Typical savings: 30-70% per call.
Official APIs charge fixed rates.
Dit.ai creates a real-time auction where 160+ providers compete for every request.
Your call routes to the best-priced qualified provider automatically.
Typical savings: 30-70% per call.
1
0
4
1/ One API Key for 50+ Models
Stop managing separate accounts for OpenAI, Anthropic, Google, and DeepSeek.
Dit.ai gives you GPT, Claude, Gemini, Grok, and dozens more through a single OpenAI-compatible endpoint.
Access everything without the vendor sprawl.
Stop managing separate accounts for OpenAI, Anthropic, Google, and DeepSeek.
Dit.ai gives you GPT, Claude, Gemini, Grok, and dozens more through a single OpenAI-compatible endpoint.
Access everything without the vendor sprawl.
1
1
5
I cut my AI API costs by 60% without changing a single line of application logic.
Same models. Same SDK. One routing layer that makes 160+ providers compete for every call.
Dit.ai is the infrastructure layer you need if you're building agents or shipping AI features. 🧵
Same models. Same SDK. One routing layer that makes 160+ providers compete for every call.
Dit.ai is the infrastructure layer you need if you're building agents or shipping AI features. 🧵
9
9
36
Run frontier MoE models on hardware you already own.
Colibrì is a pure C inference engine that runs massive Mixture-of-Experts models by treating VRAM, RAM, and NVMe as one memory hierarchy.
MoE models have hundreds of billions or trillions of parameters, but only a small ate per token.
Colibrì moves weights across storage tiers based on which experts are needed.
Hot experts stay in VRAM or RAM. Cold experts live on NVMe and stream in when the router selects them.
It tracks routing patterns, keeps frequent experts closer to compute, and prefetches the next layer before it's requested.
The result: limited VRAM changes speed, not which model you can run.
Key capabilities:
- Memory multitiering: VRAM, RAM, and NVMe as placement tiers
- Expert streaming: Load routed experts from storage instead of keeping all in memory
- Routing-aware caching: Per-layer LRU, hot-expert pinning, and one-layer-ahead prefetching
- Heterogeneous execution: CPU, CUDA, Metal, NUMA memory, partial or full expert residency
- Frontier-scale support: Qwen3.6 (35B) to Kimi K3 (2.8T), plus GLM-5.2, DeepSeek V4 Flash, and others
100% open source.
Colibrì is a pure C inference engine that runs massive Mixture-of-Experts models by treating VRAM, RAM, and NVMe as one memory hierarchy.
MoE models have hundreds of billions or trillions of parameters, but only a small ate per token.
Colibrì moves weights across storage tiers based on which experts are needed.
Hot experts stay in VRAM or RAM. Cold experts live on NVMe and stream in when the router selects them.
It tracks routing patterns, keeps frequent experts closer to compute, and prefetches the next layer before it's requested.
The result: limited VRAM changes speed, not which model you can run.
Key capabilities:
- Memory multitiering: VRAM, RAM, and NVMe as placement tiers
- Expert streaming: Load routed experts from storage instead of keeping all in memory
- Routing-aware caching: Per-layer LRU, hot-expert pinning, and one-layer-ahead prefetching
- Heterogeneous execution: CPU, CUDA, Metal, NUMA memory, partial or full expert residency
- Frontier-scale support: Qwen3.6 (35B) to Kimi K3 (2.8T), plus GLM-5.2, DeepSeek V4 Flash, and others
100% open source.
7
7
17
You can literally visualize your entire AI workflow using this amazing tool for FREE.
Your AI coding sessions generate tons of context.
But you can't see the full picture unless you ship everything to the cloud.
Jolli Local Dashboard changes that.
Visualizes your entire AI dev workflow. 100% local. Zero cloud.
Here's what you get 🧵👇
0
5
10
Stop guessing which models fit in your VRAM.
llmfit is a CLI tool that detects your hardware and ranks 206 models by what actually runs on your system.
