Published on April 25, 2026 at 18:00 CEST (UTC+2)
1-Bit Hokusai's "The Great Wave" (2023) (219 points by stephen-hill)
1-Bit Hokusai's “The Great Wave” (2023)
The author describes a personal project to recreate all 36 views of Mount Fuji by Hokusai as 1-bit pixel art on vintage Macintosh computers (like a Quadra 700 or PowerBook 100) running System 7 and Aldus SuperPaint. The goal is to capture both the original woodcut aesthetic and the retro Mac pixel look, using the exact 512×342 screen resolution. The project stalled but the author finds joy in the “flow state” and the challenge of authentic constraints. It’s a nostalgic, artistic endeavor rather than a technical innovation.
GPT 5.5 biosafety bounty (29 points by Murfalo)
GPT 5.5 biosafety bounty
OpenAI launched a biosafety bounty program for its GPT-5.5 model, inviting external researchers to identify potential risks related to biological threats. The initiative aims to proactively uncover and mitigate dangerous capabilities before the model is widely deployed. This reflects growing industry focus on AI safety and responsible release practices.
New 10 GbE USB adapters are cooler, smaller, cheaper (390 points by calcifer)
New 10 GbE USB adapters are cooler, smaller, cheaper
Jeff Geerling reviews a new $80 RTL8159-based USB 3.2 10 GbE adapter from WisdPi, comparing it to older, bulkier Thunderbolt adapters. He tests it on several laptops and a desktop, noting that actual performance depends heavily on the host’s USB controller specification. While cheaper and more compact than Thunderbolt alternatives, users may not always achieve full 10 Gbps throughput due to USB protocol limitations.
Martin Galway's music source files from 1980's Commodore 64 games (80 points by ingve)
Martin Galway's music source files from 1980's Commodore 64 games
Legendary C64 composer Martin Galway has open-sourced the assembly source files for his game music (e.g., Wizball, Athena) on GitHub. The repository includes both 1st and 2nd generation player code, allowing others to read, analyze, and remix the music. Galway retains copyright but acquired the rights from Infogrames, encouraging proper attribution for derivative works.
What's Missing in the 'Agentic' Story (3 points by ingve)
What's Missing in the 'Agentic' Story
Mark Nottingham argues that the common narrative around “agentic” AI overlooks fundamental trust and control issues. Historically, local software was assumed to do only what its creators promised—malware was the exception. But as AI agents become pervasive, we lose that assumption; users can no longer be sure what a remote, opaque agent will actually do. This calls for new governance and transparency mechanisms.
Google plans to invest up to $40B in Anthropic (707 points by elffjs)
Google plans to invest up to $40B in Anthropic
Bloomberg reports that Google intends to invest up to $40 billion in Anthropic, the AI safety company behind Claude. This massive financial commitment underscores the fierce competition among tech giants to secure access to frontier AI models and talent. It also signals a bet that safety-focused AI development will be commercially and strategically critical.
Commenting and Approving Pull Requests (34 points by jwworth)
Commenting and Approving Pull Requests
Jake Worth shares a practical PR review workflow: if all comments are non-blocking (nitpicks, suggestions, questions), he approves the PR at the same time. He emphasizes trust in the team and the value of leaving positive or constructive observations. The approach speeds up code review while maintaining quality, assuming CI is fast and approvals are not reset on new commits.
Insights into firewood use by early Middle Pleistocene hominins (11 points by wslh)
Insights into firewood use by early Middle Pleistocene hominins
This scientific paper (published in Quaternary Science Reviews) analyzes archaeological evidence to understand how hominins in the Middle Pleistocene used firewood. It investigates fuel selection, burning patterns, and implications for early human behavior and environmental adaptation. The study contributes to our understanding of hominin technology and survival strategies before the emergence of Homo sapiens.
