UchenML links: Browser demo - AlphaZero-style Dots: uchenml.tech/demos/dots Overview: uchenml.tech/introducing-uch… Using SIMD: uchenml.tech/simd-softmax-go… Making inference fast: uchenml.tech/uchenml-optimiz…

Come think of it - iOS is older than Windows. Modern Windows heritage is NT from 90s. iOS is NEXTStep from 80s.

This is quite healthy collaboration and actually saves a lot of time. Just review what your agent writes.

Was Minecraft originally an applet? So cool. Had it been made 10 years before than it would've been Shockwave Flash. 10 years after - and we would have a WASM.

At my age I really shouldn't care this much, for blood pressure reasons, but it makes zero sense to me that most iterator concepts require a default constructor. I have never once needed to conjure up a fresh iterator out of thin air.

One of my favorite anecdotes - one Lawrence Page building a web robot in Java 1.0beta2. Is that how Google began?

Late-Soviet kids were warned about vibe coding in 1979. The Adventures of Elektronik: boy finds an android that's his exact twin, outsources school to it. Android aces everything, becomes the popular one. Kid ends up bored, useless, and replaced by his own agent. We watched it as a comedy.

The wheel keeps turning. IntelliJ/Eclipse replaced the bloated IDEs and RAD tools of the late '90s. Sublime/Atom/VS Code replaced bloated Eclipse. Now it's VS Code's turn.

This really matters because the core Russian messaging is "Ukrainians and Russians are one people", "Ukrainian state never existed", "Ukrainian language is just a mangled Russian". These are the messages Russia uses as an excuse to bomb hospitals and churches. They are very harmful lies.

Linux won on POSIX compliance. Server side software used to live on “big iron” UNIXes (AIX, Solaris, HP-UX). All those developer skills and that software managed to migrate to Linux well. Meanwhile, ms was insisting on forcing Win32 api that was drastically different. That’s why Windows became a tiny niche.

Wrong way around. Most people on NVIDIA never touch anything below PyTorch. I wrote every algorithm in UchenML by hand on a Mac Studio - only based on textbooks and white papers. The hardware doesn't teach you. Typing with your fingers does. I can't imagine a better hardware for coding in C++.

I am surprised how many people are still confused with spaceship operator ("<=>") - and that's C++20.

An extremely interesting perspective from my coworker/friend @graphitemaster on LLMs/Claude: Having a lot of experience now using Claude both personally and professionally and seeing others use it in both those contexts, I'm convinced that it's output is very much a reflection of who uses it. The floor of who can make something with it has been lowered significantly, but the ceiling is still only reachable by those who are skilled in the domain and able to guide it. It can navigate much of the simpler aspects of the domain on its own that once required a skilled professional, but it struggles to navigate on its own for anything beyond that, requiring professional input. In some respects this means that our idea of what was once technically impressive is sort of misaligned now, but the idea of technically impressive and excellent work is still very much alive and real, it's just now that much more higher. In much the same way that high-school students can walk circles around Euclid and Archimedes today. The aspects that make computing and programming both interesting and a skill that takes time have not gone anywhere, they're just far more demanding now than they ever will be and there's no doubt in my mind that this will ever change. What I think this means for the @handmade_net community, and to a larger extent this industry as a whole, is that unless you're striving to do your own frontier work, or at least something close to that, anything you produce is going to have that err that Agentic tools could've produced it. To me this signals that the personal programming journey of people needs to shift towards a much higher bar of technical excellence. It's the same outcome of any field that matures. Decades ago you could cut it being a graphics programmer, these days you need to be incredibly skilled in the physics of light transport to deliver anything compelling, not to mention the amount of systems engineering experience you need as well to manage these vastly complex pieces of highly parallel hardware. This is no different.

No group of software engineers has bigger delusions of grandeur than game devs. One machine, one clock, one owner: that's how it was done in the 1980s. The industry evolved. Game devs didn't. Web software is torn into pieces running on their own time across the planet, failing independently, and it still has to look like one thing to the user.

I am excited about the future of software. Problems that used to feel insurmountable are now very approachable. Iteration is much cheaper too - resulting in better software.

Is Anthropic saying that their effort levels past med are a scam?

It's quite clear someone made a deal with Ahamkara.

Google Highway has that too. You write width-agnostic algorithm and then at compile time it gets correct lane count.

One day. I am telling you - one day we will get computers fast enough so Slack window can be resized without lag. But not today.

How can ARM CPUs compete with Intel/AMD when NEON is only 128-bit, while x86 can go all the way to 512-bit with AVX-512? Because 512-bit SIMD does NOT mean 4x the CPU performance of 128-bit SIMD. AArch64 NEON uses fixed 128-bit vectors. AVX2 can use 256-bit vectors, while AVX-512 can use 512-bit vectors. But vector width is only one part of the equation. Performance also depends on: -> How many vector operations can execute in parallel -> Execution throughput and latency -> Clock frequency -> OoO execution -> Cache and memory bandwidth -> How well the workload can actually be vectorized -> Power and thermal limits A CPU with multiple 128-bit execution pipelines can process several vector operations in parallel. Meanwhile, AVX-512 is incredibly powerful when the workload has enough parallel data to actually use those 512-bit operations. And ARM is no longer just 128-bit NEON either. SVE/SVE2 introduced scalable vectors, while SME/SME2 adds scalable streaming vectors and matrix processing. Apple's newer M4/A18-generation CPUs support SME/SME2, so describing modern Apple Silicon as simply "128-bit SIMD" isn't accurate anymore. The important point is: 128-bit vs 512-bit tells you how much data ONE vector instruction can describe. It does NOT tell you how fast the entire CPU is. AVX-512 can win massively in the right vector-heavy workload. But for general CPU performance, a wider SIMD instruction alone doesn't guarantee a faster CPU. That's how a CPU built around 128-bit NEON can still compete with, or sometimes outperform, a CPU that supports 512-bit AVX-512. SIMD width tells you how wide an instruction can be. Microarchitecture + workload tell you how much of that width actually turns into performance.