High Learning Rate
Articles about High Learning Rate
Codex Gets More Room Before Compaction - 2026-10-12
A larger default context, plus a debugging lesson about testing hypotheses instead of rerunning guesses.
A Close Pass Rate Can Hide a Bad Judge - 2026-10-11
One agent-evaluation experiment worth inspecting, plus a calibration toolkit to watch.
When Agents Cross the Boundary - 2026-10-10
Agent boundaries, decision models, a 3D experiment, and the real cost of longer contexts.
Pretty Interfaces Still Need Tests - 2026-10-09
A generative-UI teardown, Sol's speed premium, opt-in OSS scanning, and Google's persistent work agent.
Check the Grader, Not Just the Agent - 2026-10-07
A verifier audit, multimodal embeddings, inspectable mathematics, and a practical lesson in GitHub polling.
Plans You Can Inspect, Decisions Without Tokens - 2026-10-06
A visual planning skill, image-based decisions, and the limits of text watermarks.
Restarts and Revenue Rankings - 2026-10-05
vLLM's deployment changes, a sales-agent evaluation proposal, and one engineer's shell-formatting failure.
The Delay Before the Model - 2026-10-04
A gateway latency fix, a firsthand software-factory workflow, and a state-constrained agent demo.
Switch Coding Agents Mid-Thread - 2026-10-03
Context handoffs, memory pruning, a non-rendering browser, and practical lessons from production inference.
Finding Bugs Without a Ticket - 2026-10-02
Proactive debugging, crash-safe agents, better evaluations, and making model output easier to understand.
Decisions Without Generated Text - 2026-10-01
Clef's decision models, Claude Code mods, contextual retrieval, and Gemini Argon's staged rollout.
Agents Rewrite Their Context - 2026-09-30
Editable context, pretrained tables, background agents, and cheaper Sol tokens.
Sonnet 5.5, Same Token Price - 2026-09-29
Task cost, subscription allowances, voice latency, and the limits of prompt-only containment.
A Million Hours, Better Audio Filters - 2026-09-28
YODAS v3 gives voice builders more data and metadata to select the right audio.
Kernel Optimization Without Words - 2026-09-27
A PTX-only experiment highlights why valid instructions still need correctness tests.
Small Models, Real Rewards - 2026-09-25
Verifiable RL tasks, constrained decisions, faster vision decoding, and cloud agent sandboxes.
Spend Effort, Guard the Exits - 2026-09-26
Effort experiments, cache-aware routing, and new evidence about agent data transfers.
Agents Need Hard Boundaries - 2026-09-24
Access controls, speaker attribution, scientific validation, and robot policies: four deployment lessons.
Cheaper Models, Migration Traps - 2026-09-23
Sol, Luna and Opus 5.5 change agent economics; security evidence and local inference deserve attention too.
MiMo Opens Up, Tokenization Speeds Up - 2026-09-22
Xiaomi releases MiMo V2.6 weights, and Hugging Face targets the CPU work before inference.
Open Weights, Restricted Use - 2026-09-21
Qwen's license, ZCode's security response, and self-hosted decision models.
Cheap Judges Need Better Tests - 2026-09-20
Cheaper agent judges, cache-aware sessions, a new Step model, and a research corpus worth auditing.
Claude Code Reads AGENTS.md - 2026-09-19
Shared coding-agent rules, local typed decisions, and a speech-model migration worth testing before switching defaults.
Check the Benchmark First - 2026-09-18
Benchmark audits, smaller local models, proof-checked code, and historical search change what builders can verify.
When Agents Rewrite Their Instructions - 2026-09-17
Compaction failures, useful engineering feedback, tool-aware caching, and Claude's shift toward background work.
Voice Agents Keep Working - 2026-09-16
Voice agents, lab-trained models, safer command execution, and a model retirement deadline.
Agent Code Outruns CI - 2026-09-15
CI backlogs, tighter agent permissions, portable coding sessions, and Firebase's delayed spending controls.
FlashREINFORCE and an Open AI Infrastructure Book - 2026-09-14
Critic-free asynchronous RL, plus a practical guide to memory, bandwidth, and distributed training.
Outside Evaluators, Grok in Copilot - 2026-09-13
Anthropic and OpenAI commit to embedded evaluators; Copilot's Grok preview adds a new data-processing decision.
Plugin Evals, Fusion, Astra Fixes - 2026-09-12
Test plugin value, split agent work, revisit Astra failures, and check DSPy's new execution paths.