The Groundy archive · Page 34 of 43
Browse Groundy's complete archive of 1027 articles on AI, developer tools and infrastructure. Page 34 of 43.
793–816 of 1027 articles · Newest first
LLM-written CUDA kernels compile, run, and pass smoke tests while returning wrong numerics, so crash-free execution is not enough to trust AI-generated GPU code.
The RoboSSM preprint swaps the transformer backbone of in-context robot imitation for a Longhorn state-space model and claims LIBERO gains. The full paper is still pending.
Routing architecture decides whether MoE experts can be pruned safely: hard routers keep calibration per expert, soft routers need an aggregate guard.
A June 2026 arXiv paper encodes agent obligations and prohibitions as deontic policies enforced by a logic engine outside the LLM, producing a record auditors can inspect.
GLM-5.2 and Kimi K2.7 Code shipped a day apart as open-weight coding models. One leads the open-weights index; the other wins on price and a claimed token-efficiency gain.
Per-block linear recoverability in transformer FFNs swings from near-linear to strongly nonlinear between layers, so compression tools must probe each block per checkpoint.
Cursor is being acquired by SpaceX for $60B. Windsurf was absorbed into Cognition and rebranded Devin Desktop on June 2. Both vendors have changed; the tools have too.
A June 2026 preprint finds essay quality is linearly decodable from LLM internals, but cannot show whether that signal tracks argument quality or just length and fluency.
Every GLM-5.2 benchmark figure, from 62.1% SWE-bench Pro to 99.2% AIME, traces to Zhipu's own evaluations. This guide separates vendor-run scores from external evals.
GLM-5.2 ships under MIT, removing the Llama usage-threshold audit burden, but finance and healthcare teams still face compliance gaps when self-hosting this 753B MoE model.
GLM-5.2 scores 81.0 on Terminal-Bench 2.1, trails Claude Opus 4.8 (85.0) on shell tasks, but wins on 1M-token monorepo context. Here is how to route tasks.
GLM-5.2 posts 62.1% on SWE-bench Pro and 81.0 on Terminal-Bench 2.1, four points behind Opus 4.8. MIT weights are self-hostable; flat plan starts at $18/month.
GLM-5.2 ships 753B parameters under MIT with strong coding benchmarks, but the full MoE weight load makes self-hosting far heavier than the license terms imply.
GLM-5.2 weights are live on HuggingFace under MIT license: 753B MoE, 1M-token context, FP8 and BF16 variants. How to pick a deployment framework and model the hardware cost.
GLM-5.2 migration starts with the client's provider settings. Cline and Roo Code have different documented configuration paths, and neither path establishes that every Cursor feature follows the same endpoint swap.
STAR uses attention-derived spatial maps to replace uniform scalar reward in diffusion RL, shifting the bottleneck from preference pair volume to reward localization.
Zhipu shipped GLM-5.2 with a 1M-token context window the day after the US ordered Anthropic to cut foreign access, with MIT-licensed weights promised within a week.
A June 2026 preprint localizes Gemma 4 repetition loops to a few MLP neurons and removes them with a one-time weight edit while benchmark scores hold.
Zhipu shipped GLM-5.2 on June 13 with a 1M-token window and an Anthropic-compatible endpoint, but published no benchmarks and keeps the hosted API on a paid plan.
AWS Bedrock's provider_data_share gate for Mythos-class models removes the in-AWS data boundary regulated teams bought it for, pushing them toward self-hosted serving.
Vercel's Remend packages the partial-markdown repair every streaming chat team hand-rolls, but the shared heuristic can mask upstream token-boundary defects.
Moonshot's Kimi K2.7 Code loses 11 of 12 benchmark cells to GPT-5.5 and Opus 4.8, leading on token efficiency and price, which pushes buyers to run their own evals.
Two June 2026 preprints claim formal safety guarantees hold without a capability tax in low-dimensional robotic control, sharpening the attestation-versus-verification gap.
A June 2026 MLSys paper breaks vLLM cold start into six CPU-bound boot phases, showing why scale-to-zero serving forces operators back into warm GPU pools.