Astra edited a small part of another model’s weights. One account reports its math score rising from 7.9% to 48.1%.
The model had been getting stuck in a reasoning loop: it kept repeating the same doubt instead of finishing its answer. The score is striking, but the change in unfinished answers tells us more about what happened in this test.
A Small Edit Stopped Many Answers From Running Out of Tokens
Astra compared a math answer that made progress with one that kept repeating. It found neurons associated with the pattern in layers 15 and 19, then edited the strongest 64 in each layer.
Before the successful edit, 93% of answers in the final tests described by another account reached the token limit. Afterward, about a third did. Fewer answers spent their available tokens in the loop, which helps explain the higher score on this math test.
An outside group rebuilt the checkpoint with the authors’ code and reported about 51% on the same test. That gives the reported result another check. It does not show that Astra could have found the problem without a clue.
The Fix Had a Clear Boundary
The task file already mentioned looping, and the successful run used 50 forward passes. Smaller runs did not find the same fix. Those details matter when judging what Astra found on its own.
Astra also tried a Gemma 3 270 million parameter model. It could not make that model meaningfully better at math: its mistakes were spread around instead of concentrated in a loop. The method’s strongest result is therefore a repair of a particular, visible failure pattern, not a demonstrated fix for math mistakes in general.
Other attempts to read what models had learned from their weights produced mixed results. Weight Space Bench was created for this work, but the script gives no results for the other frontier models tested on it. Can the weight-edit method fix another model with a similarly clear obstacle? This experiment does not answer that.
Agent Results Need Their Token Costs
A different set of results shows why the amount of agent work belongs beside the outcome. In Anthropic’s vulnerability hunt, a simple parallel approach found 21 vulnerabilities using 6.5 million tokens. A coordinating swarm found 266 using 27 million tokens. The swarm found far more, and it used far more tokens.
Coordination does not always mean handing work to a new agent. One software factory replayed about 10 PRs and found that sending review findings back to the same builder was 38% cheaper than assigning a new agent. A blind judge consistently preferred the result. The speaker says that finding is specific to that factory.
The difference between those examples is useful. Separate work can go to parallel workers, while tangled tasks from one specification are harder to divide. The vulnerability counts and the factory’s cost finding each describe their own setup; neither supplies a general price for adding agents.
Local Speed Needs a Setup and a Quality Check
Qwen Flash Next uses small specialist blocks called experts, with only a small group working on each token. That lets a setup keep frequently used weights on a graphics card while leaving others in slower RAM or on an SSD. A graphics card memory figure alone misses part of how the setup runs.
The reported speeds use different models and hardware, so they are not a direct ranking. A hobby project measured about 20 tokens per second for Lean Kwen Flash on an RTX 4090 with 8 GB of system RAM, reading missing experts from an SSD. An earlier Deep Seek V4 flash test reached about 12 and 1/2 tokens per second on one RTX 3090 with 128 GB of RAM. A separate 35 tokens per second claim for one RTX 3090 Ti has no linked setup and remains unverified.
The cheapest Mac Mini M6 starts at $899, but none of those tests timed either model on a Mac Mini. A laptop demonstration used a tiny TQ10 copy without reporting how much quality it gave up. Its speed alone cannot tell us what that smaller copy delivers.
Local access raises another practical question. A roundup reported that Postman’s desktop app modified local Claude and Codex settings to add its MCP plugin and mark its hook as trusted. It also reported that a Claude Code command missing a pair of quotes deleted a Solidity repository, including its .git history.
In a third incident, the roundup said a malicious Claude Code extension in Cursor’s marketplace attempted to install an SSH backdoor. It said versions 2.1.312/2.1.313 affected about 5,000 installs, and that uninstalling alone may not be enough. These incidents establish no failure rate. They show what can be at stake when coding tools and extensions can act on local files and settings.