60 years of war, and they've never actually seen the enemy... until now. 👽 This watchtower holds a darker secret than just the threat outside.
#SciFi #MovieRecs #PlotTwist #Thriller
1
2
5
2,451
CRAZY🗿🇶🇦 retweeted
With a gallery visit finished, I stopped under a canopy. After lunch, the Model S wore broken city reflections. It lifted the contrast. The route was ready. Quick power felt natural. The clean scene left the afternoon feeling more polished. #ElectricPerformance
1
1
1
10
CRAZY🗿🇶🇦 retweeted
A Robot’s Failures Are a Map
What if a robot failing at a task is not a problem to be eliminated, but a signal telling us where the model needs to learn next?
Consider a simple scenario.
A robot succeeds at the same task 95% of the time.
The obvious reaction is to collect more successful demonstrations.
But what about the remaining 5%?
Those failures may reveal something far more important:
the boundary of the model’s current capability.
Maybe the robot struggles when the object is slightly displaced.
Maybe it fails under a different visual configuration.
Maybe a small change in friction, orientation, or context is enough to break the policy.
These are not merely bad trajectories.
They are evidence of what the model has not yet learned to generalise.
And this changes how I think about failure data in Physical AI.
A successful trajectory tells us:
“The model can do this.”
A failure can tell us:
“The model cannot reliably do this yet.”
That second piece of information can be much more actionable.
This is where @axisrobotics ($AXIS) becomes interesting.
Instead of treating failure as the end of a trajectory, Axis is building toward a closed-loop system where model weaknesses can feed back into data generation:
Model → Evaluate → Identify Weakness → Collect Corrections → Train → Model
The important part is not simply collecting more failures.
It is turning failures into information about where the next data should come from.
In other words:
A robot’s failure is not just an error.
It is a map of the model’s capability frontier.
And if that map continuously determines what the system learns next, data collection becomes far more than a scaling problem.
It becomes an adaptive learning process.
That, to me, is where the real compounding effect starts.
Connect X with @KaitoAI / Yaps: s.kaito.ai/jd7nwhO
When Does Data Become Worthless?
Imagine a robot model that has just learned how to reliably pick up a cup.
Now you give it another 10,000 successful trajectories of the same task.
The dataset gets bigger.
But does the data become more valuable?
Not necessarily.
This is something I think is easy to overlook in Physical AI:
The value of data is not static.
A trajectory can be extremely useful when a model doesn't understand a capability.
But once the model masters that capability, thousands of nearly identical examples may have very little marginal value.
Meanwhile, a single trajectory showing something the model consistently struggles with could be much more valuable.
That means the same dataset can have very different value depending on what the model knows right now.
And this changes how we should think about data collection.
The question isn't simply:
“Is this trajectory good?”
It should be:
“Is this trajectory useful for the model we have today?”
This is where @axisrobotics ($AXIS) gets interesting.
Axis is pushing toward a system where data generation can adapt to the model's current weaknesses.
The loop becomes:
Collect → Train → Evaluate → Find Weakness → Generate Better Data → Train Again
So instead of treating data as a static asset, the system treats its value as something that changes with the model.
And that leads to a bigger idea:
"The future of Physical AI may not depend on collecting more data.
It may depend on continuously discovering which data is becoming valuable next"
Yesterday's valuable data may become tomorrow's redundant data.
And the model itself may be the best signal for knowing the difference.
Connect X with @KaitoAI / Yaps: s.kaito.ai/jd7nwhO
82
5
116
18,890
CRAZY🗿🇶🇦 retweeted
More Data Is No Longer the Answer
Imagine you're a robot model.
You already know how to pick up a cup.
Then someone gives you another 10,000 trajectories of… picking up cups.
You get more data.
But do you actually get much smarter?
This is the problem I think Physical AI is slowly running into.
For a long time, the assumption was simple:
More data → better model
And at the early stage, that makes sense.
But as a model improves, the value of additional data is no longer equal.
If the model already understands a task, another 10,000 similar successful trajectories may add very little.
Meanwhile, the data it doesn't have could be far more valuable:
→ unfamiliar object configurations
→ difficult environments
→ edge cases
→ failed attempts
→ capabilities where the model is still weak
So the question changes.
It is no longer:
“How much data can we collect?”
It becomes:
“What data should we collect next?”
This is where I find the direction of @axisrobotics ($AXIS) interesting.
Axis is moving toward a closed-loop system where the model doesn't just consume data.
The model's weaknesses can help determine what data should be generated next.
Collect → Train → Evaluate → Adapt Data → Train again
That changes the role of a data engine.
It's no longer just about scaling the amount of data.
