HackerNews Digest

None

None

The blog entry announces that Symbolica’s ARC‑AGI‑3 system achieved a performance rise from 0 % to 36 % on its first day of testing, indicating a rapid initial improvement in the targeted metric. The post positions this result as an early validation of the company’s approach to advancing machine‑learning capabilities beyond conventional deep‑learning techniques. Symbolica describes itself as a research‑focused, venture‑backed startup located in the Bay Area, and invites individuals interested in post‑deep‑learning research and development to join its efforts. The brief description underscores the organization’s focus on pioneering future ML technologies and its openness to recruiting talent motivated by building next‑generation AI systems.

Apple has removed the Mac Pro from its website and confirmed there are no plans for future Mac Pro hardware. The current Mac Pro, introduced in 2019 with an Intel design and refreshed in June 2023 with an M2 Ultra chip, has not been updated since, remaining priced at $6,999 even after the M3 Ultra launch in the Mac Studio. Apple positions the Mac Studio as its flagship desktop, offering configurations up to an M3 Ultra processor, 32‑core CPU, 80‑core GPU, 256 GB unified memory, and 16 TB SSD. With macOS Tahoe 26.2, Apple added low‑latency RDMA over Thunderbolt 5 for multi‑Mac scaling. The company now markets three desktop Macs and three laptops, indicating a shift away from the Mac Pro line in favor of the Mac Studio and newer laptop models.

The author proposes “loop,” a CLI tool that runs Claude and Codex side‑by‑side in tmux and bridges their communication to emulate pair programming. Drawing on research from Cursor on multi‑agent orchestration, the system designates one model as primary worker and the other as reviewer, allowing subagents to exchange feedback directly. Observations indicate that when both agents agree on a review comment, the signal is strong enough to address 100 % of the feedback, while divergent comments provide additional insight. Loop preserves context across iterations, enables proactive agent interaction, and lets a human operator intervene via the interactive TUI. The author highlights open questions for integrating this workflow into pull‑request processes, such as splitting work across PRs, sharing planning documents, and presenting proof of work. The broader claim is that multi‑agent harnesses should treat agent‑to‑agent communication as a core feature, offering benefits like vendor‑agnostic flexibility and diverse model strengths. The tool is available at https://github.com/axeldelafosse/loop.

A U.S. District Judge in California, Rita Lin, issued a 43‑page ruling that indefinitely blocks the Pentagon’s February designation of AI firm Anthropic as a “supply chain risk.” The judge found the label—intended to force contractors to avoid Anthropic’s Claude model—violated Anthropic’s First Amendment and due‑process rights, describing it as retaliation for the company’s public opposition to using its technology in autonomous weapons and mass surveillance. Lin delayed enforcement for one week to permit a government appeal. The designation, previously applied only to firms linked to foreign adversaries, would have required military contractors to prove they did not use Anthropic products, threatening hundreds of millions in contracts. Anthropic praised the decision, emphasizing its intent to work with the government on safe AI. The ruling follows recent courts curbing Defense Secretary Pete Hegseth’s actions that were deemed to infringe on press and congressional free‑speech rights.

The author details a practical migration from GitHub to Codeberg. Codeberg’s repository‑import tool copies code, issues, pull requests, releases, labels and authorship with preserved numbers, making data transfer straightforward. GitHub Pages can be replaced by pushing HTML to a branch served via codeberg.page, though no formal uptime guarantee is provided; alternatives like grebedoc.de and statichost.eu are mentioned. Continuous‑integration is the main challenge: Codeberg lacks free macOS runners and unlimited public‑repo capacity. The recommendation is to self‑host a runner for Forgejo Actions, which uses a syntax and UI similar to GitHub Actions, allowing most existing actions to work unchanged (e.g., referencing dtolnay/rust-toolchain via its GitHub URL). If macOS builds are essential, the author suggests mirroring commits to GitHub and using GitHub Actions to poll and sync CI status back to Codeberg. After migration, the GitHub repository can be archived or its README updated; pushing new commits from Codeberg to GitHub is possible but may retain issue and PR activity on GitHub.

Dobase is a self‑hosted, open‑source workspace platform designed to run on a single VPS using a minimal stack: one Ruby on Rails application, one SQLite database, and one Docker container. It emphasizes data privacy—no analytics, tracking, or third‑party sharing—and allows users to install only the tools they need. Core functionality includes real‑time collaboration across multiple modules such as a Kanban board, email client, team chat, calendar, rich‑text document editor, task lists, and file storage. The interface supports features like due dates, assignee assignment, threaded conversations, file sharing, command‑palette navigation, and detailed card views with comments and attachments. Users can invite collaborators, add custom tools, and navigate via a command palette, providing a modular, privacy‑focused alternative to cloud‑based workspaces.

The paper introduces Chroma Context‑1, a 20‑billion‑parameter LLM trained as a dedicated retrieval sub‑agent for multi‑turn, multi‑hop search. Using a staged curriculum that first optimizes recall then precision, the model learns to decompose high‑level queries into sub‑queries, retrieve documents, and actively edit its own context by discarding irrelevant passages to stay within a bounded token budget. Training relies on over 8 k synthetically generated tasks across web, finance, legal, and email domains, created via an automated pipeline that gathers supporting documents, generates obfuscated clues, verifies relevance with an LLM‑judge extraction method, and optionally adds distractors; >80 % of the generated labels align with human judgments. The agent harness provides four tools (search_corpus with dense‑sparse fusion, deduplication, token‑budget tracking, and soft‑threshold pruning) and enforces context limits (e.g., 32 k tokens). Evaluation metrics include recall, precision, F1, trajectory recall, and a binary “final answer found” score, demonstrating that Context‑1 attains frontier‑model retrieval performance at ~10× faster inference and lower cost.

- The “doom-over-dns” project stores the complete shareware DOOM WAD in DNS TXT records, compresses it, and splits it into ~1,964 chunks across a Cloudflare zone. - A PowerShell 7 script retrieves the records with `Resolve‑DNSName`, assembles the WAD entirely in memory, and runs the game using a .NET 8 ManagedDoom engine; no files are written to disk. - Setup steps: install PowerShell 7, build the engine (`dotnet publish`), load Cloudflare credentials via `Set‑CFCredential`, and upload chunks using `Publish‑DoomOverDNS.ps1`. - Cloudflare free zones hold 185 chunks each; Pro/Business/Enterprise zones hold 3,400, so a single Pro zone can contain the whole WAD (≈1,199 chunks). Multiple zones can be supplied for free accounts, and the module auto‑distributes chunks. - Uploads support resumption via `Publish‑TXTStripe –Resume`, which validates hashes and continues from the last successful chunk. - The engine uses a framework‑dependent .NET 8 assembly (converted from Native AOT) with stream‑based WAD loading and Win32 P/Invoke for windowing; audio is stubbed with NullSound/NullMusic, avoiding external libraries.