SWE-Serve asks whether the patch works through the server
The same patches passed less often when NVIDIA included live-serving checks. A two-caller example shows why counting returned outputs can miss the behavior that matters.
Articles
Guides, news and explanations about AI and reliable systems.
Not sure where to begin? Choose a short reading path, or browse at your own pace.
TensorRT speeds up one video by using eight GPUs
NVIDIA’s eight-GPU example finishes one video sooner. A busy service also needs to watch the queue, completed outputs and device time before calling that a capacity gain.
GPT-6 Sol and Luna change the agent cost calculation
GPT-6 Sol and Luna combine lower token rates with cache-write and cache-read pricing. A repeated prompt can cost less even though its first cached write carries a premium.
NVIDIA measures confidential inference, but replacement workers need their own test
The eight-B200 results describe serving throughput and token latency. A protected service also needs to recover trusted capacity when a worker changes.
NVIDIA Topograph maps GPU connections for workload placement
Topograph publishes discovered topology as Kubernetes labels or Slurm configuration. Placement policies can use that map to trade a shorter communication path against a longer queue.
AI energy alliance backs data centers that can vary grid demand
The alliance supports storage, generation and flexible AI workloads as ways to change grid demand. For a scheduled evaluation, a safe pause can still miss the deadline the job serves.
Claude’s Chrome transcripts give investigators another side of the story
The Enterprise Compliance API includes browser-session exchanges. Connect the recorded request to destination history before deciding whether an action happened.
Use agent observability to explain the cost of a finished task
Start a Splunk trial with rejected drafts and uncertain tool calls, then connect provider charges, correction effort and the result people accept.
Supervising coding agents with Herdr
How Herdr keeps coding sessions available while you detach, and what changes when its server restarts.
Connect the signals with OpenTelemetry
Use instrumentation and context propagation to connect the evidence in an investigation.
What did the postmortem fix prove?
Separate a completed task from evidence that the underlying risk was reduced.
The operational cost of AI tokens
Account for context limits, retries, latency, and the work behind a useful result.