No more downloading 50GB models that crash. 🧵
llmfit is a CLI tool that detects your hardware and ranks 206 models by what actually runs on your system.
No more downloading 50GB models that crash. 🧵
1
4
15
Three major releases in six months — Tencent Hunyuan is moving fast, and Hy4 preview has entered the open-source big leagues. 🚀
workbuddy.ai
@TencentHunyuan @TencentAI_News @WorkBuddy_AI
It brings 770B total parameters, 49B active parameters, and a 1M+ token or coding, productivity, research and real-world agentic tasks.
Tencent’s “model + product” co-design approach also brings it closer to products like WorkBuddy, with training shaped around real-world workflows.
In a blind test across 203 engineering tasks with 163 experts, Hy4 preview scored 2.99/4, ahead of GLM 5.3 and Kimi K3.
On public benchmarks, it clearly beats GLM 5.2 and sits right alongside GLM 5.3.
And the API pricing is the real surprise: $0.834/M input, $2.501/M output, and $0.042/M cache hits.
Top-tier performance, open-source access, and a seriously budget-friendly price.
Hy4 preview on WorkBuddy is currently free to try for two weeks for a limited time. 🚀
If you’re curious about what it can actually do, give it a try while the free access is available. 👀
workbuddy.ai
@TencentHunyuan @TencentAI_News @WorkBuddy_AI
It brings 770B total parameters, 49B active parameters, and a 1M+ token or coding, productivity, research and real-world agentic tasks.
Tencent’s “model + product” co-design approach also brings it closer to products like WorkBuddy, with training shaped around real-world workflows.
In a blind test across 203 engineering tasks with 163 experts, Hy4 preview scored 2.99/4, ahead of GLM 5.3 and Kimi K3.
On public benchmarks, it clearly beats GLM 5.2 and sits right alongside GLM 5.3.
And the API pricing is the real surprise: $0.834/M input, $2.501/M output, and $0.042/M cache hits.
Top-tier performance, open-source access, and a seriously budget-friendly price.
Hy4 preview on WorkBuddy is currently free to try for two weeks for a limited time. 🚀
If you’re curious about what it can actually do, give it a try while the free access is available. 👀
4
4
11
Long-horizon agency requires more than a larger context window.
An agent must evaluate its own progress, learn from unfamiliar environments, remember what matters, coordinate tools, and adapt when users or external conditions change.
dots3-note preview brings these capabilities ght multimodal model:
• Recursive self-critique with TEMPO
• Test-time learning and memory updates
• Text, vision, and speech understanding
• Complex reasoning and tool use
• Multi-stage task execution
• Adaptation across evolving real-life scenarios
Read the technical write-up:
studio.dots.ai/dots/dots3-en.…
Try dots3-note preview free on OpenRouter:
openrouter.ai/dots-studio/do…
Follow @dotsstudioai for more research, demos, and updates.
An agent must evaluate its own progress, learn from unfamiliar environments, remember what matters, coordinate tools, and adapt when users or external conditions change.
dots3-note preview brings these capabilities ght multimodal model:
• Recursive self-critique with TEMPO
• Test-time learning and memory updates
• Text, vision, and speech understanding
• Complex reasoning and tool use
• Multi-stage task execution
• Adaptation across evolving real-life scenarios
Read the technical write-up:
studio.dots.ai/dots/dots3-en.…
Try dots3-note preview free on OpenRouter:
openrouter.ai/dots-studio/do…
Follow @dotsstudioai for more research, demos, and updates.
1
0
1
6/ Travel planning doesn’t stop when the itinerary is created
User preferences can emerge gradually. Weather changes. Prices move. Inventory disappears. Flights and schedules may be updated without warning.
In this simulated End-to-End Travel Concierge scenario, dots3-note state across stages, coordinates services through tools, and adapts the itinerary as new constraints appear.
The challenge is not producing one polished travel plan.
It is staying useful throughout the journey—remembering earlier decisions, detecting when assumptions are no longer valid, and revising the plan without losing sight of the traveler’s broader goals.
