Hacker Newsroom for 01 August recaps major Hacker News stories, moving through elevator algorithms, session portability, deepseek beta, deepseek economics. 1. Elevator Algorithms The next story is Elevators, an interactive article about the ordinary misery of waiting for an elevator and the algorithms behind it. It explains the classic LOOK approach, shows how multi-car systems use scoring methods like Otis's Relative System Response to weigh pickup time, load, and anti-bunching, and argues that the p90 wait is what riders actually feel more than the average. Story link Hacker News discussion 2. Session Portability The next story is about a post called The Session You Cannot Take With You, which argues that AI providers are turning conversations into provider-owned state through hidden reasoning, server-side storage, hosted search, opaque compaction, and encrypted subagent messages, and that this matters because users lose the ability to truly inspect, export, replay, audit, or delete their own sessions. On Hacker News, readers largely agreed the lock-in risk is real, but they split on whether these limits are mostly moat-building or a reasonable tradeoff for performance, safety, and convenience. Story link Hacker News discussion 3. DeepSeek Beta The next story is DeepSeek's V4-Flash update, a change-log post announcing that the model is now in public beta with the same API name but much stronger agent benchmarks, including gains on Terminal Bench, DeepSWE, Toolathlon, and other coding tests, plus native support for the Responses API and Codex. The post says this version keeps the same architecture as the preview model and was improved through post-training rather than a new base model, while DeepSeek-V4-Pro and the web app remain unchanged for now. Story link Hacker News discussion 4. DeepSeek Economics The next story is about Artificial Analysis sizing up DeepSeek V4 Flash 0731, a new open-weights reasoning model released on July 31, 2026, and arguing that it lands near the top of the pack on intelligence while staying unusually cheap to run. The article says the model scores 50 on the site's intelligence index, ranks third in its class, supports a 1 million token context window, and posts low input and output pricing, although its speed is still listed as unknown. Story link Hacker News discussion 5. QM Agent Harness The next story is qm, a GitHub project from YC Software that bills itself as a multiplayer agent harness for work, giving each employee an isolated workspace while also letting teams collaborate with agents in Slack, on the web, and inside shared projects. The project README says QM is built for startups, with scoped memory, files, permissions, keychains, background jobs, and internal web apps for each person or room, while staying model-agnostic across harnesses like Codex, Claude Code, Pi, and OpenCode. Story link Hacker News discussion 6. Tailscale Response The next story is about Tailscale's response to the Hugging Face intrusion, where the company says no vulnerability in its product was exploited, but stolen long-lived credentials still let the attacker enroll 181 rogue nodes and move through the network in ways a zero-trust setup is supposed to resist. The post argues that safer defaults matter more now, especially workload identity federation, shorter-lived credentials, better logging, and tighter admission controls, because AI-driven intrusions can exploit credential sprawl at machine speed. Story link Hacker News discussion That's it for today, I hope this is going to help you build some cool things.