「Open to Projects」 • Dabbler • Learner • Explorer • https://xitter.twitterapi.xyz/t.co/rMAvL5souA student • https://xitter.twitterapi.xyz/t.co/7FBysT47M6 solver • Dabbling in Embodied AI • 自学中文 // Teaching myself Chinese
Hong Kong/Shenzhen/GBA
Joined September 2022
- Tweets1.6K
- Following1K
- Followers257
- Likes390
Salman // 萨尔曼 retweeted
NVIDIA released Foundation4D. Give it multi-camera driving video and a single forward pass returns depth, camera poses, 3D boxes, semantic occupancy and Gaussians you can render. One shared backbone, a decoder per task.
Salman // 萨尔曼 retweeted
A recent UX pattern I've started using, I don't know if I got this from someone else, is I ask AI to allow me to highlight text in an HTML such that the highlight remains and it gets saved to a JSON file so that I can communicate with AI about the things I highlighted. I've found that it makes it a lot easier and quicker for me to get through a lot of text or data during AI-review. I haven't looked at actual numbers, but I would assume it also saves tokens.
Salman // 萨尔曼 retweeted
really is wild that AI is poised to reach PhD level first, then Starbucks barista level after that.
Salman // 萨尔曼 retweeted
AI is a tool. You can use it to think for you. Or you can use it to learn and try things you wouldn't have attempted before.
I tell everyone at Figma: we hired you because you can think. We expect you to think. Don't outsource your thinking to AI.
Social media is not reality. Things are never as good or bad as they seem. Seeing things as they are and seeking truth are skills to develop.
Curiosity must be cultivated. Pulling the level of a slop/slot machine can be weirdly exhausting and depleting. It can be hard and even anxiety inducing to step away when things are moving so fast, but time away from the computer is a must right now.
@zoink how do you keep your independent thinking in this tsunami of tech-bro-ai gloom and doom?
Salman // 萨尔曼 retweeted
On writing with AI help:
1. Write it yourself.
2. Ask AI for feedback.
3. Select only valid feedback.
4. Ask AI to address selected feedback with minor changes.
This will retain your voice while protecting against AI-speak takeover.
Salman // 萨尔曼 retweeted
There's something missing from benchmarks. Kimi K3 often writes much better code than GPT-6 Astra. It doesn't write useless verbose tests, it doesn't give confusing explanations. To be a good software engineer, you have to be able to communicate well, and RL doesn't capture that.
Salman // 萨尔曼 retweeted
This isn’t cloth.
It’s titanium powder, fused layer by layer by a Chinese metal 3D-printing company, into a mesh that wrinkles in your hand.
“Stronger than steel” is specific strength, not tank armor. The real gap is geometry: LPBF can grow interlocking cells a mill can’t cut.
That same print capacity is already shipping production parts to 30+ Chinese commercial space customers.
Materials aren’t catalog SKUs. Form is a process choice.
#AdditiveManufacturing #Titanium #LPBF #CommercialSpace #PhysicalAI #ChinaHardware
Salman // 萨尔曼 retweeted
Always before, in mathematics, solving a problem was nearly entirely aligned with finding useful reusable teachable ideas that pushed the field forwards.
Now that's not always true any more.
I think that's a really important and interesting insight! :)
Salman // 萨尔曼 retweeted
I love Astra BUT
I genuinely get ick whenever people use these models to tokenmaxx on some task they could solve with like 5k tokens using existing CV stack like @huggingface transformers or @roboflow under 200 LoC and low latency (real time)
I used GPT-6 Astra Ultra to track a tennis ball. The results are impressive until I tell you the following.
Ball annotation, using 3 parallel agents took 8m 35s and 17,611 tokens.
Complete workflow: 11m 49s and 30,481 tokens.
The complete workflow also processed 313,914 uncached input tokens and 7,528,320 cached input tokens.
That's a total of 7.87 million total, mostly context reused across frame-inspection calls.
Salman // 萨尔曼 retweeted
I’m seeing lots of tweet replies asking basically “but how do I get two models to talk to each other? Without me manually copy/pasting results between apps?”
And I’ve heard this same question repeatedly from friends who use AIs mainly through GUI apps.
Yes, there are various apps starting to be released for just this problem. But it’s worth knowing that in a pinch you can already do all this yourself, on the command line, just by using tmux.
Here’s how.
On the command line start a named tmux session with the command “tmux new -s chat”.
Create a second window within that session. (The default keystroke for this is doing “ctrl-b c”. You can switch between windows by doing “ctrl-b n” for the next window.)
Launch the codex CLI in one window, and the claude CLI in the other.
Then you can literally say to the AI in window 1, “please read the analysis of the AI in window 2 of the tmux session ‘chat’. Wdyt?” And vice versa. Now you have cross-harness review.
Or you can instruct one AI to delegate and manage the other AI over multiple turns, or monitor it over time, or whatever. This works because each AI can use the tmux CLI to read and write to the other’s window.
This does not on its own solve coordination problems like deciding who is in charge, or who talks first, or who waits for whom, but it establishes the basic communication primitives of “read the other AI’s output” and “write to the other AI”. You can build more complex workflows on top of that mostly by prompting.
