cyberhandcrafting boltwright /// byteslinger humbly serving atomwranglers /// prev cofounded @pulley and grew it to $15M ARR /// by all means build
Calgary
Joined December 2008
- Tweets5.9K
- Following2.2K
- Followers1.5K
- Likes60.4K
"v high precision" that's that trick innit. even for @pangram it's nearly impossible to reach high precision on such short inputs
although hopefully they'll run automated checks against the entire reply history? that should make it easy to get to v high precision
case in point:
1
867
astonishing result in this paper. it's like if a neurosurgeon understood what each neuron was for while operating on you.
kinda makes you wonder what would happen if you gave astra access to a TMS machine pointed at your head and asked it to make you smarter. 😏
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
41
openai is holding fast so they can drop a megamodel the day of the anthropic ipo 😉
don't shoot until you see the whites of their eyes
1
1
110
added "crack jokes and use lots of profanity" to my system prompt today.
best idea i've had in a while.
75
ironic that so many of the replies on this post are AI reply bots 😂
i keep blocking the AI reply bots but there's just an endless supply of them. i guess i should automate filtering them out with an extension?
the signature is funny, claude has been RL'd so hard to follow the anti-spam laws to the letter.
2
152
i love the smell of ultracode in the morning
DON'T FORGET: 2 things for Claude Max subscribers
First, if you haven't used your banked reset, they expire in ~36 hours.
Second, you also should claim and use your $250 Cloud Credits! These expire Nov 5th.
Cloud credits are great for managing issues/PRs in your repo. Ask Claude to kick off a workflow to manage your Github -- give it your acceptance criteria, ask it to evaluate, and suggest fixes for high quality submissions.
Cloud sessions are just 👌 sooo much nicer than spinning up 20 worktrees trust me. The $250 went further than expected.
1
3
165
@scottastevenson is spellbook still on venn? how do you like it?
trying to decide between that and a few other options like wealthsimple business and airwallex.
is it agent-friendly?
2
2
204