Try
Useful tools & techniquesN-gram hints cut NanoGPT training to 6.7 seconds
Herman Brunborg reports reducing the NanoGPT Speedrun record from 39.9 to 6.7 seconds by adding retrieval-augmented n-gram hints from training data. The method lets the model learn how much to trust repeated text; the CPU builds a 37-billion-token hint corpus alongside GPU training. The author says AI wrote most of the code but struggled with the major algorithmic ideas.
Useful systems lesson: exploit idle hardware and tailor data structures to the workload; reproduce results before relying on them.
LiveNerf begins measuring post-launch Claude drift
The open-source LiveNerf project tracks Claude Opus 5.5 through Claude Code using a pinned CLI, frozen tasks, exact graders, and preregistered statistics. At day 36, its first two 10-day comparison windows showed no change; the third was underway. Its author says the setup can detect about a 7.5-point accuracy shift per 10-day window, but may miss smaller model swaps.
Use its design to build a controlled, repeatable regression check for your own model workflow.
Know
What changedKimi K3.1 Pro cuts price, lifts agent scores
Artificial Analysis reports Kimi K3.1 Pro scored 55 on its Intelligence Index, with 1M-token context and text/image input; weights were not yet published. Listed API prices are $2/$10 per million input/output tokens and $0.20 per million cache-hit tokens, 33% below Kimi K3. The vendor evaluator reports 67% on Terminal-Bench 4.0 and $1.98 per task. A separate 14-model comparison found 35 of 41 answers converged on humans favoring belonging over truth.
Evaluate it on your own agent tasks; treat third-party benchmark results as a screening signal, not a guarantee.
Claude Duet opens a US research-preview waitlist
Claude Duet is a proactive cloud-computer agent for US Pro, Max, and Ultra subscribers, currently on a waitlist; Team and Enterprise are excluded. It works through connected apps, can draft replies and prepare documents, and asks approval before sensitive actions. Background checks count against the shared Claude plan limits, and Anthropic warns it can miss updates or misjudge urgency.
If you qualify, test it on low-risk delegated work and review its permissions and usage impact.
Thousands of agents analyze brain-cell data in 32 minutes
OpenAI, PerturbAI, and Scripps Research say thousands of GPT-Rosalind agents analyzed a mouse-brain CRISPR dataset and literature in 32 minutes, producing a reasoning atlas of candidate research leads. In blinded review, experts preferred distributed-agent reports over single-agent reports 75% of the time; at the largest workload, the teams cited over 700 references versus 11. The atlas is available for researchers to explore.
For large research or code-analysis tasks, test parallel, scoped agents against one agent and compare evidence coverage.
OpenAI reorganizes its Model Spec around five values
OpenAI’s updated Model Spec centers on faithfulness to human intent, harm minimization, honesty, human agency, and care. It adds guidance on context, oversight, and side effects, revises behavioral defaults, and retires much detailed style guidance. The company says the Spec describes intended behavior, not behavior its models fully meet today.
Review the updated behavior expectations when designing prompts, safeguards, and user controls for OpenAI-powered products.