AI models editing AI weights: what MIT's GPT-6 Astra result means
Aggiornato
Foto illustrativa
A frontier model that reads and rewrites weights
MIT researchers led by PhD student Rikiya Takehi report that GPT-6 Astra could read and directly adjust the weights of a smaller model. The test subject was Qwen 3.5-0.8B, and the target was reasoning. Astra touched only 0.016% of the weights. That tiny edit, per the researchers, lifted Qwen's math benchmark scores sixfold and instruction-following threefold. The abstract argues frontiermodels are surprisingly good at natively reading and writing neural network weights, beyond what humans or conventional algorithms manage.
The same paper claims Astra can tell what a network was trained to do just by looking at its weights. Without any environment feedback, it wrote small game-playing networks from scratch that outplayed networks trained with PPO for 100 million steps. The researchers flag interpretability as dual-use: models that are good at optimization are good at optimizing themselves.
Anthropic's filing puts a number on the worry
Anthropic's updated S-1, released Monday morning, added language on page 29. According to aggregator Andrew Curran, internal estimates suggest the company may imminently meet its own RSI danger threshold: as much progress in one year as two years of the 2018–2024 baseline. The filing also says Anthropic does not anticipate safely using Claude to fully autonomously design and build successive frontier models without meaningful human supervision or understanding. The filing adds that the company is considering updates to its responsible scaling policy.
The chart investors are staring at
The Wall Street Journal ran an Anthropic chart showing AI now leads 26% of research. The chart was first published in September on August numbers and was reshared over the weekend by former AI Czar David Sacks. Anthropic co-founder Jack Clark called RSI possibly the most important technical trend in the world, per the Journal's framing. Financial professionals are treating RSI as a key factor for the IPO.
Why the subsidy debate is back
The RSI story lands next to a reopened argument about AI subsidies. The bear case: once companies charge what tokens actually cost plus margin, users may not value the intelligence enough to keep spending at current levels. That would put valuations in question. The AI Daily Brief host says he is not an AI bear but welcomes the bears, arguing loud criticism lowers bubble risk. Separately, Elon Musk's response to Grok Bot users hitting usage limits landed badly. Users compared it to paying for a 12-inch pizza and getting six.
What to watch before you believe the curve
The MITresult is a single experiment on one small model, and the numbers come from the researchers, not from independent replication. The Anthropic threshold is an internal estimate in a filing, not a published measurement. So treat the sixfold and threefold gains as a signal, not a settled fact. What's worth tracking next: whether other labs reproduce weight-level editing, and whether Anthropic's responsible scaling update lands with concrete supervision rules.
Link affiliati: possiamo ricevere una commissione sugli acquisti idonei, e non cambia i nostri dati né le classifiche. Aprono i risultati di ricerca di Amazon: prezzi e disponibilità si vedono lì.