Lambda Calculus Benchmark for AI (69 points by marvinborner)
Lambda Calculus Benchmark for AI
“LamBench” is a benchmark suite designed to test the reasoning and elegance of AI models using lambda calculus problems. It evaluates AI on speed, elegance, and correctness in solving functional programming challenges. The goal is to provide a more rigorous, mathematically grounded measure of AI’s symbolic reasoning capabilities beyond typical NLP benchmarks.
A web-based RDP client built with Go WebAssembly and grdp (51 points by mariuz)
A web-based RDP client built with Go WebAssembly and grdp
This open-source project (grdpwasm) implements a Remote Desktop Protocol client that runs entirely in the browser using Go WebAssembly. A lightweight Go proxy bridges WebSocket connections from the browser to a TCP-based RDP server, overcoming browser limitations. It enables plugin-free remote desktop access from any modern browser.
Massive capital concentration in frontier AI safety companies
Google’s planned $40B investment in Anthropic is one of the largest single bets on an AI lab, reflecting the high strategic value of safe, aligned AI models. This trend shows that “AI safety” is no longer a niche research field but a core business imperative for Big Tech. Expect similar mega-deals as companies race to secure access to the best models and talent, potentially creating an oligopoly.
Biosafety bounties signal a shift from after-the-fact moderation to proactive capability testing
OpenAI’s GPT-5.5 biosafety bounty program moves beyond red-teaming by incentivizing external researchers to probe for dangerous capabilities (e.g., designing biological weapons) before deployment. This “bug bounty for harms” model could become standard for high-risk AI systems. Implication: future AI releases will likely require public, independent safety audits, and companies will compete on transparency to build trust.
The “agentic” narrative lacks governance and trust mechanisms
Mark Nottingham’s critique highlights a growing unease: as AI agents act autonomously on behalf of users, we lose the “local software” assumption of predictable behavior. This insight points to a critical gap in current AI development—there are no established protocols for auditing, verifying, or constraining agentic behavior. Expect new standards (e.g., agent manifests, behavioral attestations) to emerge, likely driven by regulatory pressure.
Lambda calculus benchmarking targets AI’s reasoning limitations
The LamBench project focuses on symbolic reasoning (lambda calculus) rather than pattern matching on natural language. This signals a shift toward evaluating AI on formal, verifiable tasks where failure modes are clear. It matters because current LLMs often struggle with algorithmic precision; such benchmarks could expose weaknesses that matter for code generation, math, and logic-based applications. Companies that improve on these grounds will gain an edge in enterprise automation.
AI is driving infrastructure innovation (10 GbE, WebAssembly) even indirectly
The new cheaper 10 GbE adapters and the WebAssembly RDP client are not AI themselves, but they enable the heavy data movement and remote access needed for distributed AI workloads (e.g., training clusters, remote inference). The trend is that AI’s insatiable demand for bandwidth, latency, and remote execution is accelerating hardware and web standards innovation. For developers, this means lower-cost high-speed networking and new browser-based tools for AI model management are coming.
Retro computing and open-source preservation intersect with AI analysis
Martin Galway’s C64 music source release and Hokusai pixel art project may seem unrelated to AI, but they hint at a growing interest in using AI to analyze, generate, or remix historical digital artifacts. For example, AI could reconstruct incomplete retro game code or generate new music in a composer’s style. This intersection opens up both creative opportunities and copyright challenges—AI companies will need to navigate fair use of vintage creative works.
The gap between AI safety rhetoric and practical deployment remains wide
Despite billion-dollar investments (Anthropic) and bounty programs, most AI products still lack the transparency and control that Nottingham argues for. The LamBench benchmark shows that fundamental reasoning gaps persist. The overall trend is that the industry is investing heavily in “brand safety” while the actual technical challenges of alignment, verification, and trust are only beginning to be tackled. Actionable takeaway: developers should prioritize building auditable, constraint-based AI interfaces rather than simply wrapping black-box models.
Analysis generated by deepseek-reasoner