It's about increasing the information value of every next trajectory.
My thesis:
"The real moat in Physical AI may not be who collects the most data
It may be who knows what data is worth collecting next"
And I think this is a much more interesting question than simply asking who has the biggest dataset.
Connect X with @KaitoAI /Yaps: s.kaito.ai/jd7nwhO
HUMANOID ROBOT: DATA PROBLEM THỰC SỰ LÀ GÌ?
@axisrobotics x @KaitoAI
Nhiều người nhìn humanoid và nghĩ:
“Chỉ cần model đủ thông minh → robot sẽ tự làm được mọi thứ.”
Nhưng với Physical AI, model không phải bottleneck duy nhất. Data mới là một trong những vấn đề lớn nhất.
1. Humanoid cần loại data nào?
Robot không chỉ cần biết “làm gì”, mà phải học được “làm như thế nào trong thế giới vật lý.”
Một trajectory tốt có thể chứa:
Observation → State → Action → Timing → Outcome
Ví dụ:
“Pick up a cup”
Với AI thông thường, đây chỉ là một instruction.
Nhưng với humanoid:
Tay phải di chuyển bao xa?
Lực nắm bao nhiêu?
Góc cổ tay thế nào?
Bước chân ra sao?
Nếu chiếc cốc trượt thì xử lý thế nào?
Nếu vật nằm lệch vị trí thì sao?
👉 Đây là physical interaction data, không phải text data.
2️. Vấn đề lớn: Data ngoài đời cực kỳ đắt:
Muốn thu thập real-world trajectory, cần:
Robot + Hardware + Operator + Environment + Time
Và mỗi robot chỉ có thể thực hiện một số lượng trajectory nhất định.
Trong khi đó, một model muốn trở nên general-purpose phải trải qua hàng triệu tình huống khác nhau.
Đây chính là nghịch lý:
Robot càng muốn general → càng cần data đa dạng → nhưng real-world data càng khó scale.
3️. Humanoid còn có một vấn đề khó hơn: Diversity:
Robot công nghiệp thường hoạt động trong môi trường được kiểm soát.
Humanoid thì ngược lại.
Nó phải đối mặt với:
🏠 nhiều môi trường
🧱 nhiều loại vật thể
👨👩👧 nhiều cách con người tương tác
💡 ánh sáng khác nhau
📷 camera/sensor khác nhau
🦿 embodiment khác nhau
⚠️ vô số failure modes
Vì vậy:
10 triệu trajectory giống nhau ≠ 10 triệu trajectory hữu ích.
Data cần diversity + quality + coverage, không chỉ volume.
4️. Và đây mới là phần khó nhất: Failure Data:
Robot học tốt không chỉ từ những lần thành công.
Nó cần biết:
“Tại sao lần này tôi thất bại?”
Ví dụ:
Robot → Attempt → Failure → Human Correction → New Data → Retrain
Một failure có thể tiết lộ:
policy sai ở đâu
trạng thái nào chưa được model hóa
environment nào chưa có trong training distribution
action nào khiến robot mất ổn định
Đây là lý do mình đánh giá cao hướng closed-loop data engine.
Thay vì:
Data → Model → Done
thì phải là:
Data → Model → Robot → Failure → Correction → Data → Better Model
🎯 Insight:
Theo mình, cuộc đua humanoid không chỉ là cuộc đua về hardware hay foundation model.
Một câu hỏi quan trọng hơn là:
Ai có thể xây dựng được data flywheel đủ lớn để robot liên tục học từ thế giới thực?
Hardware có thể copy.
Model architecture cũng có thể được cải tiến.
Nhưng một hệ thống thu thập → kiểm định → đa dạng hóa → học từ failure → đưa data quay lại training ở quy mô lớn thì khó hơn rất nhiều.
Đó mới là infrastructure layer mà Physical AI cần.
Humanoid không thiếu intelligence.
Humanoid đang thiếu enough high-quality physical experience.
Axis hub: hub.axisrobotics.ai/login?in…
Connect X với Kaito/Yaps: s.kaito.ai/jd7nwhO
#AxisRobotics #PhysicalAI #Humanoid #Robotics #AI #Kaito
78
3
105
4,152
CRAZY🗿🇶🇦 retweeted
RoboDojo - Deposit Coin (usability v2) - Guide to doing the task (3s)
Under the comment👇
33
1
58
1,326
CRAZY🗿🇶🇦 retweeted
🚨COUNTDOWN - 10 POSTS SUBMITTED
@axisrobotics x @KaitoStudio_ @KaitoAI
Only a few more hours until I find out if I made it into the top 500 ($AXIS)
But here’s the part I’m still wondering:
- Are these 10 posts strong enough to get me into Top 500? 👀
- That’s what makes the final stretch interesting.
