Visruth Srimath Kandali
@visruth.com
Cal Poly Statistics student interested in statistical computing/learning and its applications. Especially interested in R, Julia, Stochastics, and Bayesian methods.
visruth.com
visruth.com
Current training schemes don't value uncertainty and I think it's a killer capability. There was that recentish OpenAI paper that talked about this:
www.visruth.com/blog/2025090...
www.visruth.com/blog/2025090...
Ternary Truths and Hallucinations
Visruth Srimath Kandali's Website
www.visruth.com
December 5, 2025 at 8:22 PM
Current training schemes don't value uncertainty and I think it's a killer capability. There was that recentish OpenAI paper that talked about this:
www.visruth.com/blog/2025090...
www.visruth.com/blog/2025090...
Never underestimate uni students when food is on the line
December 5, 2025 at 8:19 PM
Never underestimate uni students when food is on the line
Just read your change and I like it too and don't understand why core rejected it and said that, strange decision but hopefully people will find your thread. Maybe worth making a quick blog post?
December 3, 2025 at 9:33 AM
Just read your change and I like it too and don't understand why core rejected it and said that, strange decision but hopefully people will find your thread. Maybe worth making a quick blog post?
I've always thought NaN means something was supposed to be a number but a function threw an error somewhere and now you have something which is Not a Number but is expected to be; and that NA covered other scenarios and was never presupposed to be a number.
December 3, 2025 at 9:31 AM
I've always thought NaN means something was supposed to be a number but a function threw an error somewhere and now you have something which is Not a Number but is expected to be; and that NA covered other scenarios and was never presupposed to be a number.
December 3, 2025 at 9:12 AM
Whats your walk in song?
December 3, 2025 at 9:11 AM
Whats your walk in song?
That's right! Seems like companies are maximizing utilization of MCP internally before synergizing its benefits to the public.
Is there clout like karma on Twitter or why do people farm? Other bot followers? Ragebait?
Is there clout like karma on Twitter or why do people farm? Other bot followers? Ragebait?
December 3, 2025 at 9:10 AM
That's right! Seems like companies are maximizing utilization of MCP internally before synergizing its benefits to the public.
Is there clout like karma on Twitter or why do people farm? Other bot followers? Ragebait?
Is there clout like karma on Twitter or why do people farm? Other bot followers? Ragebait?
You can inline dependencies in scripts iirc, so I'm sure there's some way to do that with marimo notebooks since they're just Python.
November 30, 2025 at 11:12 PM
You can inline dependencies in scripts iirc, so I'm sure there's some way to do that with marimo notebooks since they're just Python.
I don't think you'd be using notebooks, but if you do you should try marimo. uv and marimo are my favorite Python tools in a while and make working in Python more enjoyable.
November 30, 2025 at 11:02 PM
I don't think you'd be using notebooks, but if you do you should try marimo. uv and marimo are my favorite Python tools in a while and make working in Python more enjoyable.
Oop if I read the quotes post it answers in the positive if I'm reading it right
November 27, 2025 at 6:27 PM
Oop if I read the quotes post it answers in the positive if I'm reading it right
Haven't skimmed it yet; was the RL distributed like their last model? Very interesting even if it wasn't of course, but distributed RL is super cool.
www.primeintellect.ai/blog/intelle...
www.primeintellect.ai/blog/intelle...
INTELLECT-2: The First Globally Distributed Reinforcement Learning Training of a 32B Parameter Model
Today we are launching INTELLECT-2: the first 32B parameter globally decentralized Reinforcement Learning training run where anyone can permissionlessly contribute their heterogeneous compute resource...
www.primeintellect.ai
November 27, 2025 at 6:27 PM
Haven't skimmed it yet; was the RL distributed like their last model? Very interesting even if it wasn't of course, but distributed RL is super cool.
www.primeintellect.ai/blog/intelle...
www.primeintellect.ai/blog/intelle...
Haha true. ChatGPT R2 coming to you, soon #deepseeksniped
November 26, 2025 at 8:29 PM
Haha true. ChatGPT R2 coming to you, soon #deepseeksniped
I thought there were no more reasoner only releases and they were folded into the mainline model? I.e they wouldn't separately release an o4/o5?
November 26, 2025 at 8:00 PM
I thought there were no more reasoner only releases and they were folded into the mainline model? I.e they wouldn't separately release an o4/o5?
