alex (@alexeeps) profile picture

computers and the like

Joined June 2020
Pinned Tweet
out larping
1
14
1,116
wdym im compiling rust in my browser
weekend quest: <unkernel> A real Linux kernel, compiled to WebAssembly. Programs run on a fast RISC-V CPU with a wasm JIT. It works on Node.js, Browsers, and even iOS.
3
138
4,272
OpenAI was doing this btw but had to stop since it started tinting their images yellow
why can’t the ai data centers just use pee instead
16
105
1,664
14,843
cool little transformer trained on my listening history
3
163
i somewhat doubt the next generations will be able to tell real and AI imagery apart
1
66
oh shut up
3
86
cf cli is fucking awesome, my agent just deployed this website http://localhost:5173
1
9
314
it actually deployed my app and handled the DNS and setting up a preview environment as well tho, it’s really cool
17
you can now commit using tinygit!
i got refs working yesterday here's git log! (well, some of it)
8
261
first meeting with my supervisor and i want to DROP OUT that’s normal right? it builds character or something
1
9
190
is there some sort of @pangram bot for github? I don't personally want it because I don't have big OSS projects but it feels like there's a market gap in detecting slopped up issues / PR descriptions which is one of the reasons we see so many projects disabling them
2
121
parquet packfiles (or parqfiles as I like to call them) are 17% smaller, but packfiles are generally faster to work with this is storing non deltified objects in parquet and relying on zstd (level 9) for compression (ps: implementation is vibe coded, I don't have time)
parquet git storage backend?
69
parquet git storage backend?
1
1
116
i feel so sigmoid
5
173
Sometimes on campus you accidentally walk by a Comp Sci class and the professor is writing like SELECT title, genre, price FROM books WHERE (genre = 'Fantasy' OR genre = 'Sci-Fi') AND price < 12.00; on the board and everyone is taking notes like its actual school
40
158
24
3,737
105,881
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math. rikiyat.github.io/blog/weight-sp… github.com/RikiyaT/lm_und… huggingface.co/Qwen/Qwen3.5-0… huggingface.co/RikiyaT/Qwen3.…
2,367
7,260
5,751
55,935
10,049,883
who's building a self improvement layer for apps? i don't want it, i just think it'd be cool a library you integrate into your app (whatever form of app it is) which takes care of managing feature requests / bug reports, maybe monitoring itself and redeploying
1
2
116
1/3 A CPU inside an LLM is one thing, but running another LLM inside an LLM would be taking things a bit too far, wouldn't it? Anyways, here's an LLM running inside an LLM : QwenPU, an LLM (@karpathy's stories260k) running on an NPU compiled inside Qwen 3.5-4B's weights.
1/4 Meet Qwenputer: Qwen3.5-4B with a full RISC-V CPU inside its weights, using the model's own Gated DeltaNet layers as RAM. It runs C and MicroPython in its forward passes, no external runtime. On DeepMind math: base 55–90%, base + Python tool 95–100%, Qwenputer 97.5–100%.
153
257
115
2,243
177,489
i love formal verification but i really do feel like it's a trend that'll go by like graphql
1
5
178
Replying to @gnukeith
resource constraints drive innovation
3
8
234
23,097
will anyone i know be at FOSDEM in January?
1
70