Here's some @SloydAi-generated 3D pixel art for the weekend.
Mars station signing off.
19
Avi Latner | Sloyd retweeted
4× more of our PRs pass first review now, since we started running one skill at the end of every coding session.
We just open sourced it. It's called make-native.
AI love writing new code. And bad at making it look like it belongs: a flag here, an if (isNew) there, a copy of a component "just for this case".
make-native runs at the end of a session and:
→ finds the bolted-on code
→ finds the helpers, patterns and types that already do the job
→ rewrites the feature to use them and deletes the rest
Works with Claude Code, Codex, Copilot, Opencode ++
Free. Link below 👇
This is getting very interesting 💡
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.…
20
Avi Latner | Sloyd retweeted
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.…
Just talked with a fellow game dev who said the AI debate in game dev is over. Yet Kenney is sticking to his guns. I respect him, he's made great stuff. But the industry is adjusting, and it's time to adjust with it.
Don't let billionaires convince you slop is the future. They're used to managing & directing, not making things themselves. Of course they can't see the difference. They don't understand your love for the craft. Keep making personal, beautiful games.
You've got my support 💖
74
@typesafeai Jev image classification is 3 days old. I benchmarked it against Haiku 5.5 on real production data. We know it's faster, but are the results as good?
Turns out they are.
2
2
67
By 2030 there will be ten more AI labs. But most enterprises will run open-weight models on their own GPUs.
43
Here is one of my favorite methods to improve decision making.
We plot every feature on a chart showing adoption in its first 60 days. The rate of change, even after two weeks, tells us if we got our priorities right.
1
42
I used Sloyd's video AI to make a video about Sloyd's video AI.
Very meta 😂
@SloydAi
We started Sloyd with 3D modeling, but it's only one type of creators need. Overtime, we've expanded into image, animation, skyboxes, scenes.
Now, we're adding AI video and audio.
1 location for media gen is the base for a harness: a layer for using memory rich with context.
Sloyd makes great 3D models.
And now they move. And talk. 🐸
Video + Audio is live in Sloyd. 3D, images, video and audio in one subscription.
Zero model shopping.
1
61
In product and at startups, working from first principles was all that mattered.
Tools were not important. You had to use best practices, but being the first to adopt the newest tools was a distraction.
Now, with AI, tools make a huge difference. You gotta keep up with the latest
1
26
בתי כלא = מתקני עינוים
פלנגות במימון המדינה בגדה
פשעי מלחמה בעזה
ומהשבוע רדיפה והפקרת דם של עיתונאים
אנחנו גרועים כמו האויבים הכי גרועים שלנו
אנחנו האויבים הכי גרועים שלנו
20
I left my coding agent on Opus 5 + Max reasoning and forgot about it.
A tiny task took 5 min and 25 credits. Same task on Auto: 40 sec and 2 credits.
Max for the hard stuff. Auto by default. That’s 7.5x less waiting and 12.5x fewer credits.
28
