Sabitlenmiş Tweet
Transitioning to sharing technical insights, code fragments, and independent research updates in AI/ML. Focusing strictly on core engineering and substance over noise.
Feel free to connect to discuss systems or papers.
English
Chamuditha Gayashan
55 posts
Chamuditha Gayashan
@chamuuuz
Research in AI , Machine Learning , Robotics
Sri Lanka Katılım Nisan 2017
61 Takip Edilen7 Takipçiler
physics and compute density are naturally solving this. At current TDP levels for next gen accelerators, evaporative water cooling isn’t even viable anymore. the industry is fundamentally shifting toward closed loop liquid and phase change cooling to sustain performance, which completely changes the consumption narrative.
Sri Lanka 🇱🇰 English
Data centers aren’t stealing your water.
Even if the total water draw of data centers triples by 2030, they’d require just 8% of the water consumed by American golf courses.
@dodgeblake interviewed @AndyMasley, the man who’s been debunking AI water doomerism. Full story 👇
Naval@naval
The latest IQ test involves data centers and water.
English
Chamuditha Gayashan retweetledi
Welcome to Gemini 3.5 Flash, our most powerful model to date. It pushes the frontier of intelligence, speed, and cost putting 3.5 Flash in a class of its own.
We spent the last 6 months making sure Flash is great for real world use cases. It's available everywhere now!
English
@docmilanfar Yes, that's true to some extent. Maintaining relationships with people is an advantage in our professional journey.
English
graduates of the most elite colleges have superior career outcomes not primarily because of the education they receive, but because of the people they meet and the connections they make
English
developers…..What’s your go-to AI model right now?
- Gemini 3.1
- Opus 4.7
- Sonnet 4.6
- Codex
- GPT 5.5
English
❓What does an LLM actually do?
Simply -> it predicts the next token.
But that "prediction" comes from a probability distribution. Every token in the vocabulary gets a score: "this is how likely I am to come next."
Temperature 0 => always picks the top probability..
Temperature 1+ => samples from the distribution
That's why the same prompt gives you different responses.
✔ it's statistics...
English
@draft_ofkritika ->"4" * 2 = 44
->"3" + "44" = 344
Answer = 344
English
went back to stats, probability, and linear algebra fundamentals today.
every time I do this, ML clicks more:
PCA = eigenvectors of covariance
logistic regression = MLE
backprop = chain rule on matrices
L2 regularization = gaussian prior
the magic is in the connections, not the topics...
English
Software engineers, imagine AI becomes 10x better at coding in 3 years. what are you doing next?
English
@ravikiran_dev7 most devs use Windows because it's cheaper and it's what their company already runs not because the macbook isn't good.
English
counterfactual reasoning and true causal inference. current deep learning architectures excel at high dimensional statistical correlation, but mapping a structural causal model from sparse data or reasoning about 'what if' scenarios completely outside the training distribution is still fundamentally a human capability.
English
speed, because optimizing for scalability before validating the core product hypothesis is a textbook engineering anti pattern. building a highly distributed infrastructure for a system that might pivot in two weeks wastes critical engineering runway. Scalability is a high class problem you solve once you have verified retention.
English
I've noticed a similar cognitive shift in my own workflow. the execution constraint is no longer the syntax or the boilerplate generation time, but the mental context switching required to review and integrate the code. We are moving from a state of active writing to one of continuous code review and orchestration.
English
A 33% increase in top speed for actually smart summon suggests significant optimizations in inference latency or temporal consistency within the end to end model. In chaotic parking lot environments, even a modest speed adjustment requires much tighter control loops and faster object tracking to manage the diminished reaction window.
English
Tesla FSD V14.3.3 just started rolling out, and it comes with the Spring Update!
Actually Smart Summon’s top speed has also been increased to 8 mph (from 6 mph)!
Downloading it on my Model Y right now. Software version 2026.14.6.6.
English
@YashHustle_22 It really hinges on the scope of work. claude’s extended context window and reasoning capabilities are superior for mapping out system architecture and maintaining consistency across a large codebase
English
This valuation isn't just a reflection of hardware demand, it underscores the dominance of the CUDA ecosystem. While competitors are narrowing the gap in raw TFLOPS, the software maturity and deep integration with existing ML frameworks make NVIDIA the standard substrate for production grade computer vision and LLM pipelines.
English
🚀 NVIDIA Hit $5.5 TRILLION Market Cap!
From a $300B company in early 2023 to over $5.5T now. Mind-blowing growth.
English