This is what long-horizon agency looks like when the environment refuses to stand still.
User preferences can emerge gradually. Weather changes. Prices move. Inventory disappears. Flights and schedules may be updated without warning.
In this simulated End-to-End Travel Concierge scenario, dots3-note state across stages, coordinates services through tools, and adapts the itinerary as new constraints appear.
The challenge is not producing one polished travel plan.
It is staying useful throughout the journey—remembering earlier decisions, detecting when assumptions are no longer valid, and revising the plan without losing sight of the traveler’s broader goals.
This is what long-horizon agency looks like when the environment refuses to stand still.
1
0
1
5/ Now move from a controlled environment into business operations
Running an e-commerce business is not a one-step task.
E-commerce operations unfold across multiple stages, where user needs, service updates, and external conditions can change over time. Each new event may ere made earlier.
In this simulated Coffee E-commerce Operations scenario, dots3-note preview works across multiple stages, uses tools to handle different parts of the workflow, and maintains the context needed to keep its actions coordinated.
When new information changes the situation, it can revisit the existing plan instead of starting over.
That continuity is essential for agents designed to support real, evolving operations.
Running an e-commerce business is not a one-step task.
E-commerce operations unfold across multiple stages, where user needs, service updates, and external conditions can change over time. Each new event may ere made earlier.
In this simulated Coffee E-commerce Operations scenario, dots3-note preview works across multiple stages, uses tools to handle different parts of the workflow, and maintains the context needed to keep its actions coordinated.
When new information changes the situation, it can revisit the existing plan instead of starting over.
That continuity is essential for agents designed to support real, evolving operations.
1
0
1
4/ Perception is only the beginning
dots3-note preview jointly understands text, vision, and speech, then connects that perception with complex reasoning, coding, tool calling, and result delivery.
In this demo, the model must interpret information across different modalities, needed, coordinate their use, inspect the results, and continue working toward the broader objective.
The key is orchestration: individual capabilities are not used in isolation. They become part of one continuous workflow.
Perceive → reason → choose tools → act → evaluate → deliver
dots3-note preview jointly understands text, vision, and speech, then connects that perception with complex reasoning, coding, tool calling, and result delivery.
In this demo, the model must interpret information across different modalities, needed, coordinate their use, inspect the results, and continue working toward the broader objective.
The key is orchestration: individual capabilities are not used in isolation. They become part of one continuous workflow.
Perceive → reason → choose tools → act → evaluate → deliver
1
0
1
3/ Watch that learning loop unfold in ARC-AGI-3
The model enters an unfamiliar interactive environment with no complete explanation of its controls, objectives, or hidden mechanics.
Across thousands of turns, it must experiment, observe the results, build a working theory, and ver that theory fails.
A wrong prediction is not simply discarded. The model can identify the source of the mistake, preserve the correction, and apply the revised rule to later levels.
This is autonomous exploration at test time:
**Hypothesize → test → detect errors → update memory → try again**
The environment keeps changing—and the model keeps learning.
The model enters an unfamiliar interactive environment with no complete explanation of its controls, objectives, or hidden mechanics.
Across thousands of turns, it must experiment, observe the results, build a working theory, and ver that theory fails.
A wrong prediction is not simply discarded. The model can identify the source of the mistake, preserve the correction, and apply the revised rule to later levels.
This is autonomous exploration at test time:
**Hypothesize → test → detect errors → update memory → try again**
The environment keeps changing—and the model keeps learning.
1
0
1
2/ Test-Time Learning in Novel Environments
How does an agent improve when it enters an environment it has never seen before?
dots3-note preview explores, forms hypotheses, tests them against the environment, and updates its memory when those assumptions fail.
Explore → aluate → update memory
On ARC-AGI-3, this learning loop continues across long interactions, allowing the model to refine its understanding as it gathers more evidence.
The chart compares average ARC-AGI-3 score against inference cost, showing the performance–cost trade-offs across the evaluated models.
