Several research groups report that frontier large language models can generate or directly edit the weights of small neural networks without gradient training or external tools, a claim that has drawn both excitement and pushback. Alex Krentsel's replication of an MIT experiment found a sharp capability jump across model generations, with the latest models producing networks that scored up to 22 times an RL-trained baseline. Hugging Face separately reproduced the paper using ML Intern, its open-source AI research harness, and released the model and a demo. In another test, a user had Fable hand-write a 100,000-parameter network to distinguish real legal citations from fake ones, and the result reportedly outperformed a trained version of the same classifier.
Scepticism is growing. Critics argue that some claims are overstated, that many lack competent baselines, and that fitting weights directly against a benchmark is essentially meaningless. A reconstruction from logs by ZhihuFrontier of GPT-6 Astra's edit of 131,072 weights in Qwen3.5-0.8B suggests the headline gain is narrower than reported. The reported MATH-500 score rose from 7.9% to 48.1%, but 576 of the 579 repaired answers had originally hit the 32K-token limit because the model was looping. On 1,152 held-out problems, accuracy rose from 10.6% to 60.4%, which the reconstruction reads as mostly existing reasoning being allowed to finish rather than new capability.
The week also brought a steady run of model and benchmark news. Tencent's Hy4 scored 75.2 on APEX-Agents against 72.1 and 72.9 for DeepSeek V4.1 Flash-1017, with API pricing of $0.834 and $2.501 per million input and output tokens, and Tencent plans to release its weights. On Artificial Analysis, Claude Opus 5.5 (max) leads all six Capability Indexes v2.0, and Kimi K3.1 Pro leads among open-weight models. For Australian businesses and professionals, the weight-editing claims change nothing practical yet, since they remain unproven outside controlled tests and should not shape product or procurement decisions. Model selection is the more relevant takeaway, because per-task costs vary widely: Hy4 cost $0.55 per task against $1.46 for Kimi K3.1 Pro (Max) in Agent Arena, with scores within a point of each other. Test candidates on your own workloads before committing, and if you send personal information to any hosted model, your obligations under the Privacy Act 1988 apply whichever model you choose.
Key Points
Several groups report frontier LLMs can generate or edit small neural network weights without gradient training or tools.
Cross-generation replications show capability rising, with the latest models' networks scoring up to 22 times an RL-trained baseline.
Critics say the claims lack competent baselines and that fitting weights directly to a benchmark is largely meaningless.
A reconstruction of GPT-6 Astra's Qwen3.5-0.8B edit suggests much of the headline gain comes from reasoning that can now finish instead of looping.
Tencent's Hy4 and Artificial Analysis updates show close performance but large cost differences across models, so workload-specific testing matters.