The fundamental reason this all works is that the Unix-era command line environment is more interoperable than the modern GUI environment, so a decades-old tool like tmux is still one of the best ways to compose modern AIs. Obviously the command line tools are a PITA in other ways. But it’s great that they are so flexible, they already exist, and they work right now, so you don’t need to wait for the vendors to get their apps to play nicely with each other (which might never happen) or for someone else to catch up and write a new app just to connect things.
Salman // 萨尔曼 retweeted
My LLM cliché highlighter is up to 38 patterns now tools.simonwillison.net/llm-…
![Screenshot of a writing-analysis tool showing paragraphs of text with phrases highlighted in two shades of yellow, brown circular "3" badges, and a black tooltip overlaying the first line reading: "No X, no Y" chains · 3 "no" items. The visible text reads: We r[obscured by tooltip]nd up. No sign-ups, no downloads, no hassle (3) — just paste your text and start writing. Everything runs locally in your browser. The reviewer read the draft twice. Did not flinch, did not blink, did not reach for the red pen (3). That's the whole review, honestly. Don't call it a rewrite — call it a rescue. The improvement is real, and it's not subtle. That loss is worth naming. Sit with that for a moment. The gains were modest, but that's not nothing. You already know the answer, of course. Consistency is the entire game, and the punchline is that nobody wants to hear it. The entire pitch is one sentence long.](https://pbs.twimg.com/media/HQzQkUpXoAAVmjH.jpg?name=orig)
ALT Screenshot of a writing-analysis tool showing paragraphs of text with phrases highlighted in two shades of yellow, brown circular "3" badges, and a black tooltip overlaying the first line reading: "No X, no Y" chains · 3 "no" items. The visible text reads: We r[obscured by tooltip]nd up. No sign-ups, no downloads, no hassle (3) — just paste your text and start writing. Everything runs locally in your browser. The reviewer read the draft twice. Did not flinch, did not blink, did not reach for the red pen (3). That's the whole review, honestly. Don't call it a rewrite — call it a rescue. The improvement is real, and it's not subtle. That loss is worth naming. Sit with that for a moment. The gains were modest, but that's not nothing. You already know the answer, of course. Consistency is the entire game, and the punchline is that nobody wants to hear it. The entire pitch is one sentence long.
Salman // 萨尔曼 retweeted
We have a huge news to share today!
Today we are unveiling the first truly accessible RL robot - welcome Microduck
A 25 cm tiny open-source biped with 15 actuators and packed with sensors (camera, speaker, LiDAR, NFC, bluetooth, wifi, etc) that you train yourself with reinforcement learning.
It's also playable out of the box with more than half a dozen fun and playful pre-trained policies to have it walk, sit, crouch, roller-skate, pick up objects with its articulated beak, and recover on its own.
And all for less than $400.
See all the details, play with the simulator and order it at: pollen-robotics.com/microduc…
(video with sound on 🔊)
Salman // 萨尔曼 retweeted
Remembering the night I got fixated on paper data storage and learned you can fit about 140 pages of (encoded) plaintext on a single sheet of paper. The page below includes the encoder legend and its source code repeated 26 times
It's illegible at a glance and breaks when photographed or shared on social media, but is totally reconstructable without guidance if printed, scanned, and handed to an LLM for decoding
Do with that what you will!
Salman // 萨尔曼 retweeted
At 1400x cheaper, I know I'm sticking to embeddings (dense, sparse, multi-vector), plus hybrid (incl. bm25) and rerankers.
Perhaps I'd even use listwise cross-encoders, they seem interesting.
can LLMs replace embedding models? in a new work "embedder's dilemma" we find LLMs now beat embedding models -- but at much higher cost. when to choose which?...
📜 arxiv.org/abs/2608.12875
Salman // 萨尔曼 retweeted
BREAKING: The Robot Lamp Wars are coming. Choose your lamp! They’re all cute, use AI and are animatronic.
🇻🇳 - @dee_hw Autonomous AI
🇺🇸 - @natalieyeo PlaiPin
🇯🇵 - @skylark_mmm Nightingale01
🇫🇷 - @Karim_RC Ongo
🇺🇸 - @SarkaryShahvir LeLamp
Which one is your favorite one?
Salman // 萨尔曼 retweeted
fall-detection-vison
Elderly Fall Detection System
A real-time video-based fall detection system that uses skeleton pose estimation and temporal deep learning to detect when a person falls and stays on the ground. Designed for elderly care facilities, hospitals, and smart home monitoring.
Salman // 萨尔曼 retweeted
A coffee shop is using AI to track every barista and customer in real time.
- How many cups each person made
- Exactly how long every customer has been sitting
- Who is fast and who is falling behind
- Movement patterns across the floor
This used to be enterprise-only tech.
It’s arriving at small businesses right now.
Salman // 萨尔曼 retweeted
Watch a human do a task. Hell, actually think about how YOU'RE doing a task like putting away groceries. You will notice that the human uses vision barely at all.