I don't have it on me now, but my LA prof handed out printed sheets with this. It's Micheal Jordan normal form as far as I'm concerned.
November 26, 2025 at 7:58 PM
I don't have it on me now, but my LA prof handed out printed sheets with this. It's Micheal Jordan normal form as far as I'm concerned.
Well, at the very least it's cheaper now but I wonder how it feels--uni only pays for ChatGPT :(
Tops some benchmarks, ofc, if you care about that kinda thing still
www.anthropic.com/news/claude-...
Tops some benchmarks, ofc, if you care about that kinda thing still
www.anthropic.com/news/claude-...
Introducing Claude Opus 4.5
Anthropic is an AI safety and research company that's working to build reliable, interpretable, and steerable AI systems.
www.anthropic.com
November 24, 2025 at 7:25 PM
Well, at the very least it's cheaper now but I wonder how it feels--uni only pays for ChatGPT :(
Tops some benchmarks, ofc, if you care about that kinda thing still
www.anthropic.com/news/claude-...
Tops some benchmarks, ofc, if you care about that kinda thing still
www.anthropic.com/news/claude-...
Footnote 1: "Similar to Cogito v2, we fork off the open-licensed Deepseek base model from November 2024 and run post-training in-house for Cogito v2.1."
Wasn't V3 released in December? Is there another model in forgetting or is this a typo?
Wasn't V3 released in December? Is there another model in forgetting or is this a typo?
November 19, 2025 at 10:58 PM
Footnote 1: "Similar to Cogito v2, we fork off the open-licensed Deepseek base model from November 2024 and run post-training in-house for Cogito v2.1."
Wasn't V3 released in December? Is there another model in forgetting or is this a typo?
Wasn't V3 released in December? Is there another model in forgetting or is this a typo?
Tbf the ease of use of S3 is unparalleled; I still remember being so shocked when @kellybodwin.com wrote something like `print.obj <- \(x) print("Hello")` and showed me that was all it took to define a print method for `obj` classed objects. I was only used to Java OOP so S3 seemed like mayhem.
November 19, 2025 at 10:39 PM
Tbf the ease of use of S3 is unparalleled; I still remember being so shocked when @kellybodwin.com wrote something like `print.obj <- \(x) print("Hello")` and showed me that was all it took to define a print method for `obj` classed objects. I was only used to Java OOP so S3 seemed like mayhem.
Absolutely. I was very pleased to see ggplot 4 using S7 and I hope that encourages more devs to use S7 for new projects. The interop is good too (S7 is built on S3 iirc). I see no reason to bite the bullet and use R6 or S7 for new projects--stricter, but saner.
"If it hurts, do it more often"
"If it hurts, do it more often"
bliki: Frequency Reduces Difficulty
"If it hurts - do it more often". Good advice if the amount of pain raises exponentially with the time between actions, such as for integrating software.
martinfowler.com
November 19, 2025 at 10:35 PM
Absolutely. I was very pleased to see ggplot 4 using S7 and I hope that encourages more devs to use S7 for new projects. The interop is good too (S7 is built on S3 iirc). I see no reason to bite the bullet and use R6 or S7 for new projects--stricter, but saner.
"If it hurts, do it more often"
"If it hurts, do it more often"
Yeah I'm with Wilke--NSE is just so nice.
I think S7 is the correct next OOP system in R and is sufficiently strict; I think the natural next (in a long while, probably) step would be type hints and their ilk, but I do agree that as-is, R OOP is not fun and Python is much better in that regard.
I think S7 is the correct next OOP system in R and is sufficiently strict; I think the natural next (in a long while, probably) step would be type hints and their ilk, but I do agree that as-is, R OOP is not fun and Python is much better in that regard.
November 19, 2025 at 10:22 PM
Yeah I'm with Wilke--NSE is just so nice.
I think S7 is the correct next OOP system in R and is sufficiently strict; I think the natural next (in a long while, probably) step would be type hints and their ilk, but I do agree that as-is, R OOP is not fun and Python is much better in that regard.
I think S7 is the correct next OOP system in R and is sufficiently strict; I think the natural next (in a long while, probably) step would be type hints and their ilk, but I do agree that as-is, R OOP is not fun and Python is much better in that regard.
Gemini 4, obviously
November 18, 2025 at 9:38 PM
Gemini 4, obviously
IDK nothing seems outta pocket or sus to me
November 8, 2025 at 12:32 AM
IDK nothing seems outta pocket or sus to me