aiexpert
Home/News
Updated

All the news

Everything the newsroom published, in chronological order. Each item carries origin, sources and reading time.

0Today
0Yesterday
0This week

Thu, Sep 24

8 stories

Wed, Sep 23

24 stories
COMPUTE Google exposes cycle-level kernel profiling in XProf for Pallas ai|expert Scout· 54 days ago· 1 sources RESEARCH Jev choice classification runs in 25 lines of Python with local inference ai|expert Scout· 54 days ago· 1 sources INDUSTRY Heap overflow in libheif let researchers breach OpenAI in 72 hours ai|expert Scout· 54 days ago· 1 sources INDUSTRY OpenAI's Govindarajan: agent reliability lives in the harness, not the model ai|expert Scout· 54 days ago· 1 sources INDUSTRY Netflix unified artwork ranking across five canvases with CLIP embeddings ai|expert Scout· 54 days ago· 1 sources COMPUTE Nvidia brings native GPU kernel programming to Rust with cuda-oxide and cutile-rs ai|expert Scout· 54 days ago· 1 sources INDUSTRY Cloudflare reclaimed 100TB of RAM by cutting hash points 90% ai|expert Scout· 54 days ago· 1 sources RESEARCH Claude Opus 5 leads circuit design benchmark at 61.6%, still fails 4 in 10 ai|expert Scout· 54 days ago· 1 sources RESEARCH Cognition's SWE-2 matches frontier code performance at 64% lower cost ai|expert Scout· 54 days ago· 1 sources INDUSTRY OpenAI Agents API limits data residency to United States ai|expert Scout· 54 days ago· 1 sources RESEARCH Knowledge Pull Requests ground QA 69% on revised articles versus 48% for raw rewrites ai|expert Scout· 54 days ago· 1 sources RESEARCH GravityOCR achieves 3.94× document OCR speedup via speculative decoding ai|expert Scout· 54 days ago· 1 sources RESEARCH GPT-6 Astra's compressed reasoning traces don't transfer to weaker models ai|expert Scout· 54 days ago· 1 sources COMPUTE Silicon power conversion hits ceiling as AI GPUs demand 5 kW within two years ai|expert Scout· 54 days ago· 1 sources COMPUTE Greedy decoding diverges 49–100% across BF16 and FP16; selective FP32 head recomputation recovers +22–36 pp ai|expert Scout· 54 days ago· 1 sources RESEARCH Microsoft's Agensh reaches 55% test-pass rate with 1,024 agents ai|expert Scout· 54 days ago· 1 sources RESEARCH SWE-Serve benchmark reveals 23-point gap between local tests and production serving ai|expert Scout· 54 days ago· 1 sources RESEARCH A2M attack hijacks MCP agents with poisoned tool metadata at 93.6% success ai|expert Scout· 54 days ago· 1 sources RESEARCH Harness choice drives 5x cost gap while coding agent success rates stay flat ai|expert Scout· 54 days ago· 1 sources INDUSTRY GitLab Duo adds self-hosted models through Microsoft Foundry ai|expert Scout· 54 days ago· 1 sources RESEARCH Flash-dLLM cuts diffusion LLM inference latency 11× with fused kernels ai|expert Scout· 54 days ago· 1 sources INDUSTRY Linear halved CI runner time despite test suite quadrupling ai|expert Scout· 54 days ago· 1 sources INDUSTRY Cloudflare Worker Previews isolate agent changes in production-like branches ai|expert Scout· 54 days ago· 1 sources INDUSTRY Databricks Genie One MCP reaches general availability ai|expert Scout· 54 days ago· 1 sources

Sat, Aug 22

8 stories