the log · 159 entries
Log-scale notes on infrastructure.
Analysis on backup, cloud, open source, security and the companies shaping them. One entry a day, filtered through twenty-five years of running the systems involved.
all 159
ai-machine-learning 16
ai-engineering 78
infrastructure 119
machine-learning 3
programming 64
software-development 40
technology 113
technology-leadership 29
tutorials 1
2026.08.177 min read
Data Center Operators Pursue IPOs as AI Demand Pushes Valuations Higher
Vantage, Switch, CyrusOne and DayOne signaled public listing plans in one week. AI demand is lifting their expected valuations, but power constraints, operating costs and investor caution…
2026.08.167 min read
AI API Prices Are Falling Fast. Model Routing Is Becoming Core Infrastructure.
Rapid AI price cuts make a fixed model choice expensive to maintain. Task-aware routing offers a way to reduce costs while keeping harder work on more capable models.
2026.08.156 min read
Qwen 3.8-27B Gives Self-Hosted Agents a Stronger Case
Alibaba's Qwen 3.8-27B combines Apache 2.0 licensing, single-GPU deployment and stronger reported agent benchmarks. The next test is how those gains hold up in local workloads.
2026.08.147 min read
Cerebras Reports 750 Tokens Per Second for GPT-5.6 Sol Ultrafast
Cerebras and OpenAI's Ultrafast tier promises up to 750 output tokens per second. On-chip memory helps explain the speed, though production performance and capacity remain open questions.
2026.08.137 min read
IBM's $240 Million Together AI Deal Puts Open-Model Inference Costs in Focus
IBM and Together AI plan a 2,000-chip Blackwell inference cluster for Q1 2027. The deal gives enterprises another option to compare on price, model quality, and operating requirements.
2026.08.127 min read
Mojo 1.0 Adds a Stability Commitment, but Replacing CUDA Still Takes Work
Mojo 1.0 brings a stability commitment and an Apache 2.0 standard library. Competitive memory-bound benchmarks make it worth evaluating, though performance gaps and a closed compiler still…
2026.08.117 min read
Meta's Muse Glimmer Brings a 30B Agent Model to a Single Consumer GPU
Meta says Muse Glimmer runs on a single 24 GB GPU, with competitive agent benchmarks and Apache 2.0 licensing. The release makes local inference more practical, though performance and…
2026.08.107 min read
ByteDance's Reported 10 Trillion-Parameter Model Faces a Hardware Test
ByteDance is reportedly training a huge AI model on domestic Chinese chips. The harder test is keeping 30,000 accelerators coordinated, supplied with data, and running through months of…
2026.08.097 min read
SCTPhantom: An 18-Year-Old Linux SCTP Flaw Can Allow Container Escape
Tencent researchers report a Linux SCTP use-after-free that can lead to root access and container escape. Exposure depends on SCTP access, with kernel patches and module restrictions…
Nothing on this page matches. Press Enter to search every entry.