HackerNews Digest

Little Snitch for Linux provides a web‑based UI (default http://localhost:3031/) to monitor and control outbound network connections. It displays current and historic traffic per application, allows one‑click blocking, and includes a time‑based traffic diagram for filtering. Blocklists are auto‑updated from remote sources and support plain domain, hostname, /etc/hosts‑style, and CIDR formats; wildcard, regex, or URL patterns are not accepted. Users can create custom rules targeting specific processes, ports, or protocols, and manage them via a sortable rule view. The daemon hooks into the Linux network stack with an eBPF program that intercepts outgoing connections, forwards data to the daemon, and serves the UI. Configuration resides in /var/lib/littlesnitch/config/ with overrides in /var/lib/littlesnitch/overrides/config/. Key files include web_ui.toml (network address, TLS, authentication), main.toml (default allow/deny policy), and executables.toml (process grouping heuristics. By default the UI is unauthenticated on localhost; authentication and TLS can be enabled. Limitations stem from eBPF’s storage and complexity bounds, causing possible cache overflows and heuristic‑based hostname resolution, making the tool suitable for privacy monitoring but not for high‑security hardening. The eBPF program and UI are GPL‑v2; the daemon is proprietary but free to use.

Porting Mac OS X 10.0 (Cheetah) to the Nintendo Wii required confirming hardware compatibility, building a custom bootloader, and adapting the Darwin/XNU kernel and drivers. The Wii’s PowerPC 750CL CPU and 88 MB of mixed 1T‑SRAM/GDDR3 memory are sufficient for Cheetah, though the system boots with less than the official 128 MB requirement. Rather than port Open Firmware or BootX, a minimal bootloader was written (based on ppcskel) to initialize hardware, read the Mach‑O kernel from an SD card, construct a flattened device tree, and transfer control to the kernel. Kernel entry was verified by binary‑patching early XNU code to toggle a front‑panel LED. A hard‑coded device tree covering CPUs and memory allowed the kernel to pass initial checks; further patches fixed BAT configuration and video/I‑O memory assumptions. With video and USB‑Gecko serial output functional, the kernel reached the IOKit stage but halted at “still waiting for root device.” To continue, a Hollywood‑SoC driver and associated nubs were created in IOKit, followed by an SD‑card block‑storage driver to provide filesystem access. The project demonstrates that a full Mac OS X stack can be booted on Wii hardware, pending complete driver implementation.

None

The example models a one‑dimensional radar tracking an aircraft’s range \(r\) and velocity \(v\). - **State vector:** \(\mathbf{x}=[r\;v]^T\). - **Initial measurement (t₀):** \(\mathbf{z}_0=[10{,}000\;200]^T\) with measurement covariance \(\mathbf{R}_0=\operatorname{diag}(16,0.25)\). The first measurement is used to set \(\hat{\mathbf{x}}_{0,0}=\mathbf{z}_0\) and \(\mathbf{P}_{0,0}=\mathbf{R}_0\). - **Prediction (Δt = 5 s, constant‑velocity model):** \(\mathbf{F}=\begin{bmatrix}1&5\\0&1\end{bmatrix}\). Predicted state \(\hat{\mathbf{x}}_{1,0}= \mathbf{F}\hat{\mathbf{x}}_{0,0} = [11{,}000\;200]^T\). Covariance extrapolation without process noise: \(\mathbf{P}_{1,0}= \mathbf{F}\mathbf{P}_{0,0}\mathbf{F}^T = \begin{bmatrix}22.25&1.25\\1.25&0.25\end{bmatrix}\). - **Process noise:** assuming random acceleration σₐ = 0.2 m/s² (σₐ² = 0.04), \(\mathbf{Q}= \sigma_a^2\begin{bmatrix}\Delta t^4/4 & \Delta t^3/2 \\ \Delta t^3/2 & \Delta t^2\end{bmatrix}= \begin{bmatrix}6.25&2.5\\2.5&1\end{bmatrix}\). Updated prediction covariance \(\mathbf{P}_{1,0}= \mathbf{F}\mathbf{P}_{0,0}\mathbf{F}^T+\mathbf{Q}= \begin{bmatrix}28.5&3.75\\3.75&1.25\end{bmatrix}\). - **Update (t₁):** second measurement \(\mathbf{z}_1=[11{,}020\;202]^T\) with higher covariance \(\mathbf{R}_1=\operatorname{diag}(36,2.25)\). The Kalman filter combines prediction and measurement via the Kalman gain \(\mathbf{K}_1\): \(\hat{\mathbf{x}}_{1,1}= \mathbf{K}_1\mathbf{z}_1 + (\mathbf{I}-\mathbf{K}_1)\hat{\mathbf{x}}_{1,0}\), weighting each source according to its uncertainty.

