From a senior to future senior Hackers

Testing Inkling 2 Mini on a 32GB Mac: Verified JavaScript Fixes, Peak Memory, and Swap
Featured

Testing Inkling 2 Mini on a 32GB Mac: Verified JavaScript Fixes, Peak Memory, and Swap

A hands-on deep technical walkthrough testing Thinking Machines' Inkling 2 Mini on a 32GB Mac, measuring peak memory and swap while evaluating tool-call reliability, latency, cost per verified JavaScript fix, and the impact of its looped transformer design.

pr0h0•
local-llmai-coding-agentsjavascriptmemory-profilinginkling-2-mini

Latest Posts

View all posts →

Latest Videos

View all videos →
An Out-of-Band Incident Channel for Microsoft 365 and N-able Teams

An Out-of-Band Incident Channel for Microsoft 365 and N-able Teams

A second chat app isn't out of band if its login, recovery, or alerts still depend on the failed stack. This walkthrough shows a hidden identity dependency, an illustrative JavaScript check for transitive dependencies, and a phone-tree path that avoids corporate identity. It also covers the response requirements and why the reported email-security scenario calls for a tabletop rather than an invented technical simulation. The supplied secondary reports describe a Microsoft 365 outage and a security incident affecting N-able cloud email security services. They do not establish root causes or complete incident timelines, and they do not support extending the N-able scope to other products. Full write-up: https://hackyjs.com/posts/an-out-of-band-incident-channel-for-teams-dependent-on-microsoft-365-and-n-able

incident-responseresiliencesaas-dependencies
AbortSignal.any: Cancel fetch on Timeout or User Action

AbortSignal.any: Cancel fetch on Timeout or User Action

Promise.race can reject without stopping fetch. Compose a timeout signal and a fresh cancellation controller with AbortSignal.any, then pass the composite to fetch. The same signal also protects the response body read. The first abort wins and preserves its reason. Check HTTP status separately, and remember that client cancellation does not roll back server work. Full write-up and runnable example: https://hackyjs.com/posts/using-abortsignalany-to-compose-timeout-and-cancellation-signals-in-fetch

javascriptfetchabortsignal
Testing Gemini 4 Neon's Long-Context Reliability Past 200K Tokens

Testing Gemini 4 Neon's Long-Context Reliability Past 200K Tokens

Gemini 4 Neon's launch coverage claims a 1M-token context window and a #1 Agent Index placement. Neither answers the question that decides whether you can ship on it: at what depth inside that window does the model still return the right fact, in the right order, without inventing one? This short walks the hands-on reliability test behind the full post: four mechanically scored tasks — needle retrieval with distractors, chained multi-hop synthesis, codebase-scale retrieval, and an agentic edit loop — run at 200K, 500K and 900K reported prompt tokens, five runs per cell, with forty questions against a real open-source JavaScript repository. No LLM-as-judge anywhere; every score is a string comparison or a test-suite exit code. What the write-up covers: - Why an agent benchmark rank is not a long-context retrieval measurement - Advertised context versus effective context, and why the gap widens with depth - Positional encoding extrapolation, attention dilution, and prefill cost as a reliability property - Lost in the middle, distractor substitution, and chained-hop drift - Token accounting from the provider's usage metadata instead of a local estimate - Why correct abstention on absent facts decides the incident - The operational position: treat 200K as the contract, and anything past it as an experiment you re-validate per workload Every number quoted here comes from that harness. The source material states that pricing exists but not the figures, so none are given. Full write-up: https://hackyjs.com/posts/testing-gemini-4-neons-long-context-reliability-past-200k-tokens

gemini-4-neonlong-contextllm-evaluation