무인 3D 개발, 다중 가속기 PyTorch, LLM 신뢰도와 모델 효율

Astra seems to be doing okay (tm) with driving Unreal Engine development headless and running verification loops. It's slow, but it's unattended and incremental!

Astra has successfully beat all 48 levels of the “I’m Not a Robot” game:

Live from Shanghai: PyTorch Foundation Executive Director @sparkycollier is speaking at the FlagOS-hosted “Open Computing: Building Open Software for Diverse Hardware,” a sponsor event co-located with KubeCon + CloudNativeCon + OpenInfra Summit + PyTorch Conference China 2026. The session brings PyTorch Foundation into a technical discussion with experts from @BAAIBeijing Shanghai AI Laboratory, @vllm_project , @sgl_project , @nvidia , and others on open system software across increasingly diverse hardware. AI hardware is becoming increasingly diverse. That creates new demands across the software stack, from operators and compilers to frameworks and inference, while repeated backend-specific integration becomes a bottleneck. The scale of PyTorch and its surrounding open source communities is measurable. torch saw 80M+ PyPI downloads last month, PyTorch had 3,929 active contributors over the past year, up 44%, and vLLM contributor growth reached 4.5x, from 740 in December 2024 to 3,400+ today. The session examines open system software spanning multiple chip architectures, including operators, compilers, training and inference frameworks, a communication library, and hardware-software co-optimization. 🔗 See the full program and session schedule: https:// lfopensource.cn/kubecon-cloudn ativecon-openinfra-summit-pytorch-conference-china/program/schedule/ …

📸 @sparkycollier provides the opening remarks for “Hardware-Aware AI: Building PyTorch Workloads Across Accelerators,” hosted by @Huawei . Huawei’s co-located session focuses on PyTorch across GPUs, NPUs, and XPUs, including framework adaptation, operators, compiler

Do LLMs passively report confidence—or do they use it to guide behaviour? Our @NatMachIntell paper provides causal evidence for the latter: boosting/suppressing confidence changes whether models answer or abstain. Open access: https:// nature.com/articles/s4225 6-026-01293-x … @GoogleDeepMind @Princeton

// Evaluating and Improving LLM Self-Modeling // Really interesting paper. Can a model answer questions about its own behavior? The questions are deliberately verifiable, such as whether a particular prompt edit would change the model's final answer. This framing avoids the

Alexandr Wang

@alexandr_wang

17시간 전

interesting eval

keep the reports coming, we do read them

We used TeamAI-CLI internally at Tencent since March. Open sourced it. It turns team knowledge into one git repo, so every agent works from the same handbook. what it does: > git-based: skills, rules and docs live in one repo, changes go through a merge request > hook-triggered: merged changes land on everyone's next session > each learning earns confidence from real usage, strong ones surface first, weak ones sink > works with claude code, codex, cursor, opencode, codebuddy and workbuddy One person's hard-won workaround can now become the whole team's default.

6 ways production LLMs encode token positions: (bookmark this) A plain Transformer has no built-in sense of token order. Without positional information or a causal mask, the same three tokens in a different order produce the same output vectors, just rearranged. This is