Published on April 29, 2026 at 06:00 CEST (UTC+2)
Ghostty is leaving GitHub (1933 points by WadeGrimridge)
Ghostty is leaving GitHub – Mitchell Hashimoto, creator of Ghostty and longtime GitHub user (joined 2008), announces he is leaving the platform. He describes his deep emotional attachment to GitHub, where he spent countless hours over 18 years, including during personal hardships and vacations. The post is a personal farewell, reflecting on how GitHub shaped his open-source identity and career, though it does not fully explain the reasons for leaving.
Before GitHub (317 points by mlex)
Before GitHub – Armin Ronacher reflects on the pre-GitHub era of open source (SourceForge, Trac, Bitbucket) and how GitHub became the social infrastructure for the community. He laments GitHub's recent decline under Microsoft, noting that it was more than just code hosting—it was where relationships and identities formed. He concludes by contemplating what might replace GitHub in the future.
How ChatGPT serves ads (198 points by lmbbuchodi)
How ChatGPT serves ads – A detailed technical analysis of OpenAI's ad platform inside ChatGPT. Ads are injected as structured JSON objects into the SSE stream during conversation responses, with encrypted click tokens and a tracking SDK (OAIQ) on the merchant side. The article provides a full attribution loop, including an example with a Grubhub ad.
Show HN: Auto-Architecture: Karpathy's Loop, pointed at a CPU (32 points by fesens)
Auto-Architecture: Karpathy's Loop, pointed at a CPU – This project applies Andrej Karpathy's autonomous research loop (propose, implement, measure) to hardware design—specifically a 5-stage RV32IM CPU core in SystemVerilog. It tests whether the AI agent methodology can generalize beyond software optimization to hardware architecture, demonstrating that the approach works across domains.
Regression: malware reminder on every read still causes subagent refusals (169 points by thomashobohm)
Regression: malware reminder on every Read still causes subagent refusals – A bug report for Claude Code (v2.1.111) where a system reminder about malware is injected into every Read and Grep tool result, causing subagents to refuse legitimate code edits. A previous fix (v2.1.92) was supposed to resolve this, but the issue has regressed, indicating the fix did not hold.
We decreased our LLM costs with Opus (59 points by shad42)
We decreased our LLM costs with Opus – Mendral describes a triager architecture where a cheap Haiku agent determines if a CI failure is a duplicate (80% of cases) before escalating to the expensive Opus model. This pattern lowers overall cost compared to using Sonnet for everything, because the frontier model is only invoked for novel issues, not for routine duplicates.
OpenAI models coming to Amazon Bedrock: Interview with OpenAI and AWS CEOs (203 points by translocator)
OpenAI models coming to Amazon Bedrock – An interview with OpenAI CEO Sam Altman and AWS CEO Matt Garman about Bedrock Managed Agents powered by OpenAI. Microsoft and OpenAI recently amended their agreement to allow OpenAI on other cloud providers. The interview discusses the competitive dynamics and how this expands access to frontier models.
We still don't have a more precise value for "Big G" (10 points by rbanffy)
We still don't have a more precise value for "Big G" – The article covers the ongoing struggle to measure the gravitational constant precisely. Despite a decade-long effort by NIST to replicate a divergent experimental result, gravity's weakness and Earth's gravitational noise make it difficult. Big G remains the least precise fundamental constant.
I won a championship that doesn't exist (100 points by SEJeff)
I won a championship that doesn't exist – Ron Stoner fabricated a fake "6 Nimmt! World Championship" on Wikipedia and demonstrated that frontier LLMs with web search quoted it back as fact. This shows a cheap, fast form of LLM supply-chain poisoning through the retrieval layer, as opposed to training-time attacks.
Behavioral timescale synaptic plasticity rewires the brain after an experience (74 points by ibobev)
Behavioral timescale synaptic plasticity rewires the brain after an experience – Quanta Magazine reports on a newly discovered form of neuroplasticity that allows the brain to learn from a single experience across longer timescales, beyond the classic "neurons that fire together, wire together" rule. The mechanism explains how the brain consolidates learning from events separated in time.
Open source platforms are losing trust; migration is accelerating – Multiple articles (Ghostty leaving GitHub, Ronacher's reflection) signal a growing dissatisfaction with centralized platforms, especially under corporate ownership. This trend may push AI/ML tooling and model repositories toward federated or self-hosted alternatives. Why it matters: For AI/ML developers, relying on a single platform for collaboration, dataset hosting, and model distribution creates single points of failure and governance risks. Expect more projects to diversify hosting (e.g., GitLab, Forgejo, or decentralized storage).
Monetization of AI chatbots is maturing with real advertising infrastructure – The detailed breakdown of ChatGPT's ad injection shows that LLM-as-a-service is moving beyond subscription fees into programmatic ad revenue. The use of encrypted tokens, tracking SDKs, and SSE-level injection mirrors traditional web advertising but adapted for conversational interfaces. Implications: AI developers must consider privacy implications and potential user backlash. For businesses, this opens new revenue channels but also risks degrading the user experience if ads are poorly integrated.
AI agents are generalizing beyond machine learning to hardware and scientific domains – The auto-architecture project applying Karpathy's research loop to CPU design demonstrates that agentic search-and-evaluate loops can automate discovery in non-software fields. This suggests that AI agents could become general-purpose research accelerators. Actionable takeaway: Companies should explore applying agentic workflows to their own domain-specific design or optimization problems—not just code generation.
LLM safety and reliability remain fragile; regressions highlight the need for better testing – The Claude Code bug regression shows that even supposedly fixed issues can reappear after updates. The malware-reminder injection causes agents to refuse legitimate work, a serious reliability flaw for agentic systems. Why it matters: As AI agents become more autonomous, regressions can cause silent failures or security holes. Developers need continuous integration testing specifically for agent behavior, not just model accuracy.
Cost optimization through model routing (cheap triage + expensive deep reasoning) is becoming standard – Mendral's architecture (Haiku for triage, Opus for hard cases) mirrors the pattern seen in other production systems. By routing 80% of queries to a cheaper model, they reduced overall cost while improving quality on the remaining 20%. Key insight: The future of LLM deployment is not one model fits all—it's hybrid pipelines. Companies should invest in routing logic and cheap classifier models to avoid wasting frontier-model capacity on trivial tasks.
Cloud AI competition is heating up; model exclusivity is breaking down – The OpenAI-AWS Bedrock deal, enabled by Microsoft's amended agreement, signals that cloud providers are no longer locked into exclusive model partnerships. This will likely lead to greater model availability and price competition. Implications: Enterprises gain more flexibility to choose inference providers. However, it also means that AI startups must compete on model quality and specialization rather than platform exclusivity.
LLM retrieval-layer poisoning is a cheap, effective attack vector – The fake championship experiment demonstrates that creating a single Wikipedia hoax can cause frontier LLMs to cite false information. Unlike training-time data poisoning, this attack exploits retrieval-augmented generation (RAG) and works quickly with minimal effort. Why it matters: As AI assistants increasingly rely on web search, malicious actors can manipulate the knowledge base at scale. Defenses like source verification, multiple-source cross-checking, and factuality scoring are urgently needed. For developers, this means RAG pipelines must include trustworthiness scoring, not just relevance ranking.
Analysis generated by deepseek-reasoner