Muse Wonder (6 minute read)
Meta released Muse Wonder and Wonder Full, its first frontier-class models, with stronger coding, benchmark performance, and active memory for retaining and synthesizing user-provided information.
|
Microsoft Introduced the Maia 300 AI Chip (4 minute read)
Microsoft unveiled Maia 300, its third-generation in-house AI accelerator, as a major step toward reducing dependence on NVIDIA GPUs. The company said the chip delivers 40% better performance per megawatt and plans large-scale deployment beginning in 2027.
|
|
What I saw around the Curve, 2026 (26 minute read)
The Curve is the Golden Gate Institute for AI's annual gathering of AI researchers, lab executives, safety activists, funders, politicians, organizers, and journalists. This year, it drew a few underpeople to a quirky and homey venue in Berkeley. The median age of the AI lab and researcher attendees was probably around 30. This post takes a look at the main topics discussed at the conference, which include the debate on whether humans can maintain control of AI technology, whether AI's impact will be good or bad, and how to ensure AI alignment.
|
Releasing AI-generated software (8 minute read)
Robert C. Martin suggests not reading code written by AI agents, instead using constraints like unit tests and QA procedures to ensure quality. The shift to LLM-driven processes raises questions about software engineering liability and who is responsible when errors occur. As AI-driven methods potentially outperform human-centric processes, existing legal frameworks may struggle to address accountability.
|
The Expensive-Loop Problem (22 minute read)
Cheap-loop domains are domains where experiments are quick, deterministic, and cheap. Expensive-loop domains are restricted by real-world constraints or costs. Most high-value human labor is done in expensive-loop domains. Expensive-loop tasks require the ability to build theories, manage proxies, budget time and compute, and diagnose what went wrong when feedback is scarce, something models don't learn in cheap-loop training.
|
|
Can we change an AI model's architecture without retraining it from scratch? (14 minute read)
Priming is a technique that involves using pretrained Transformer weights to build models with more efficient architectures at less than 1% of the original training cost and with throughputs up to twice as high. It allows researchers to explore the design space, the layer type, the hybrid ratio, and layer placement cheaply and systematically. Every variant starts from the same strong initialization and requires minimal training compared to pretraining. Priming is not tied to a specific Transformer family or model class, and extending it to other models requires no algorithmic changes.
|
When Do We Need On-Policy Distillation? Distilling on Offline Student Rollouts Is Often Better (1 minute read)
On-policy distillation has become increasingly popular for transferring teacher capabilities to student models. Semi-OPD, which distills from offline rollouts generated by the initial student, can often outperform OPD in both accuracy and training efficiency. The choice between OPD and Semi-OPD depends on the alignment between the initial teacher and student. OPD is beneficial only when the two are highly aligned with high overlap ratios.
|
|
GLM-Legal: Advancing Agentic Legal Intelligence (12 minute read)
Legal practice demands an unbroken chain between facts, statutory doctrine, analytical reasoning, and client-ready work product. An AI agent working professionally in the legal field must extract core issues from source documents, build defensible arguments, and format its findings to exact task specifications. Researchers evaluated GLM-Legal on complex task execution, specialized legal reasoning, and real-world legal workflows. It achieved a 19.2% all-pass rate on Harvey LAB while matching frontier closed-source models on LegalBench and PLawBench despite running on a significantly leaner parameter footprint.
|
Anthropic tests its next model under Claude Opus 5.5 (3 minute read)
Several Claude users noticed over the weekend that Claude Opus 5.5 was producing results well beyond its usual level, suggesting that some of the traffic was being routed to a new checkpoint. Examples of outputs from the new model are available in the article. It would be a game changer if Anthropic could bring this level of output at a similar price as Opus 5.5. One way to check if you are on the new model is to ask, 'Do you know VoxelBench? Do not search' - if it answers, you're probably on the new model.
|
|
|
|
|