- Submitting 10 posts doesn’t mean the job is done.
- Only your 6 best-performing posts will actually count.
-> So now it’s less about posting more, and more about letting the performance data speak.
How many have you submitted so far?
EP4 SETTLED - 2,992 POINTS 🔥
@axisrobotics
Check lại Portfolio hôm nay, mình có 2,992 Points cho EP4 (Aug 28 → Sep 11).
🎯 EP4: 2,992 Points
🤖 12,079 trajectories tổng cộng
⭐ 87.0 Avg Score
✅ 90.5% Verified
Nói thật, "EP4 lần này chưa khiến mình thấy quá nổi bật".
Không phải vì kết quả thấp, mà vì số lượng task trong Epoch này khá hạn chế, nên cơ hội để mình grind thêm và đẩy Points lên cao hơn cũng ít hơn những tháng trước.
Nhìn lại các Epoch:
2,697 → 2,697 → 5,399 → 2,992
EP3 từng đạt 5,399 Points, nên EP4 chắc chắn có một chút tiếc nuối.
Nhưng mình nghĩ đây cũng là một phần của cuộc chơi.
Không phải Epoch nào cũng có cùng lượng task. Điều mình kiểm soát được là sẵn sàng khi task xuất hiện và tiếp tục tích lũy data đều đặn.
EP4 khép lại.
EP5, mình tiếp tục grind.
Keep building. Keep contributing.
@plpiaoliang @KaitoAI
25
1
36
20,983
CRAZY🗿🇶🇦 retweeted
The Òṣè Ṣàngó. A carved wooden dance staff used in rituals to honor Ṣàngó, the Yoruba deity of thunder and lightning.
3
59
250
4,693
CRAZY🗿🇶🇦 retweeted
INSTEAD OF WATCHING NETFLIX TONIGHT.
Spend 1 hour with this.
Claude AI FULL COURSE that teaches you how to BUILD and AUTOMATE anything.
The people who watch this tonight will wake up tomorrow with a new skill.
Watch it and Bookmark it now.
9
116
546
38,086
CRAZY🗿🇶🇦 retweeted
50 AI tools that will save you hundreds of hours in 2026. 🤯
1. Claude — Solve any problem
2. Perplexity — Research anything
3. PortfolioTab — Create your portfolio
4. Kling AI — Create AI videos
5. Tripo AI — Create 3D models
6. Gemini — Perfect writing
7. CapCut — Edit videos
8. The AI Library — Discover useful AI tools
9. YouLearn — Summarize YouTube videos
10. Canva — Design graphics
11. ElevenLabs — Clone voices
12. Podcastle — Edit podcasts
13. ChatGPT — Brainstorm ideas
14. NotebookLM — Analyze documents
15. Lovable — Build web apps
16. Bolt new — Generate full-stack apps
17. Cursor — AI coding assistant
18. Windsurf — AI software development
19. Replit — Build apps in your browser
20. Gamma — Create presentations
21. HeyGen — Create AI avatar videos
22. Synthesia — Generate AI presenter videos
23. Midjourney — Generate AI art
24. Ideogram — Create images with text
25. Runway — AI video editing
26. Pika — Generate AI videos
27. Luma AI — Create cinematic AI videos
28. Leonardo AI — Create game assets & artwork
29. Figma AI — Design UI/UX
30. Framer AI — Build websites
31. Tally — Create smart forms
32. Otter ai — Transcribe meetings
33. Fireflies ai — AI meeting notes
34. Granola — AI meeting assistant
35. Wispr Flow — Voice-to-text
36. Zapier — Automate workflows
37. Make — Connect apps with automation
38. n8n — Open-source automation
39. Photoroom — Edit product photos
40. Remove bg — Remove image backgrounds
41. Suno — Generate AI music
42. Udio — Create original songs
43. DeepL — Translate accurately
44. Poe — Access multiple AI models
45. Grok — Real-time AI assistant
46. Genspark — AI super agent
47. Manus — Autonomous AI agent
48. Elicit — Research papers faster
49. SciSpace — Understand research papers
50. Notion AI — Write and organize notes
✅ Save this list—you'll probably use it more than you think.
52
646
1,959
221,392
CRAZY🗿🇶🇦 retweeted
This guy just explain how to Create Unlimited AI Videos for FREE | No Limits, No Watermark🤯
116
1,195
5,195
239,455
Look down X̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̳̅
Keep tweeting with ค้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้้ likes who cares
3
2
3
578