Through RL, dots3-note preview learns not only how to act, but also what is worth remembering and what should be forgotten.
How does an agent improve when it enters an environment it has never seen before?
dots3-note preview explores, forms hypotheses, tests them against the environment, and updates its memory when those assumptions fail.
Explore → aluate → update memory
On ARC-AGI-3, this learning loop continues across long interactions, allowing the model to refine its understanding as it gathers more evidence.
The chart compares average ARC-AGI-3 score against inference cost, showing the performance–cost trade-offs across the evaluated models.
Through RL, dots3-note preview learns not only how to act, but also what is worth remembering and what should be forgotten.
1
0
3
1/ Recursive Self-Critique via TEMPO
Long-horizon reinforcement learning has a fundamental problem: a rollout may continue for hours, while the final reward arrives too late to explain which of thousands of earlier decisions actually mattered.
TEMPO lets the same model and critic.
At each macro-step, the model pauses, examines its current state, reasons about the remaining task, and calls tools to estimate whether its trajectory is moving toward success.
Instead of waiting until the end to discover that a plan failed, it can generate earlier and richer signals about genuine progress.
The goal is not simply to act for longer—it is to recognize when the current direction needs to change.
Long-horizon reinforcement learning has a fundamental problem: a rollout may continue for hours, while the final reward arrives too late to explain which of thousands of earlier decisions actually mattered.
TEMPO lets the same model and critic.
At each macro-step, the model pauses, examines its current state, reasons about the remaining task, and calls tools to estimate whether its trajectory is moving toward success.
Instead of waiting until the end to discover that a plan failed, it can generate earlier and richer signals about genuine progress.
The goal is not simply to act for longer—it is to recognize when the current direction needs to change.
1
1
4
What happens when an AI task doesn’t end after one prompt—but continues for hours, days, or even weeks?
The agent must remember what happened, judge whether it is making progress, use tools, and revise its plans as the world changes.
Meet dots3-note preview, an open-weight ped by rednote’s dots studio @dotsstudioai for long-horizon agency in real life.
280B total parameters. 16B active. 512K context. Text, vision, and speech.
Here’s what it can do. 🧵
The agent must remember what happened, judge whether it is making progress, use tools, and revise its plans as the world changes.
Meet dots3-note preview, an open-weight ped by rednote’s dots studio @dotsstudioai for long-horizon agency in real life.
280B total parameters. 16B active. 512K context. Text, vision, and speech.
Here’s what it can do. 🧵
6
4
15
Promotional video creation is very smooth. Revid is able to get the core idea of the web app that I provided.
However, it could get the exact screenshot of the website, I was also not expecting that. I could have shared few screenshots. Otherwise nice video with voice over and text.
However, it could get the exact screenshot of the website, I was also not expecting that. I could have shared few screenshots. Otherwise nice video with voice over and text.
1
1
5
Claim The Spots just launched.
claimthespots.com
It's a pay-to-rank leaderboard. You submit your product or profile, bid to claim a position, and anyone can outbid you for that spot.
The mechanic is simple. Higher bids get higher visibility. The leaderboard is public. raffic to the top positions, which makes those positions worth more.
It's the same concept that made outbid blow up this week.
If you're launching something and want immediate eyeballs, this is another channel to test.
claimthespots.com
It's a pay-to-rank leaderboard. You submit your product or profile, bid to claim a position, and anyone can outbid you for that spot.
The mechanic is simple. Higher bids get higher visibility. The leaderboard is public. raffic to the top positions, which makes those positions worth more.
It's the same concept that made outbid blow up this week.
If you're launching something and want immediate eyeballs, this is another channel to test.
1
5
8
Here is my version of outbid!
claimthespots.com
No ads. No API keys. No revenue sharing. Claim your spot for $2.
Top spot is up for $3.
claimthespots.com
No ads. No API keys. No revenue sharing. Claim your spot for $2.
Top spot is up for $3.
Aaaaaaand we're live!
outbid.lol
No ads. No API keys. No revenue sharing.
