stealthcat

Marketed as “multimodal” but actually texts and images.

Multimodal dataset should be multimedia: text, audio, images, video, and optionally more like sensor readings and robot actions.

This worldly life is no more than play and amusement. But the Hereafter is indeed the real life, if only they knew. [1]

Know that this worldly life is no more than play, amusement, luxury, mutual boasting, and competition in wealth and children. This is like rain that causes plants to grow, to the delight of the planters. But later the plants dry up and you see them wither, then they are reduced to chaff. And in the Hereafter there will be either severe punishment or forgiveness and pleasure of Allah, whereas the life of this world is no more than the delusion of enjoyment. [2]

[1] https://quran.com/29:64

[2] https://quran.com/57:20

Unitree is better than Optimus.

Qwen et al is better than Grok.

BYD is better than Tesla.

Did you have some small toy experiments to prove this?

Should list most of the technical debt accumulated so far and rank them. At this stage, lots of corners have been cut.

Are there decent wind models and motor models if you want to model drones

Imagine all those millions of $$$ GPU of cloud credits for training, only to overlook this bug

Consider applying PhD program to some ML labs in Europe other than UK and France, they pay livable salaries like junior engineers. Better if in Swiss since the salaries are much higher.

European ML labs generally have less prestige than US/China ML labs but they pay much more humane wages even for people with family.

Eat lutein supplements and get as much sunlight and barefoot grounding

Kotlin Multiplatform.

In ML the problem is passing 0D scalar tensor as 1D 1-element tensor.

How do you trust numbers to “quantify extremely smart people”?

Interviewers know to hire/no-hire within first 10 minutes of talking to candidates.

Similarly, the best LLMs nowadays are probed only with some prompts.

Just last time I see LLM midwit meme on quantifying best scores on all benchmarks vs idiots/geniuses who just prompt something for some while.

Do not try to put numbers on everything.

More like GodotDesktop

Fun exercise… but for such use cases, people use Casadi and DrJit

https://github.com/casadi/casadi https://github.com/mitsuba-renderer/drjit

DrJit is made by same author of pybind11 and nanobind.

Isn’t this more of symptom of modern western values?

Try a book by Muhammad Asad (Leopold Weiss), like the Road to Mecca.

Every year during hajj in Mecca, there’s profound view of sea of humans wearing white from vast different backgrounds, wealth, health, ethnicities, continents. You can find Japanese, Africans, Australians, Russians, Indonesians, etc, variety of skin colors and builds, in one place.

Put the world in your hand, not in your heart.

Kotlin multiplatform

Quran

Good code? Code reading is more like understanding different cultures. Read big tech codebases meta and google, to startups codebases like openai and comma.ai, you will see.

Because you work in semiconductor industry.

Software engineering is hard problem. How is Julia in multiple authors codebase? A language need to consider ergonomics of multiple users working on same thing. Reading >> Writing.

C++/Python still hold ground in this regards sadly. One side for low-level control, another side for glue code.

Related: https://kidger.site/thoughts/jax-vs-julia/

Not OP, but in my experience, two methods actually work for me

1. Use SRS (spaced repetition system) like Anki to learn a difficult-to-read language. Slowly but consistently ~1 hour a day learn mandarin or japanese. Personally I managed to get JLPT N3 this way in ~2 years.

2. Specific to muslims: Memorize the Quran with the Itqan method [0]. Literally intense drilling by reciting 100+ times for a page of Quran a day by looking the page then 100+ times not looking. Can spend 2-4 hours easily for a page.

[0] https://hifdh.weebly.com/mauritanian.html

Personally Itqan method is way harder than SRS but retention is faster and stronger even without revision for months.

Depends on subreddit. Birds of feather flock together.

Join HVM and Kindelia’s Discord. https://github.com/HigherOrderCO/HVM

Read Interaction Nets paper

Last time I saw a project aiming to be like a fast decentralized REPL but with blockchain.

https://github.com/Kindelia/Kindelia-Chain

Seems like the so called Lisp hackers are Common Lisp users.

Before Matlab was cool, Lisp is VERY maths-heavy and engineering-heavy. Symbolic equation solvers, robot controls, neural nets, etc.

Even Julia which is very Matlab-like has roots to Lisp.

Find that “itch” that you really want to scratch.

Maybe building self-hosting language/compiler, or a toy linux-like kernel, or a physics engine, games then publish to itch.io, a small numpy-like library that can do autograd, whatever.

Once you start it really is hard to stop.

The spectrum is too wide, might need to dabble to see which works for you:

- Mainstream vision/text/speech, something Huggingface et al do. The classic advice where you learn SGD, stack net layers and put loss, train, deploy on web, clean and collect data, retrain

- Hardware accelerators, inference and training. Check out startups like tenstorrent, tinygrad, mythic. Also involve compilers, net graph data structure and good software frontends and abstraction to hardware to write models onto hardware.

- Anything else, applying ML for finance, manufacturing, oil and gas, all those boring but important stuff

Really, last time I read there’s zig compiler flag to not use LLVM at all (but not all features supported yet)

Religion is easy [0]. Those who commit apostasy are given time to reflect and repent to become muslim again, it is not like they are not given chance to turn back and live.

[0] https://www.abuaminaelias.com/dailyhadithonline/2011/04/30/i...

So far sadly yes? http://www.incompleteideas.net/IncIdeas/BitterLesson.html

tl;dr So far things that enable faster search and faster learning win over long run.

Recursive things like backprop in NN and optimizing reward over long trees of states, seem to win despite huge compute requirements.

Personally I think we are still on the right track of trying to do the right thing, then do the thing right, then do the thing faster.

You cannot refute the things you do not understand.