Kris Pamungkas retweeted
Replying to @lexfridman
This is a US national interest and it is influenced heavily by political goals. For example, top US AI companies don't like NVIDIA path to support open weight models.
Kris Pamungkas retweeted
As a computer scientist in AI (and also a tech founder), I see it the other way around.
People who have no idea how AI works trying to draw relations and conclusions about its abilities telling me I don’t know what I’m talking about.
My stance shouldn’t be controversial, AI is a computer program, a mathematical model doing large matrix statistics, and it is impossible for it to give rise to feelings, desires, emotions, awareness, understanding, or consciousness, and comparing it to biological systems is foolish.
And before you start telling me about the people who are saying it who are experts… they have an agenda.
Even the “Godfather of AI” (more like “God Complex”) is selling books and fear sells, as does popularity, interviews, and making batshit crazy claims.
Most of the tooling that people use to build and inference models is open source. You can read it yourself (if you know how) or you can have agent do it and explain it to you. I can assure you, its design is to emulate and generalize on its training data, not be alive.
So I’m curious how many people who know very little about how AI works will argue with me saying how little I know about consciousness, biological processes, or that my stance is somehow influenced by me being a tech bro / CEO in AI…
Kris Pamungkas retweeted
Billionaire$ claim AI will bring in 'wealth for all' while simultaneously AI is just being a tool/excuse to lay off people, the $ is loosing value (not gaining value), costs are increasing putting small businesses out of business, and wealth/control moves to the few ...
Kris Pamungkas retweeted
The only 2 explanations for what's happening in AI.
1. They're afraid of open weight models
2. They've hit a wall. They all the see the same thing.
There's no way in hell they are seeing ASI. They still ship Electron apps. If they were close to ASI then they could cure cancer
Kris Pamungkas retweeted
FreeBSD on laptops? It’s happening! 💻 FreeBSD is making strides in laptop adoption, bringing stability and flexibility to mobile users. Discover how this shift expands FreeBSD's reach in the open source world. #FreeBSD #OpenSource #Laptop #TechNews
hubs.li/Q02SFrql0
Kris Pamungkas retweeted
The most fun in programming was the autocomplete era. Much better than the current slot machine + code review. It's a pity they stopped making better and better autocomplete models.
Kris Pamungkas retweeted
Since the @bunjavascript folks recently wrote about their experience moving from Zig to Rust, it seemed like a good time to write about our experience rewriting @roc_lang from Rust to Zig!
rtfeldman.com/rust-to-zig
Kris Pamungkas retweeted
Trump becomes the first president in history to openly use its position to get a FIFA World Cup rigged.
He is also the president who made billions in cryptos on the back of his fans.
He also manipulate the markets during the war.
I know these values... i don't like them.
UEFA statement on the Balogun case: ⬇️
uefa.com/news-media/news/02a…
Kris Pamungkas retweeted
Most privacy advice tells you to add things. A VPN, a blocker, another extension, another subscription.
Here's one that works the other way: delete the surveillance machine entirely.
It's 29 years old, free, and many people think it died in 2013.
It's RSS.
1/9
Kris Pamungkas retweeted
For those wondering why there has historically been a lack of solid graphical Linux debuggers, this is why. Everything is like this. This is a glimpse into the reality of *just reading the bare minimum amount of information stored in the dominant debug info format* on Linux - never mind process control, toolchain variability, lack of standardization of *thread-local-storage* or *module loading*, for Christ's sake.
Around every corner is friction and variability. No vision that connects with the user. No design taste. No consideration for the complexity required to use data formats or APIs from scratch. Eggheaded overly-generic academic slop specifications. It's just terrible.
Luckily, we're (finally) close, but this is why it's not trivial.
Kris Pamungkas retweeted
TinyGPU v3.0 is coming with a programmable Pixel Shader!!
Tested on FPGA and submitted to the #TinyTapeout May shuttle (~290k transistors).
This Chameleon demo is a tribute to the #GeForce 3, the first programmable GPU released 25 years ago. github.com/pongsagon/ttsky26…
Kris Pamungkas retweeted
Perfect microcosm of Wikipedia: I tell Jimmy Wales that JangaFX was written in Odin. He asks for a source. A JangaFX founder replies to him and confirms that it was. Jimmy ignores his (and my) response, while replying to later posts in the thread:
Kris Pamungkas retweeted
Do you use Odin commercially? Whether running a business or building a commercial application or game?
We'd love to feature you on the official Odin Programming Language Website.
Reach out at [email protected] or directly on Twitter/X.
Kris Pamungkas retweeted
I've just updated the "Odin's Fiasco with Wikipedia" article to include my discussion with Jimmy Wales:
gingerbill.org/article/2026/…
Replying to @wookash_podcast
PTX, SPIR-V, and DXIL are not stacked on each other. They’re separate IRs. On NVIDIA, the driver lowers them to NVIDIA machine code for the target GPU. Vulkan shaders do not go through the CUDA path, and same story for DXIL on Windows. PTX is the CUDA-side IR. SPIR-V and DXIL are graphics/portable API-side IRs. So, when you look at it closely in a shader disassembly, you will notice that one DXIL may map to one or multiple SASS instructions (Nvidia's machine code for multiple architectures).
Closest gamedev analogy I can think off right now in general is something like: FBX vs glTF. Different source formats, same engine import/backend path, same final runtime representation. Here the driver is both the importer and the final compiler backend.
