HackerNews Digest

The Federal Trade Commission and attorneys general from Arizona, Illinois, Michigan, Minnesota and Wisconsin secured a right‑to‑repair settlement with John Deere (Deere & Co.) that obligates the company to provide diagnostic and repair tools to equipment owners and independent repair shops, and bars its dealers from retaliating against those who choose non‑dealer service. The order, filed in Illinois and pending Judge Iain D. Johnston’s approval, also requires Deere to pay $1 million in antitrust enforcement costs to the five states and subjects the company to strict compliance oversight for ten years. This settlement follows an earlier $99 million class‑action settlement with farmers in April 2024. The FTC’s antitrust suit, brought in January 2025, alleged that Deere restricted access to service‑software, forcing customers to use authorized dealers for tractors, forestry, landscaping and construction equipment. Deere’s vice president of aftermarket and customer support said the agreement expands flexible repair options, while Arizona AG Kris Mayes highlighted the reduction of dealer monopoly over repairs.

Chatto, a group‑chat application, has been released as open‑source software and can be self‑hosted. Install via Homebrew (`brew install chattocorp/tap/chatto`) or download binaries for Linux (x86_64/ARM64), macOS, and Windows. The server runs as a single executable, serves its own frontend, and requires minimal resources. All personal and chat data are encrypted at rest with per‑user keys that are deleted upon account removal; no federation or third‑party analytics occur. Clients can connect to multiple independent servers simultaneously. Built‑in voice/video calls and screen‑sharing are end‑to‑end encrypted and scale with the host’s capacity. Chatto Cloud, entering public beta soon, offers paid hosting on European‑owned infrastructure with automatic scaling, nightly backups, and zero‑downtime upgrades; hosted servers remain fully compatible with self‑hosted deployments. Current version is 0.4, considered production‑ready. Roadmap targets 0.5 (content reporting, moderation, multi‑server UI improvements) and a 1.0 release within 6–12 months, with possible breaking changes. Community support and a low‑volume newsletter are provided for updates and feedback.

None

No content was provided to summarize.

Remote attestation uses a TPM to provide cryptographic proof of a host’s boot state. Measured boot records hashes of hardware, firmware, kernel, init image, and optionally the root filesystem into Platform Configuration Registers (PCRs), which can only extend forward. The TPM’s Endorsement Key (EK) is manufacturer‑signed; an Attestation Key (AK) derived from the EK signs TPM quotes of PCR values, proving the host booted as expected. A locally scoped Device ID (LDevID), a child of the AK, can sign arbitrary data but is typically sealed to known “golden” PCR values to prevent use after a compromised boot. TPM sealing ties encrypted material (e.g., root‑fs keys) to specific PCR states, so a host cannot access its data unless the boot measurements match. Remote attesters verify quotes and issue mTLS certificates only to attested hosts, effectively gating workloads. Upgrade handling requires TPM2_PolicyAuthorize with an AuthKey to allow new PCR values. Attestation secures the boot foundation but does not replace runtime protections such as EDR or LSM policies.

Robostral Navigate is an 8‑billion‑parameter, in‑house model for embodied robot navigation that operates from a single RGB camera without depth or LiDAR inputs. Trained entirely in simulation on ~400 k trajectories across 6 k scenes, it achieves 76.6 % success on the R2R‑CE (Room‑to‑Room Continuous Environments) unseen validation benchmark—exceeding the best single‑camera baseline by 9.7 points and surpassing the top multi‑sensor systems by 4.5 points. The model predicts navigation actions via “pointing” (image‑coordinate targets with desired orientation), falling back to local displacement commands when the target lies outside the field of view. Training uses a prefix‑caching, tree‑based attention mask that compresses full episodes into single sequences, reducing token count by 22× and shortening training from months to days. After supervised learning, online reinforcement learning with the CISPO algorithm adds 3.2 % improvement by mitigating behavior‑cloning distribution shift. Robostral Navigate generalizes across wheeled, legged, and flying robots, tolerates varying camera intrinsics, and is positioned as a foundational capability for broader embodied AI applications.

The author spends several hours daily interacting with large language models (LLMs) for both work and personal projects, using Claude Code, Codex, ChatGPT, Gemini, and Qwen. Their workflow has shifted from writing code to designing code, prompting LLMs for implementations, reviewing generated code, and finalizing it manually. They also develop a framework for large‑scale unsupervised code generation and frequently review outputs from autonomous agents. While LLMs increase productivity and expose the author to unfamiliar approaches, repeated exposure to characteristic errors—hallucinations, false assumptions, staccato phrasing, excessive emojis, and uniform stylistic quirks—has caused growing fatigue and dread when reading AI‑generated text. The author notes that these repetitive issues, unlike occasional human errors, become wearing due to their consistency across different models and sources. They acknowledge the usefulness of personalization features but lack control over content produced by others, and they are currently uncertain how to mitigate the emerging burnout.

MechCommander 1 places all “large” weapons in the same mech location (identified as the left arm, location 4), causing a severe loss when that arm is destroyed. The game classifies weapons as large via `getWeaponLarge()`, which lists specific weapon IDs; most Desperate Measures weapons are omitted, defaulting them to “small”. Weapon distribution code checks locations in different orders for small (4→5→2→3) and large (2→3→4→5) weapons. For large weapons the fallback index is 0, forcing every large weapon into location 4. The fix changes the fallback value for large weapons from 0 to 2, aligning its algorithm with the small‑weapon placement order. This is implemented by inserting a code cave at address 0xEB601 that sets EDI = 2 before returning. Additional patches re‑classify Desperate Measures weapons as large by modifying the `getWeaponLarge` routine (addresses 0xEB5B0 and optionally 0xEB1AD). Hex edits: replace bytes at 0xEB48C, 0xEB601, and 0xEB5B0 as described. The patches have been tested to distribute large weapons across both arms and torso slots, eliminating the left‑arm‑only loss. Use at own risk.

No content was provided beyond the title, so a summary cannot be generated.

Unicode UTS #35 transliteration rules, as implemented in ICU, support unbounded rewriting with optional contexts, captures, quantifiers, and a cursor that can reposition after each replacement. By compiling a universal 2‑tag system into such rules, the author shows that a single rule file can simulate any Turing‑complete computation; the example encodes Liesbeth De Mol’s 2‑tag system for the Collatz function, demonstrating step‑by‑step transformation of unary strings. The construction guarantees at most one matching rule per step, halting exactly when the underlying tag system halts, establishing undecidability of termination for arbitrary rule sets. ICU imposes a pragmatic guard of 16 rewrites per input code point, though the specification itself imposes no limit. Beyond tag systems, the same mechanism implements Rule 110 (14 rules) and Wolfram’s prime‑generating cellular automaton (223 rules). Consequently, transliteration rule files are effectively programs, and accepting external rule data entails executing code, necessitating review and runtime bounds. Supporting files (collatz.txt, rule110.txt, primes.txt, uts35.py/.sh) illustrate these capabilities.