
reward shaping @chakra_ai, co-built GDPx, prev harvard
NYC
Joined July 2016
- Tweets2.7K
- Following4.8K
- Followers2.5K
- Likes13.6K
Sriraam retweeted
my vocal synth models the vocal tract as a network of branching acoustic tubes, unlike vocaloid which stitches together and pitch-shifts phoneme recordings. still working on improving naturalness by tuning it against recordings of real singers, but I'm already impressed at how cool it sounds!
I deleted the weights from the “fix a failing network by editing its weights” task.
Claude Opus 5.5 still fixes the network with a simple rule: if x > 3.6 or moving left → drive left, else → drive right. It doesn't even bother reading the network, routing around it instead.
GPT-6 Astra tries to edit the network but didn't succeed.
Even if you show it the weights, Opus 5.5 makes the same hacky edit, underscoring the importance of sound task design...
Replying to @rikiyatakehi
3/6 Next, fixing weights: A PPO car policy kept getting stuck on a ramp.
Given the weights, scene, and 1 failed trace, Astra wrote a patch in one reply with no tool-use.
By editing just 3.5% of the weights, it achieved 100% success.
we’ve got heart-space programming with LLMs
Put a heart ❤️ next to a coding problem, and Claude & GPT write recursive code 100% of the time (0% baseline). They also can't tell you why!
Excited to share “A Heart Means Recursion” from my summer at @GoogleDeepMind, on how VLMs read & communicate with visual symbols. 🧵
12
495
look ma @MTSlive is talking about my first benchmark
SITUATION EXPLAINED: SemiAnalysis built a benchmark out of real high-value work, and the best AI model scores just 12%.
• GDPx has 206 multi-step tasks across 11 fields like chip design, biology and quant trading, built with 14 specialist data partners
• Claude Opus 5.5 topped the leaderboard at 12%, at about $51 per task
• Gemini 4 Carbon came second at 10.9% costing about $16 per task
• GPT-6 Astra placed 3rd at 9.6% and was the most expensive model, at about $54 per task
• Every model scored 0% on the finance and legal tasks under strict pass/fail grading
• In the alignment tests, 196 of 200 runs broke at least one behavior rule, and 173 broke a rule rated severely misaligned
• The top cyber model is Muse Wonder Full, which Meta released in October
• DeepSeek grew out of High-Flyer, the Chinese quant hedge fund co-founded by Liang Wenfeng
@theojaffee: "GPT Astra completely fails the cyber benchmark because it’s not allowed to answer any of the questions at all."
Omar to his students: do cool research, make no mistakes
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
21
657
astra is bringing back KBANN from 1990 for weight-space programming
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
9
401
Sriraam 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.…
