Why multiple APIs at all? Different contracts. CUDA was built as NVIDIA’s compute stack. Vulkan, DirectX, and Metal are broader graphics + compute APIs with portability and pipeline state baked in. ROCm is AMD’s closest CUDA analogue. OpenCL and SYCL tried to be more universal abstractions, with mixed real-world traction. why doesn't NVIDIA itself unify them, given they own the hardware? The real answer is somewhere along the line of: graphics APIs need render state, pipeline objects, fixed-function configuration and the compute just wants to dispatch kernels.
A lot of the same programmable cores run the math regardless of which API got you there. What changes is which fixed-function or specialized blocks get involved.
Our gaming Graphics can fire TMUs, raster, ROPs, RT cores. AI compute workloads heavily lean on tensor cores and the SM/l1/l2 cache hierarchy.
Also, so far I've seen no use of FP4 units as much as I've seen fp16units in graphics, so that is still way more common to AI workloads.
Graphics is starting to reach tensor-style hardware more directly, too, through things like Cooperative Matrix and related matrix APIs.
That’s also why H100 looks nothing like a 4090. H100 drops RT cores and display hardware and pours silicon into tensor throughput, HBM, NVLink, and cache/bandwidth for AI/HPC. 4090 makes the opposite trade: full graphics pipeline, 128 RT cores, display stack, and 72 MB L2 versus H100’s 50 MB.
And yes this is quite Opaque by design? The public IR stays stable, while the vendor keeps the real backend private so the same shader can be retargeted across architectures and driver generations.
Also, one cool thing is that Direct X will start deprecating DXIL in favor of SPIR-V, thankfully.
Kris Pamungkas retweeted
“Stop” might be the right word for concurrent collectors that aren’t on-the-fly, and even then it’s kind of sketchy.
Folks often end up with concurrent GCs where the time the mutator is “stopped” is less than the time that a C or Rust mutator would be “stopped” waiting for mmap to return the next available page for an allocation that is growing the heap.
For example in FUGC (which is on the fly), the closest thing to “stopping” is a thread scanning its own stack. But what if C++ exceptions have to scan the stack? We don’t say C++ “stops” for that, even though C++ exception unwinding is slower than FUGC stack scans
The discussion of “stopping” and “pausing” is detrimental to a good understanding of GC overheads and *why* GCs are often a bad idea. The reason why they’re a bad idea:
- GCs will use more memory than a program that uses malloc/free.
- GCs cause memory leaks. Relatedly, GCs make it impractical to estimate max memory footprint in systems where this needs to be known statically
- GCs increase page demand (the rate at which a program asks for a page that isn’t mapped).
- GCs cause a tremendous amount of memory traffic. (A concurrent on the fly GC will have an effect on latency that looks like a pause because during the time it’s running, it’s slamming the memory subsystem).
- GCs have positive feedback. They are likely to be triggered when the system is under stress. Hence, they make stress conditions worse.
- GCs tend to be inelastic. That is, programs that suddenly free up a bunch of memory will tend to sit on it because the GC hadn’t run yet.
- GCs play poorly with real time locking and scheduling (even real time GCs that have a story for this have a bad story in the sense that schedulability analysis gets way nastier than it would be otherwise)
These are all good reasons to avoid GCs. Folks should know these reasons. Folks working on GCs should try to address these reasons directly.
Talking about “stopping” and “pausing” just confuses the situation by making it harder to see the real problems.
Kris Pamungkas retweeted
Replying to @samuellhuber
Where others use AI to “create”, at TigerBeetle we’re more excited about how to use the machines to “destroy”.
Some see AI as a way to “type faster” or “increase productivity”, but at TigerBeetle I tell the team we’re already “too productive” (through TigerStyle), we want calm focused work not burnout… and we know that all the huge gains come from understanding anyway.
So “destruction” aka testing, i.e. as a foil or sparring training partner, is where we see it’s at with AI.
Not to create. That stays with the humans so we don’t atrophy our understanding, which is more valuable than “LOC”.
But to increase quality through defense in depth in testing. But even there, the gains with AI are marginal, a few percent, compared to the 90% power of our DST, which again, came from systems thinking.
So I encourage our team to “keep playing the violin yourself”, to keep practicing, keep training, keep investing in understanding.
Kris Pamungkas retweeted
Replying to @oliviscusAI
How about a 544kb pdftotext binary that runs on six OSes w/o dependencies? cosmo.zip/pub/cosmos/bin/
Kris Pamungkas retweeted
Reuters confirmed that the elementary school in the Iranian city of Minab was bombed TWICE by the US military, 40 minutes apart.
This was intentional. You don't "accidentally" bomb a school TWO TIMES.
168 children and 14 teachers were killed. It was a massacre.
Kris Pamungkas retweeted
I'm hearing that a major cloud provider is giving their C-suite 18-month severance packages while the 400+ ICs getting axed next week only get standard 2-week notice. One VP is apparently walking away with $3.2M while senior engineers who built their entire ML platform are getting escorted out with a box of swag
Source close to the situation says leadership spent three months negotiating their own exit packages before even telling middle management about the "workforce optimization." The same executives who preached about "tightening our belts" during all-hands are now cashing out while people who actually shipped code get nothing

