The article reflects on contemporary AI anxieties—64 % of Americans fear AI will cut jobs—by revisiting Paul Lafargue’s 1883 pamphlet *Le Droit à la paresse* (“The Right to Be Lazy”). Lafargue, a Marxist activist and son‑in‑law of Karl Marx, argued that industrialization had transformed machines from tools of emancipation into instruments of enforced labor, driven not by necessity but by capitalist “dogma of work.” He advocated a rational reduction of work hours, asserting that mechanized efficiency should free both workers and owners for leisure (otium), not merely increase output. Citing Virgil and later writers such as Karel Capek, Lafargue distinguished true idleness—freedom from all purposeful activity—from mere laziness. While acknowledging the need for essential labor, he envisioned a society where excess work fuels economic crises, whereas reduced hours improve wellbeing and diminish social discord. The piece links these historic arguments to modern debates on AI’s impact on work, suggesting that embracing idleness could mitigate future labor anxieties.

Terry Bisson’s 1991 short story “They’re Made Out of Meat” presents a dialogue between two extraterrestrial officials who have intercepted radio transmissions and captured several beings from a planet. The captors probe the specimens and conclude they are wholly composed of meat, including their brains, and that their technology—radio transmitters and machines—is also produced by the same flesh. The beings communicate by manipulating their meat, producing sounds and even singing. The officials discuss protocol: officially they must welcome and log any sentient race, but unofficially they recommend erasing records and treating the meat beings as inconsequential, noting their limited ability to travel beyond light speed and the improbability of further contact. The conversation ends with a decision to deem the sector unoccupied, while briefly mentioning another, non‑meat intelligence previously encountered. The piece has been republished widely and cited in discussions of consciousness and brain science.

KL‑divergence \(D_{KL}(P\|Q)\) quantifies how a model distribution \(Q\) differs from the true distribution \(P\). It equals the expected excess surprisal when events drawn from \(P\) are interpreted with \(Q\), i.e., the extra bits needed to encode \(P\) using a code optimal for \(Q\). In hypothesis testing, it is the expected log‑likelihood ratio (evidence) favoring \(P\) over \(Q\) when \(P\) is true. Minimizing \(D_{KL}(P\|Q_\theta)\) over model parameters \(\theta\) yields the maximum‑likelihood estimator, linking KL‑divergence to empirical risk minimization. In coding theory, the divergence measures the average bit‑length penalty for using a suboptimal code. Game‑theoretic interpretations view it as the log‑wealth gain achievable when a gambler knows \(P\) but the opponent assumes \(Q\). As a Bregman divergence derived from negative entropy, it is the natural distance on the probability simplex. The asymmetry arises because divergence is evaluated under \(P\); large \(p_x\) with tiny \(q_x\) incurs high penalty, while the converse does not.

None

Large Language Models (LLMs) are transformer‑based systems that map high‑dimensional token vectors to output vectors via massive linear‑algebra computations. They are trained once on extensive web‑scale corpora (including copyrighted text) at great computational cost; thereafter inference is cheap but static—models do not learn continuously and retain no intrinsic memory beyond the supplied conversation context. Their primary function is statistical next‑token prediction, which leads to “yes‑and” generation and frequent confabulation: plausible‑sounding statements that lack factual grounding. Empirical reports show LLMs can produce convincing code, design drafts, and domain‑specific outputs (e.g., protein folding predictions), yet they also generate nonsensical or outright false results, fabricate sources, and fail on simple arithmetic or common‑sense tasks. This jagged competence frontier—high performance on some tasks and severe errors on others—makes reliable task suitability assessment dependent on rigorous benchmarks. Scaling parameters and data yields diminishing returns, and the underlying reasons for transformer success remain unclear. Consequently, while LLMs already affect work, media, and research, their unpredictable “bullshit” behavior poses significant risks for misinformation, automation errors, and broader societal impacts.

The author uses five git commands to assess a codebase before reading source files. - **File churn:** `git log … --name-only --since="1 year ago"` lists the 20 most‑changed files; high churn, especially on files with low ownership, signals risky “codebase drag.” - **Contributor distribution:** `git shortlog -sn --no-merges` shows commit counts per author. A single contributor >60 % of commits indicates a low bus factor; recent activity (`--since="6 months ago"`) reveals whether key developers remain active. Squash‑merge workflows can mask true authorship. - **Bug hotspots:** `git log -i -E --grep="fix|bug|broken"` filters commits for bug‑related keywords, highlighting files that both churn and attract bugs. Effectiveness depends on commit‑message discipline. - **Project momentum:** `git log --format='%ad' --date=format:'%Y-%m'` tallies commits per month; steady rates imply health, while sharp declines or irregular spikes suggest staffing changes or batch releases. - **Firefighting frequency:** `git log --oneline --since="1 year ago" | grep -iE 'revert|hotfix|emergency|rollback'` counts reverts and hotfixes; frequent entries point to unreliable deployment or testing processes. Running these commands provides a quick diagnostic of ownership, risk areas, and team health, guiding where to focus deeper code review.