From a senior to future senior Hackers

2.4T Parameters, 60% Cheaper Per Task: Auditing Qwen4-Max After Retries
Featured

2.4T Parameters, 60% Cheaper Per Task: Auditing Qwen4-Max After Retries

An audit of Qwen4-Max's 2.4T-parameter MoE and its claimed 60% per-task cost cut, using a retry-adjusted cost-per-successful-task worksheet to test whether the vendor's economics survive eval-driven model routing.

pr0h0•
llm-economicsmoeqwenmodel-routinginference-cost

Latest Posts

View all posts →

Latest Videos

View all videos →
Testing Email, Teams, and Copilot Fallbacks Against an M365 Control-Plane Outage

Testing Email, Teams, and Copilot Fallbacks Against an M365 Control-Plane Outage

On 15 November 2026 a reported global Microsoft 365 outage affected Outlook, Teams and Copilot. Microsoft linked it to a network upgrade issue and said it would publish a full review. The engineering problem isn't the outage itself — it's that most fallbacks turned out to share an identity provider, a DNS resolver and a tenant policy boundary with the primary path. This short walks through the testable model: - Control plane versus data plane, and why the fault pattern is warm sessions surviving while every cold start fails. Existing sessions buy you roughly one access-token lifetime, not a day. - Email, Teams and Copilot as separate dependency chains. Inbound mail often keeps landing while users cannot read it, and Copilot needs identity, a licence, Graph, the semantic index, permission trimming and tenant policy before it emits a token. - A lab-scoped harness in a disposable tenant with synthetic users: curl's --resolve flag pins one hostname to a blackhole address for a single request, modelling a DNS answer that resolves but never connects, without touching the machine's resolver. - A baseline probe captured before the drill, so you can tell a degraded probe from one that was always slow. Everything here assumes a tenant you own, users you created, and faults injected on your side of the wire. Do not fault-inject against Microsoft endpoints from a production network. Reporting summarised here is headline-level. Root cause, exact duration in UTC, regional scope beyond the reported US and Japan impact, and the causal status of the reported hotel warning-system failure all remain unconfirmed until Microsoft publishes its review. Full write-up: https://hackyjs.com/posts/testing-email-teams-and-copilot-fallbacks-against-an-m365-control-plane-outage

microsoft-365outageresilience
Chrome 157.0.8103.50: What CVE-2026-125350 Patched and How to Confirm It's Running

Chrome 157.0.8103.50: What CVE-2026-125350 Patched and How to Confirm It's Running

Chrome 157.0.8103.50 carries the fix for CVE-2026-125350, a flaw that reporting says was being exploited before that build shipped. This short covers the part that actually decides your exposure: whether the fixed build is the one executing. What the public record establishes: the CVE is tracked as CVE-2026-125350, the fixed version is 157.0.8103.50, and a CERT-In warning described the flaw as actively exploited. The vulnerability class, the affected component and the full vulnerable range are not stated in the available sources — this video does not guess at them, because a wrong bug-class assumption leads to the wrong interim mitigation and the wrong detections. What the video shows: - Why "Chrome is up to date" is a statement about the update check, not about the process in front of you - How the updater stages new bits on disk while the running process keeps the old code in memory - Reading the running build from Chrome's internal version page, plus the executable path that reveals which install answered - Per-platform checks for the installed build: Google Update's registry key on Windows, bundle metadata on macOS - The exposure gap between an installed version and a running version, and why per-user and side-by-side installs break your asset count All content is defensive and lab-scoped: no exploit code, no target, no credentials. The commands shown only read version information from a machine you administer. Full write-up with the per-platform commands and output: https://hackyjs.com/posts/what-the-chrome-1570810350-update-fixes-in-cve-2026-125350-and-how-to-confirm-it-is-applied

chromezero-daycve-2026-125350
Auditing MCP Servers and Clients for Hidden Session State Before 2026-07-28

Auditing MCP Servers and Clients for Hidden Session State Before 2026-07-28

A short audit of the implicit session state in MCP servers and clients: the initialize handshake, Mcp-Session-Id, SSE replay buffers, server-initiated requests, and the caches and sticky routing teams add themselves. The HackyJS post labels what is confirmed from the 2026-07-28 commentary, what is inferred, and points to the normative text before migration planning. Full write-up: https://hackyjs.com/posts/auditing-your-mcp-server-and-client-for-hidden-session-state-before-the-2026-07-28-stateless-revision

mcpstateless-protocolssession-management