Just outbid your competitors to rank #1 and consider marketing done for today 🚀
3
2
6
Elon is coming after Microsoft, and Cursor is actively targeting GitHub.
Here is what went down yesterday:
The Blackout: GitHub was down for 7.5 hours across Actions, Copilot, and core code workflows.
The Counter-Strike: Cursor dropped their new product, Origin, right in the middle of the outage.
Here is what went down yesterday:
The Blackout: GitHub was down for 7.5 hours across Actions, Copilot, and core code workflows.
The Counter-Strike: Cursor dropped their new product, Origin, right in the middle of the outage.
Just recorded a quick demo of Cursor Origin! 🚀
Syncing a GitHub repo literally takes a few seconds and then you can immediately start asking Cursor agents questions about the whole codebase. @cursor_ai
Check out how fast it works (demo is at 2x speed with few trims)👇
1
1
8
How to get progress bar in Python?
Here is the code snippet to include progress bar for your Python loops.
Let me know what you think about it in the comments.
Here is the code snippet to include progress bar for your Python loops.
Let me know what you think about it in the comments.
0
6
12
Intro to Deep Agents Course for FREE!
@LangChain academy is offering this foundation course for FREE.
You will learn how to build long running agents for complex workflows with Deep Agents.
Find the course link in the comments!
@LangChain academy is offering this foundation course for FREE.
You will learn how to build long running agents for complex workflows with Deep Agents.
Find the course link in the comments!
2
3
16
The weekend is here and if you want to dive into AI Agents, check out this book.
You can read this book for FREE.
Find the link 🧵👇
You can read this book for FREE.
Find the link 🧵👇
3
7
23
Dreamina Seedance 2.5 has officially launched on Dreamina, the dedicated AI creation platform by ByteDance.
If you are planning to build with Seedance, this is the exact platform I highly recommend using.
There are two massive upgrades that really stand out to me:
🎬 Up to 50 sets (images, videos, and audio) processed in a single generation.
⏱️ Up to 30-second continuous video generation, which unlocks longer and incredibly coherent storytelling.
When you fuse these capabilities with Dreamina's built-in editing workflow, multilingual support, and the expansive Seedance ecosystem, it genuinely feels like a major leap toward production-ready AI video.
This hub consolidates the entire Seedance lineup in one place—spanning from Seedance 2.0 Mini straight through to Seedance 2.5. It features creator-friendly pricing, making it one of the most cost-effective official routes to experiment, rapidly iterate, and produce content at scale.
Less time juggling different apps.
More time actually creating.
#Dreamina #Seedance25 #AIVideo #DreaminaPartner
If you are planning to build with Seedance, this is the exact platform I highly recommend using.
There are two massive upgrades that really stand out to me:
🎬 Up to 50 sets (images, videos, and audio) processed in a single generation.
⏱️ Up to 30-second continuous video generation, which unlocks longer and incredibly coherent storytelling.
When you fuse these capabilities with Dreamina's built-in editing workflow, multilingual support, and the expansive Seedance ecosystem, it genuinely feels like a major leap toward production-ready AI video.
This hub consolidates the entire Seedance lineup in one place—spanning from Seedance 2.0 Mini straight through to Seedance 2.5. It features creator-friendly pricing, making it one of the most cost-effective official routes to experiment, rapidly iterate, and produce content at scale.
Less time juggling different apps.
More time actually creating.
#Dreamina #Seedance25 #AIVideo #DreaminaPartner
0
2
4
🚨 Dreamina Seedance 2.5 is coming soon to the Dreamina platform.
ByteDance’s next AI video upgrade is bringing a more complete creation workflow.
What’s coming:
- Up to 50 multimodal references in one generation
- Up to 30 seconds of continuous video
- White-model and control
- Multilingual creation for global content
- Precise local editing for targeted revisions
This is a move from short AI clips toward longer, more controllable visual storytelling.
👉 aisecret.me/seedance25
#dreamina #dreaminapartner #seedance #dreaminaseedance25
ByteDance’s next AI video upgrade is bringing a more complete creation workflow.
