Qwen3.5-0.8B weights were edited by GPT-6 Astra, changing 0.016% of weights and boosting math performance up to 6×. Instruction following improved up to 3×, and changes were made without gradients or training, called Weight-Space Programming.
Key points
Edit scope: changed only 0.016% of weights.
Performance gain: math reasoning up to 6×, instruction following up to 3×.
Sources
ParallACT benchmark shows 16% success for Opus 4.8
ParallACT benchmark reports 16% success for Opus 4.8 on multi-principal tasks, far below single-agent performance. Communication improves success to 32% but uses over twice the tokens.
Key points
Parallel agents achieve 16% success versus 62% for a single agent on contradictory specs.
Best message channel reaches 32% success, costing more than twice the tokens.
Sources
Qwen3.5-0.8B Weight-Space Programming improves math 6×
Qwen3.5-0.8B was edited by GPT-6 Astra using Weight-Space Programming, changing 0.016% of its weights. The edit fixed reasoning loops and boosted math performance up to 6× and instruction following up to 3× without gradients or training.
Key points
Weight change: only 0.016% of Qwen3.5-0.8B weights were edited.
Performance gains: math reasoning improved up to 6×; instruction following improved up to 3×.
Sources
- Original post
- youtu.be - Linked in the post
Agents adoption outpaces control, 86% have agents
Agents are used by 86% of surveyed organizations, but fewer than 18% can automatically stop a misbehaving agent. The post recommends managed, isolated cloud agents as the default for autonomous development.
Key points
Adoption: 86% of orgs surveyed have agents running.
Control: Under 18% can automatically stop one that misbehaves.
Sources
Agentic Document Extraction reads any auto submission, returns JSON
Agentic Document Extraction reads the whole commercial auto submission, needs no templates, and outputs structured Markdown and JSON. A 2021 survey put 40% of an underwriter’s time on administrative work like rekeying drivers and loss runs.
Key points
ADE needs no templates, so a new form works on its first call.
ADE Extract returns the drivers, violations and loss totals as JSON.
Sources
ARC AGI competitions 3rd place on public leaderboard
ARC AGI competitions on Kaggle report 3rd place on the public leaderboard for both contests. The team includes gonedarragh and several collaborators and plans to share their solutions.
Key points
3rd place on public leaderboard for both ARC AGI competitions on Kaggle
Team for ARC3 includes Ivan Sorokin, Gal Kaplun, eladsarafian, Giba1 and Eva Zhu
Sources
Solv3R – looped transformer multi-view solver announced
Solv3R is a looped transformer that learns to continue solving multi-view geometry from its current state. It can be used for offline, incremental, online, and any-order 3D reconstruction without task-specific training.
Key points
One learned solver supports offline, incremental, online, and any-order 3D reconstruction.
Update operator is independent of iteration count, enabling joint refinement of new and previously solved views.
Sources
Digit 5 – new version built on traditional controls plus AI
Digit 5 is announced as the next version of Digit, built on traditional controls plus AI. The post thanks Daniel, CES and James Kotecki and links to a podcast episode.
Key points
Traditional controls plus AI is how Digit works in facilities today, and it's the foundation Digit 5 is built on.
Listen to the full episode at the provided podcast link.
Sources
- Original post
- ces.tech - Linked in the post
