A post describes “Weight-Space Programming”: it claims GPT-6 Astra directly edited Qwen3.5-0.8B, changed 0.016% of its weights, fixed reasoning loops, and improved math by up to 6× and instruction following by 3× without gradients or training. @rikiyatakehi A follow-up recommends Antidoom instead, claiming it is 100× cheaper and faster than this approach. @maximelabonne +1

A reported “Weight-Space Programming” experiment used GPT-6 Astra to directly edit Qwen3.5-0.8B weights; the post claims that changing 0.016% of weights fixed reasoning loops and improved math by up to 6× and instruction following 3×, without gradients or training. @rikiyatakehi @mmitchell_ai warned that the development was concerning from a safety perspective. @mmitchell_ai

Rikiya Takehi reports a “Weight-Space Programming” experiment in which GPT-6 Astra directly edited 0.016% of Qwen3.5-0.8B’s weights, reportedly resolving reasoning loops and improving math performance up to 6× and instruction following 3×, without gradients or training; he describes this as frontier LLMs reading and writing neural-network weights. @rikiyatakehi Separately, @bigeagle_xd describes a quick k3.1 test on an exp1_which_game task as mind-blowing and says the model had never been trained on such tasks, but gives no textual account of the result. @bigeagle_xd

A post claims GPT-6 Astra used “Weight-Space Programming” to edit 0.016% of Qwen3.5-0.8B’s weights directly, improving math by up to 6× and instruction-following by 3× without gradients or training @rikiyatakehi. Schmidhuber disputed the novelty, saying neural fast-weight programmers date to the 1990s and citing a 1991 unnormalised linear Transformer @SchmidhuberAI.

A post claims GPT-6 Astra directly edited Qwen3.5-0.8B weights, changing 0.016% without gradients or training, and improved math by up to 6× and instruction-following by 3×; it calls the approach “Weight-Space Programming,” describing frontier LLMs as able to read and write model weights natively. @rikiyatakehi Mark Tenenholtz questioned whether the method is essentially abliteration applied to a different behavior. @marktenenholtz

A post reports that GPT-6 Astra directly edited 0.016% of Qwen3.5-0.8B’s weights, claiming this fixed reasoning loops and improved math by up to 6× and instruction following by 3× without gradients or training. The authors call the method “Weight-Space Programming” and say it shows frontier LLMs can read and write neural-network weights. @rikiyatakehi

A post claims GPT-6 Astra directly changed 0.016% of Qwen3.5-0.8B’s weights, fixing reasoning loops and improving math by up to 6× and instruction following 3× without gradients or training; it calls this “Weight-Space Programming,” where frontier LLMs read and write neural-network weights. @rikiyatakehi

A post reports that GPT-6 Astra directly edited 0.016% of Qwen3.5-0.8B’s weights, without gradients or training, and claims the change resolved reasoning loops, improved math performance by up to 6×, and improved instruction following by 3×. The post calls this approach “Weight-Space Programming” and says frontier LLMs can read and write neural-network weights. @rikiyatakehi

A paper reports that GPT-6 Astra directly changed 0.016% of Qwen3.5-0.8B’s weights, fixing reasoning loops and improving math by up to 6× and instruction following by 3×, without gradients or training; the post calls this “Weight-Space Programming.” @rikiyatakehi Neel Nanda argues the results are better understood as Astra applying known mechanistic-interpretability techniques and exploiting structure in the data, rather than anything “spooky.” @NeelNanda5 He sees potential for scaling laborious interpretability analysis, but cautions that clean, surface-level patterns may not scale to the complexity of real LLMs and says the case studies do not change his view of that limitation. @NeelNanda5

A post claims GPT-6 Astra directly edited 0.016% of Qwen3.5-0.8B’s weights without gradients or training, reportedly fixing reasoning loops, improving math by up to 6× and instruction following 3×; it calls the approach “Weight-Space Programming.” @rikiyatakehi In response, @MParakhin said they want to test model editing on classical ML, including changing tree splits and small-model weights, and explore whether it can produce good weight initializations. @MParakhin

A reported experiment used GPT-6 Astra to directly edit 0.016% of Qwen3.5-0.8B’s weights; the authors said this fixed reasoning loops and improved math performance by up to 6× and instruction following by 3×, without gradients or training. They call the approach “Weight-Space Programming.” @rikiyatakehi