What’s coming:
- Up to 50 multimodal references in one generation
- Up to 30 seconds of continuous video
- White-model and control
- Multilingual creation for global content
- Precise local editing for targeted revisions
This is a move from short AI clips toward longer, more controllable visual storytelling.
👉 aisecret.me/seedance25
#dreamina #dreaminapartner #seedance #dreaminaseedance25
4
4
10
Go annual on Pro or above → bonus AI credits on your usage:
→ $100 on Pro
→ $150 on Max
→ $200 on Ultra
→ $100 on Pro
→ $150 on Max
→ $200 on Ultra
1
0
2
Launch offer: Subscribe to one agent → get the other's Max plan free for 30 days.
Run OpenClaw and Hermes side by side on real work before you commit to either one.
Run OpenClaw and Hermes side by side on real work before you commit to either one.
1
0
2
MyClaw started with OpenClaw. Now Hermes is live. Claude Code and Codex are coming next.
Pick the agent runtime that fits your workflow. No Docker, no local setup, no server maintenance. Just managed agents that finish the job.
Pick the agent runtime that fits your workflow. No Docker, no local setup, no server maintenance. Just managed agents that finish the job.
1
1
5
ChatGPT gives you an answer.
MyClaw gives you the finished thing.
Give an agent a goal and the completed work lands in your WhatsApp, Telegram, or inbox. A sourced market report. Triaged emails. A competitor alert the moment something changes.
Not a chat window you forget to les.
MyClaw gives you the finished thing.
Give an agent a goal and the completed work lands in your WhatsApp, Telegram, or inbox. A sourced market report. Triaged emails. A competitor alert the moment something changes.
Not a chat window you forget to les.
1
4
9
This literally change the way we write docs for Agents.
OpenWiki is a CLI that writes and maintains agent documentation for your codebase.
Open source
OpenWiki is a CLI that writes and maintains agent documentation for your codebase.
Open source
4
5
9
Speech-to-text models keep guessing wrong on the words that actually matter.
Brand names become common words. Medical terms get butchered. Project codes vanish.
AssemblyAI Universal-3.5 Pro fixed this with contextual prompting.
Feed it your domain context and it stops t works 🧵👇
Brand names become common words. Medical terms get butchered. Project codes vanish.
AssemblyAI Universal-3.5 Pro fixed this with contextual prompting.
Feed it your domain context and it stops t works 🧵👇
3
4
12
Not gonna lie — I was skeptical when I saw this robot in my feed. Another humanoid? Another hype cycle?
But the more I looked into it, the more I realized I'd never seen anything quite like Q1.
It's at the AI for Good Summit in Geneva. It's 80cm tall. It's the first personal lt — and it actually looks like something you'd want in your home, not just in a museum.
I've been burned by too many overpromised tech products. But this one feels different. Like they actually thought about the person on the other end, not just the investors.
Anyone else feel like personal robots are finally becoming real? Or am I just getting my hopes up?
But the more I looked into it, the more I realized I'd never seen anything quite like Q1.
It's at the AI for Good Summit in Geneva. It's 80cm tall. It's the first personal lt — and it actually looks like something you'd want in your home, not just in a museum.
I've been burned by too many overpromised tech products. But this one feels different. Like they actually thought about the person on the other end, not just the investors.
Anyone else feel like personal robots are finally becoming real? Or am I just getting my hopes up?
0
2
6
I tested Tencent Hunyuan Hy3 with a small agent-building prompt on OpenRouter.
Prompt I used:
“Build a simple AI agent that turns a user request into a step-by-step action plan.”
I wanted to see if Hy3 could turn a simple instruction into something closer to real product clean.
Hy3 created an agent flow that started from the user request, understood the intent, broke the task into steps, and prepared a final action plan.
That’s useful because most AI agents need more than just answers. They need structure.
My takeaway:
Hy3 feels strong for agent-style workflows where the model needs to reason step by step and keep the output organized.
