Wait, what?! let me try.
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.…
1
48
Those are rookie numbers, push Claude it can do 40tok/s
DeepSeek-V4.1-Flash, running at home on my RTX 5090.
~27 tok/s decode, ~225 tok/s prefill, 274 GB of local weights.
Speed still needs work, but this is very promising. Looking forward to more.@coldniko
muse.ai/s/report-xoj6qgwxvxx…
1
35
Those are rookie numbers, push Claude it can do 40tok/s
DeepSeek-V4.1-Flash, running at home on my RTX 5090.
~27 tok/s decode, ~225 tok/s prefill, 274 GB of local weights.
Speed still needs work, but this is very promising. Looking forward to more.@coldniko
muse.ai/s/report-xoj6qgwxvxx…
32
Chrome is the best engineered sandbox on this planet....
app.opencve.io/cve/?product=…
If models can crack this typical enterprise app written over 20years is f....d.
86
We've taken a wrong path in a way to AGI, Super Intelligence just answers the question, it does not spawn a team of PhDs that fixate on random small problem in the data for 4h!
25
David Deutsch fixing a dishwasher and you think with AI plumbers will have a job? think again.
I just fixed my dishwasher with the help of ChatGPT. A trivial task. I had been about to order a new one. So this software has increased the country's real wealth yet decreased the measured GDP. The main economic indicator is structurally incapable of registering the thing that actually makes people better off, namely the growth of knowledge
1
71
Looks like we need to cancel Ant, OAI and Grok subscriptions and pay some Chinese labs ... Regulatory capture is real, just think how difficult it is to start a new pharma company
25
All this while EU is buying almost 100% of Yamal LNG production ...
Russia is escalating attacks on Ukrainian civilians.
July and August were the deadliest months since the war started.
But Ukrainians are standing strong. We must stand with them.
That’s why air defence is our top priority.
Some countries have interceptors that will soon expire. Those could be sent to Ukraine to help protect Ukrainian lives.
My press remarks ahead of the meeting of Defence Ministers ↓
1
56
Running only one agent at a time without git worktree, NGMI in 2026, we moved ok
Yes each of my sites is a Termius server with this kind of start command
cd /srv/http/hotelist.com && tm
tm is a script made by Claude Code which puts it in a tmux session tied to the project name (from its folder)
That means I always log back into my tmux session for each project!
My big sites like Photo AI are on their own VPS, one site per VPS, but the rest is on one VPS, all on Hetzner btw, so anyway I don't even notice this, it's just all tabs in Termius
If my phone disconnect I just reconnect and am back where I was
Every tmux session has Claude Code open
1
85
We really need to learn why these hypothesis are important ;-)
8 days ago, while jogging, I asked Claude to solve the Riemann Hypothesis
It didn’t. 1.5 days later, it proved >= 67% of the zeros are on the line (prev: 41.6%)
Still not sure what that means, but some analytic number theorists seem excited
anthropic.com/research/riema…
1
149
Morning Idea: Xteink like device, e-ink screen + microphone, for the moments when I just want to ask a research question and read a short book on a topic
110