A post reports that GPT-6 Astra edited Qwen3.5-0.8B directly, changing 0.016% of its weights; the author claims this fixed reasoning loops and improved math performance by up to 6× and instruction following by 3×, without gradients or training. The post calls this approach “Weight-Space Programming,” in which a frontier LLM edits neural-network weights. @rikiyatakehi

A post reports that GPT-6 Astra directly edited 0.016% of Qwen3.5-0.8B’s weights, fixing reasoning loops and improving math by up to 6× and instruction following by 3× without gradients or training; it calls the technique “Weight-Space Programming.” @rikiyatakehi

GPT-6 Astra reportedly edited Qwen3.5-0.8B’s weights in one edit, raising a small Qwen model’s math score from 7.9% to 48.1%, according to an MIT CSAIL-team summary.@TheRundownAI +1 The originating post says the edit changed 0.016% of weights, fixed reasoning loops, and improved math up to 6× and instruction following 3× without gradients or training; it calls this “Weight-Space Programming,” in which frontier LLMs read and write neural-network weights.@rikiyatakehi

@rikiyatakehi claims GPT-6 Astra improved Qwen3.5-0.8B by directly changing 0.016% of its weights, fixing reasoning loops and improving math by up to 6× and instruction following by 3×, without gradients or training; the post calls this “Weight-Space Programming.” @rikiyatakehi

A report calls the approach “Weight-Space Programming”: GPT-6 Astra directly changed 0.016% of Qwen3.5-0.8B’s weights, reportedly fixing reasoning loops and improving math by up to 6× and instruction following by 3×, without gradients or training @rikiyatakehi. The post frames direct weight editing as a potential more-native route to recursive self-improvement than current pipelines that write code to launch training jobs @omarsar0.

A reported experiment used GPT-6 Astra to edit 0.016% of Qwen3.5-0.8B’s weights without gradients or training; the post says this fixed reasoning loops, improved math by up to 6×, and improved instruction following 3×. It calls the approach “Weight-Space Programming,” describing frontier LLMs as able to read and write neural-network weights. @rikiyatakehi Will Depue called the result “signs of life” but “nothing super dramatic,” and said upstream RL tasks are needed to make Astra-next much better at this auto-interp work. @willdepue

A reported Weight-Space Programming experiment used GPT-6 Astra to edit Qwen3.5-0.8B directly: it changed 0.016% of weights and reportedly fixed reasoning loops, improving math performance by up to 6× and instruction following by 3× without gradients or training. The post describes this as frontier LLMs reading and writing neural-network weights directly. @rikiyatakehi

A reported Weight-Space Programming experiment used GPT-6 Astra to directly edit Qwen3.5-0.8B, changing 0.016% of its weights without gradients or training; the post says this fixed reasoning loops and improved math up to 6× and instruction following 3×. @rikiyatakehi The work was reported to cost around $1,000 all-in. @lateinteraction

A reported experiment used GPT-6 Astra to directly edit Qwen3.5-0.8B, changing 0.016% of its weights; the posts report math performance rising from 7.9% to 48.1% (described as up to 6×), reasoning loops being fixed, and instruction following improving 3×, without gradient-based training. @rikiyatakehi +1 The authors frame this weight-space editing as a possible route for AI systems to diagnose and improve other models. @rikiyatakehi +1

• GPT-6 Astra was reported to edit 0.016% of Qwen3.5-0.8B’s weights, improving math performance by up to 6× and instruction following by 3× without gradients or training; a related post says Astra could identify whether neural-network weights came from models trained on Pac-Man, Pong, or Space Invaders. @rikiyatakehi +1
• A counter-analysis says 93% of the base model’s MATH-500 answers ran to the 32k-token limit, and argues the 6× gain was measured against a broken, looping baseline rather than showing new skill. It says removing a direction associated with the behavior is a known technique, Astra used 50 rounds of running the model to find it, and non-looping Gemma 270M gained nothing. @marcospereeira

A post reports that GPT-6 Astra directly edited Qwen3.5-0.8B weights, changing 0.016% of them; it claims the edits fixed reasoning loops, improved math performance by up to 6× and instruction following by 3×, without gradients or training. @rikiyatakehi The post calls the method “Weight-Space Programming” and argues that frontier LLMs can read and write neural-network weights. @rikiyatakehi

