If you're not using the most powerful model for your therapist prompt, you don't love yourself.

founder Sanity = traction − comparison...boy, I'm deep in the negative right now.

man, the #claude ux is not friendly to mcp connectors serving images. my bit.ly/tensordiagram mcp server can add tensor visuals to help with deep learning explainers, but the end ux here is not seamless...

nobody asked for somebody to create a a visual explainer for torch.linspace. I did it anyways. #TensorViz

burned too many tokens to get CC to draw a tensor with thanksgiving colors using bit.ly/tensordiagram

I came across this meme over a year ago, and it still makes me laugh.

Got addicted to @srushnlp.bsky.social 's Tensor Puzzles, so I wrote a sequel with more puzzles: github.com/hardik-vala/.... Example:

using None as an index value for torch tensors adds a new dim. ludicrously simple when visualized.

finally made sense of all the ways you can index tensors in pytorch / numpy #TensorViz

"Quit the wrong stuff. Stick with the right stuff. Have the guts to do one or the other." — Seth Godin, The Dip A questionnaire I've developed up over years to navigate major stick or quit decisions: hardikvala.com/blog/stick-o... Answer with your gut, not your brain.:

Creating these visuals is addictive, but getting out of hand. Had to see the difference between torch chunk and unbind, but this visual doesn't need to exist. #TensorViz

Not my best work. Created a visual of torch's transpose op over 3d tensors, but I don't know if the random coloring is instructive or confusing. #TensorViz

TIL `tensor[None]` is just a shorthand for `tensor.unsqueeze( 0)` in torch.

seriously!? every time i open the app

Me: “who’s your favorite entrepreneur?” GF: “you” Me: “I’m not an entrepreneur. I’m a fraud.” GF: “that’s most entrepreneurs” Me: “good point”

Visual explainer for the pytorch op, `torch.unbind` #TensorViz

From Seth Godin's book, The Dip.

This is a cool resource: github.com/nayafia/lemo... . Documents a bunch of ways to make money with open source, with pros and cons of each approach.

Finished the visual explainer for the torch op, `chunk`. #TensorViz

Started a visual explainer for the torch op, `chunk`. The full explainer will have different arg combinations for no. of chunks and dim.

Framework lock-in is definitely not a day 0 / pre-PMF concern. mlechner.substack.com/p/why-we-sta...

whytorch.org is quietly addictive. It's been fun playing with pytorch tensors in a visual format. Not affiliated with the project. Just a fan.

torch.outer(a, b), visualized #TensorViz (drawn w/ bit.ly/tensordiagram)

Any time you see t[:, None] on a 1d tensor, it's just transpose #TensorViz

if you, like me, forget how tensors get reshaped. (drawn with bit.ly/tensordiagram)

Fed up with trying to imagine chains of tensor transformations in my head, wrote a library to relieve my pain: github.com/hardik-vala/.... Lets you visualize torch, jax, tf, numpy tensors in notebooks and other python contexts, for understanding and debugging. Enjoy.

is playground mode a new feature in colab? liking it for messing around with a notebook you authored that's also shared

JAX / TPU docs are incomplete. For anyone looking to run Pallas kernels (JAX's kernel lang) on TPUs, make sure you're running Python 3.10+.

Tensor Logic does not seem all that different from torch and tf, and don't see how it solves hallucinations. The unification of symbolic logic, database operations, and tensor transformations was intellectually instructive tho. arxiv.org/abs/2510.12269

When you create a typo in the process of fixing one. Entropy refuses to go down.

LFG

A remix on llama2.c: ModernBERT, in one file of pure C: github.com/hardik-vala/...

How do people visualize tensors? especially when reading foreign pytorch / jax code.

Takeaways from the Flower AI conference last week on decentralized / federated AI: • 𝗦𝗵𝗶𝗳𝘁 𝗶𝗻 𝗯𝗿𝗮𝗻𝗱𝗶𝗻𝗴 from "federated" and "privacy", to "decentralized" and "open".

Took a break from CC, and took Gemini CLI for a spin. Good DX still lacking: * No indicator in terminal telling you # code lines highlighted in IDE * Tends to make spurious formatting changes * Gets stuck in tool calling loops

One day, I would like my name to be immortalized as an LLM vocab token.

Find myself frequently saying, "Back when it was called ML...".

omg, the ONNX runtime is wicked fast. ran some perf benchmarks for ModernBERT on cpu vs gpu (rtx 4070) and ONNX was 3x faster than vanilla torch. not superior on cpu tho.

So any model trained/fine-tuned with llama outputs — even if it's not based on llama's architecture at all — must be prefixed with "llama" 😲 . e.g. huggingface.co/ai4privacy/l... (ModernBert fine-tune) License: www.llama.com/llama3_1/lic...

It's funny how the accessibility features in browsers make websites more accessible for browser agents.

i thought including the onloading/offloading of tensors to gpu mem would add significant lag when running batch llm inference, but it turns out those ops are negligible. makes sense when your gpu's mem bandwidth is 1450 gb/s. throughput: w/ mem ops: 155.7 tok/s w/o mem ops: 156.1 tok/s

I can't decide if #CompoundAI is fatuous term that refers to something obvious. It's like marketing cars as Compound Engine. Of course AI is integrated with other components to build complete applications.

OpenAI's hallucination paper is gaining attention, but in case you missed it FAIR published AbstentionBench a few mos ago: Benchmark for measuring LLM capacity to *abstain* against questions with underspecification, false premises, subjective interpretations, etc. arxiv.org/abs/2506.090...

PyTorch installs have unseated code compilations.

Thanks @mlopscommunity.bsky.social for the shoutout in today's newsletter! www.controlpla.in/blog/prompt-...

minusx.ai/blog/decodin... - Takeaways: * single-agent > multi-agent * small models > large models (for low entropy tasks) * llm search > RAG * PLEASE, NEVER, ALWAYS, IMPORTANT, etc. are still the best ways to steer LLM behavior

We tried prompt injection as a way to make our agents reliable and by god, it works! www.controlpla.in/blog/prompt-...

One lesson from a YC founder who shutdown his co after 5 years: "Go into markets you know deeply or problems you are truly passionate about solving."

Do like how GPT-5 isn't as trigger happy with calling tools as earlier versions. It does a good job confirming and gathering arguments before firing a tool call. #AIAgents

"What has been will be again, and what has been done will be done again; there is nothing new under the sun." Ecclesiastes 1:9