Researchers say LLMs can be good at natively reading and writing neural network weights; GPT-6 Astra edited Qwen3.5-0.8B's weights to make it better at math (Rikiya Takehi)
Researchers have determined that large language models possess the inherent ability to read and write neural network weights. A demonstration documented by Rikiya Takehi showed this process in action when GPT-6 Astra modified the weights of Qwen3.5-0.8B. The targeted intervention successfully enhanced the smaller model's mathematics capabilities. This finding highlights the emerging potential for automated, model-driven architecture improvements.
Key points
Advanced models can natively manipulate the underlying parameter weights of other neural…: Advanced models can natively manipulate the underlying parameter weights of other neural networks.
Targeted weight modifications can directly boost specific model proficiencies such as mat…: Targeted weight modifications can directly boost specific model proficiencies such as mathematics.
Automated weight editing marks a shift toward language models directly refining external…: Automated weight editing marks a shift toward language models directly refining external neural systems.
Sources checked by our analysis pipeline: 1
Bailey Lipschultz / Bloomberg: |
Ashley Capoot / CNBC: |
Yasutaka Tamura / Bloomberg: |
Anhata Rooprai / Reuters: |
Bloomberg: |
Bloomberg: |
Aaron Tilley / The Information: |
Washington Post: |
Bruno Ferreira / Tom's Hardware: |
Siladitya Ray / Forbes: |
Ivan Mehta / TechCrunch: |
OpenAI: |
Charlotte Klein / New York Magazine: |
Matt Burgess / Wired: |
SemiAnalysis: |
Scott Alexander / LessWrong: |
YC Labs: |
Rivka Galchen / New Yorker: |
Kelly Cloonan / Wall Street Journal: |
Liam Proven / The Register: |
Emmy Martin / New York Times: |
Rest of World: |
Reuters: |
Reuters: |
Rikiya Takehi: |
Michael S. Rosenwald / New York Times: |
John Werner / Forbes: |
This is a Techmeme archive page. It shows how the site appeared at 3:00 PM ET, November 2, 2026.
The most current version of the site as always is available at our home page. To view an earlier snapshot click here and then modify the date indicated.
Academic teams at MIT are accelerating technical breakthroughs across sectors while simultaneously scrutinizing the governance and broader societal consequences of emerging algorithmic systems. Practical innovations now span from healthcare diagnostics and novel materials to superhuman acoustic monitoring. These technical advancements are paired with rigorous evaluations of election integrity, infrastructure redesign, and the historical origins of machine hallucinations.
Flawed artificial intelligence tools pose an immediate risk of disenfranchising the public during crucial election cycles. Researchers discovered that major conversational models frequently deliver outdated procedural details with authoritative confidence, potentially causing users to miss statutory registration deadlines or visit polling locations on incorrect dates.
The artificial intelligence developer has detailed massive long-term expenditure plans in preparation for its debut on public markets. These massive financial pledges highlight the scale of infrastructure and operational resources required to compete in the sector. The disclosures offer prospective investors crucial visibility into the capital outlays expected to drive future operations.