For builders working on AI assistants, automation tools, or workflow agents, this is worth testing.
Try Hy3 on OpenRouter:
openrouter.ai/tencent/hy3:fr…
@TencentHunyuan
#Hy3 #Hunyuan #TencentAI #AIAgents
Prompt I used:
“Build a simple AI agent that turns a user request into a step-by-step action plan.”
I wanted to see if Hy3 could turn a simple instruction into something closer to real product clean.
Hy3 created an agent flow that started from the user request, understood the intent, broke the task into steps, and prepared a final action plan.
That’s useful because most AI agents need more than just answers. They need structure.
My takeaway:
Hy3 feels strong for agent-style workflows where the model needs to reason step by step and keep the output organized.
For builders working on AI assistants, automation tools, or workflow agents, this is worth testing.
Try Hy3 on OpenRouter:
openrouter.ai/tencent/hy3:fr…
@TencentHunyuan
#Hy3 #Hunyuan #TencentAI #AIAgents
1
4
15
How to become an AI Engineer in 2026?
Andrew Ng just dropped this course on DeepLearning Platform.
- 9+ hours
- 31 video lessons
- 7 Code Examples
Get started for Free. Find the course link in 👇🧵
Andrew Ng just dropped this course on DeepLearning Platform.
- 9+ hours
- 31 video lessons
- 7 Code Examples
Get started for Free. Find the course link in 👇🧵
1
12
34
Python and ML Books for FREE!
- Intro to ML
- ML Projects
- Think Python
- Python for Data Analysis
Like, RT, and comment “Books,” and I’ll DM you the links.
- Intro to ML
- ML Projects
- Think Python
- Python for Data Analysis
Like, RT, and comment “Books,” and I’ll DM you the links.
23
33
74
9 Free Python courses from top universities 🚀
Want to elevate your Python skills? Explore these 9 highly rated courses, covering basics to advanced topics like OOP and data science!
A thread 🧵👇
Want to elevate your Python skills? Explore these 9 highly rated courses, covering basics to advanced topics like OOP and data science!
A thread 🧵👇
5
35
127
Every creative medium went through this:
Film was expensive → accessible.
Photography was expensive → accessible.
Music production was expensive → accessible.
Now AI video on Dreamina.
Film was expensive → accessible.
Photography was expensive → accessible.
Music production was expensive → accessible.
Now AI video on Dreamina.
1
0
2
Starting at just $0.02/sec for 720p. 55% off membership plans. Nearly 50% lower credit costs.
What this actually means:
You can try the weird idea.
You can render it and hate it and try again.
You can iterate until it matches the vision in your head.
What this actually means:
You can try the weird idea.
You can render it and hate it and try again.
You can iterate until it matches the vision in your head.
1
0
2
Officially launched by ByteDance's official product Dreamina: Dreamina Seedance 2.0 mini! Delivers on-par performance with Dreamina Seedance 2.0, but at a lower cost and faster speed!
It lives right inside the Dreamina platform — images and video in one workflow, no switching a surreal visual concept I'd been sitting on for weeks. Dreamina Seedance 2.0 mini handled it beautifully.
Same quality. Zero hesitation on my end this time.
🔗 aisecret.co/dreaminaseedan…
It lives right inside the Dreamina platform — images and video in one workflow, no switching a surreal visual concept I'd been sitting on for weeks. Dreamina Seedance 2.0 mini handled it beautifully.
Same quality. Zero hesitation on my end this time.
🔗 aisecret.co/dreaminaseedan…
1
0
4
AI video had its opening moment.
Not because the tools suddenly do more.
Because the fear of wasting credits stopped owning my creative decisions.
Here's why Dreamina Seedance 2.0 mini on Dreamina is the model I've been waiting for.
Not because the tools suddenly do more.
Because the fear of wasting credits stopped owning my creative decisions.
Here's why Dreamina Seedance 2.0 mini on Dreamina is the model I've been waiting for.
2
13
25