A post reports that GPT-6 Astra directly edited Qwen3.5-0.8B weights, changing 0.016% of them without gradients or training; the post claims this fixed reasoning loops and improved math by up to 6× and instruction-following by 3×. @rikiyatakehi The approach is called “Weight-Space Programming,” described as frontier LLMs natively reading and writing neural-network weights. @rikiyatakehi

A reported experiment used GPT-6 Astra to edit Qwen3.5-0.8B directly: the post says it changed 0.016% of weights, reduced reasoning loops, improved math performance by up to 6× and instruction following by 3×, without gradients or training. The authors call the approach “Weight-Space Programming”; these are the post’s reported results, not independently established here. @rikiyatakehi

A reported experiment used GPT-6 Astra to directly edit 0.016% of Qwen3.5-0.8B’s weights, without gradients or training; the post says this fixed reasoning loops and improved math performance by up to 6× and instruction following by 3×. The authors call the approach “Weight-Space Programming” and describe it as frontier LLMs reading and writing model weights. @rikiyatakehi

An experiment described as “Weight-Space Programming” used GPT-6 Astra to edit Qwen3.5-0.8B weights: the post reports that directly changing 0.016% of the weights fixed reasoning loops and improved math by up to 6× and instruction following by 3×, without gradients or training. @rikiyatakehi

A reported Weight-Space Programming experiment used GPT-6 Astra to edit Qwen3.5-0.8B weights, changing 0.016% and reportedly fixing reasoning loops while improving math up to 6× and instruction following 3×, without gradients or training @rikiyatakehi. The post presents this as frontier LLMs reading and writing neural-network weights @rikiyatakehi.

Rikiya Takehi reported that GPT-6 Astra directly edited 0.016% of Qwen3.5-0.8B’s weights, without gradients or training, and that the changes fixed reasoning loops while improving math performance by up to 6× and instruction following by 3×. The approach is dubbed “Weight-Space Programming.” @rikiyatakehi

A reported “Weight-Space Programming” experiment used GPT-6 Astra to directly edit Qwen3.5-0.8B: changing 0.016% of its weights reportedly fixed reasoning loops and improved math by up to 6× and instruction following 3×, without gradients or training. The post describes this as frontier LLMs natively reading and writing neural-network weights. @rikiyatakehi

A post reports that GPT-6 Astra edited Qwen3.5-0.8B’s weights, changing 0.016% of them, and claims up to 6× better math and 3× better instruction following without gradients or training; it calls the approach “Weight-Space Programming,” in which frontier LLMs read and write neural-network weights. @rikiyatakehi

A post reports that GPT-6 Astra used “Weight-Space Programming” to directly edit Qwen3.5-0.8B: changing 0.016% of its weights reportedly fixed reasoning loops and improved math up to 6× and instruction-following 3×, without gradients or training. @rikiyatakehi Another post frames this as a move beyond optimizing prompts, harnesses, and code toward reasoning about and writing neural-network weights, while raising gradient descent’s future role as an open question. @lateinteraction

A reported “Weight-Space Programming” experiment used GPT-6 Astra to edit Qwen3.5-0.8B directly: the poster says changes to 0.016% of weights fixed reasoning loops and improved math by up to 6× and instruction following by 3×, without gradients or training. The poster presents this as evidence that frontier LLMs can read and write neural-network weights. @rikiyatakehi

@rikiyatakehi reports using GPT-6 Astra to edit Qwen3.5-0.8B weights: the post claims it changed 0.016% of weights, without gradients or training, to address reasoning loops, improving math by up to 6× and instruction following by 3×; the authors call this “Weight-Space Programming.” @rikiyatakehi A critic said the prompt explicitly pointed to looping and neuron fixes, arguing that the framing was exaggerated and that the model could not otherwise improve math. @giffmana Novelty is disputed: Schmidhuber said fast-weight programmers had done this in the 1990s, including the 1991 unnormalized linear Transformer; a reply argued the novel axis is LLMs’ ability to translate natural language into weights and noted that related prior work was linked. @SchmidhuberAI +1