Published on June 15, 2026 at 18:00 CEST (UTC+2)
Iroh 1.0 (126 points by chadfowler)
Iroh 1.0 introduces a new internet abstraction that replaces IP addresses with user-controlled cryptographic keys, enabling devices to be securely addressable regardless of network changes or firewalls. The project, now stable after four years of development, is already used for streaming video, training LLMs, and agent communication across millions of devices. It emphasizes open standards and built its own QUIC multipath implementation for resilient connections.
CrankGPT (303 points by rishikeshs)
CrankGPT is a satirical take on AI consumption, proposing human-powered, fully local AI models that run on manual cranks, pedals, or gym partnerships instead of energy-intensive datacenters. It mocks the climate hypocrisy of tech companies and promotes privacy by keeping data on-device, offering tiered models from basic question-answering to enterprise agent swarms. The project highlights the absurdity of current AI energy use while advocating for "rightsized" AI.
Copper transport drug restores memory and clears toxic Alzheimer's proteins (22 points by bookofjoe)
A new copper transport drug has been shown to restore memory and clear toxic Alzheimer’s proteins in preclinical studies at Monash University. The compound targets copper regulation in the brain, which is linked to amyloid plaque formation and cognitive decline. This research opens a potential therapeutic pathway for Alzheimer’s disease by addressing metal homeostasis.
Show HN: Exploiting Slack's video embeds to achieve E2EE communication (5 points by victorio)
This project exploits Slack’s video embed block to achieve end-to-end encrypted communication by encrypting messages client-side and embedding the decryption key in a video URL that Slack treats as an iframe. The server never sees the unencrypted key, allowing users to send encrypted messages through Slack without its knowledge. The implementation uses browser crypto APIs and TypeScript, demonstrating a creative workaround for E2EE in a non-encrypted platform.
Hetzner increased dedicated server prices 3-4x (164 points by enescakir)
Hetzner has dramatically increased prices for dedicated servers by 3-4x (e.g., AX102 from €124 to €454), citing rising hardware costs for RAM and disk. This follows a previous ~30% hike, sparking discussion about how hyperscalers like AWS and GCP manage to keep prices stable while smaller providers struggle. Commenters worry about the concentration of compute resources among large cloud providers and the broader decline of personal computing.
Apple Foundation Models (357 points by MehrdadKhnzd)
Apple Foundation Models now integrate with Claude via a Swift package, allowing developers to use Claude as a server-side language model within Apple’s Foundation Models framework. The package conforms to the LanguageModel protocol, supporting streaming, guided generation, and tool calling, with requests going directly to Anthropic’s API (not Apple). This enables apps to seamlessly switch between Apple’s on-device model and Claude depending on the task.
Your ePub Is fine (796 points by sohkamyung)
The article argues that EPUB files are often fine, but Kobo devices reject them due to Adobe’s DRM and rendering engine issues, not the file itself. It explains how Adobe’s global DRM system can break valid EPUBs, forcing users to convert formats unnecessarily. The author calls for blame to be placed on Adobe rather than the EPUB standard.
Salesforce to Acquire Fin (formerly Intercom) for $3.6BN (193 points by colesantiago)
Salesforce is acquiring Fin (formerly Intercom) for $3.6 billion to bolster its AI agent capabilities with Fin’s proprietary Apex model, which outperforms frontier models on customer support queries. The acquisition aims to accelerate Salesforce’s vision of an "agentic enterprise" by integrating Fin’s multi-channel AI agent into its Agentforce platform. This marks a major consolidation in the AI customer service space.
Show HN: Can Europe train a frontier AI model on the compute it owns? (34 points by smashini)
EuroMesh explores whether Europe can train a frontier-class AI model by federating existing public compute (EuroHPC supercomputers and national AI factories) instead of waiting years for new gigawatt datacenters. Using low-communication techniques like DiLoCo, the model shows that federated training on tens of exaflops of compute is feasible as a stopgap. The report highlights the gap between planned datacenter timelines and immediate AI compute needs.
Openrouter Fusion API (146 points by tdchaitanya)
OpenRouter’s Fusion API turns a single prompt into a multi-model deliberation: a panel of expert models analyzes the query in parallel with web search, then a judge model synthesizes their responses into a structured output covering consensus, contradictions, and blind spots. It is priced as the sum of all underlying completions, offering higher accuracy for research and critical tasks where a single model may be insufficient. The API is OpenAI-compatible and supports customization of panel members and judge models.
1. Decentralized compute and identity are becoming critical for AI workloads
- Trend: Projects like Iroh and EuroMesh emphasize moving away from centralized IP-based addressing and cloud datacenters toward user-controlled keys and federated compute.
- Why it matters: AI training and inference increasingly require reliable, low-latency connectivity across distributed nodes. Current IP infrastructure is brittle behind firewalls, and hyperscaler-dominated compute is expensive and slow to expand.
- Implications: Federated training (e.g., DiLoCo) and decentralized networking (e.g., QUIC multipath) will enable smaller players to pool resources for frontier AI, reducing dependence on a few big cloud providers.
2. Energy and hardware costs are reshaping AI deployment strategies
- Trend: Hetzner’s 3-4x price hike and the satirical CrankGPT both highlight soaring hardware costs and energy demands for AI. Meanwhile, EuroMesh argues for using existing public compute to avoid waiting for new power-hungry datacenters.
- Why it matters: The AI boom is straining supply chains for RAM, disk, and GPU, and grid capacity for new datacenters. Costs are passed to end users, making AI less accessible.
- Implications: Expect more efficient training methods (model distillation, low-communication federated learning) and a push toward “rightsized” AI—using smaller, local models where possible. Also, energy-aware pricing and carbon offsetting will become competitive differentiators.
3. AI acquisitions signal consolidation in enterprise agent market
- Trend: Salesforce’s $3.6B acquisition of Fin (Intercom) shows major CRM players buying specialized AI agent startups to gain proprietary models and proven resolution rates.
- Why it matters: The customer support AI market is maturing, with consistent revenue models (per-resolution or subscription). Large incumbents want to own the stack end-to-end, from front-end chat to back-end workflow automation.
- Implications: Smaller AI agent startups will face acquisition pressure or need to differentiate on niche verticals. We may see a “platform war” among Salesforce, Zendesk, and others over agent orchestration.
4. Multi-model reasoning is becoming a first-class API product
- Trend: OpenRouter’s Fusion API enables multi-model deliberation, where a panel of models analyzes a prompt and a judge synthesizes the output. This is a shift from single-model calls to ensemble reasoning as a service.
- Why it matters: Single models have blind spots and biases. Fusion-style architectures improve accuracy for complex research, medical diagnosis, and legal analysis by surfacing contradictions and consensus.
- Implications: Expect more APIs offering “model routers” and “deliberation engines” as standard features. This also incentivizes model developers to expose intermediate reasoning for better judge performance.
5. On-device and private AI gains traction through creative integrations
- Trend: Apple Foundation Models now let developers use Claude alongside Apple’s on-device model, while CrankGPT parodies the privacy/energy trade-off, and the Slack E2EE hack shows demand for private communication in popular platforms.
- Why it matters: Users and regulators increasingly demand data sovereignty. Running AI locally (or through trusted APIs) reduces attack surface and avoids training on sensitive data.
- Implications: Apple’s move will pressure other OS vendors (Google, Microsoft) to offer similar hybrid local/cloud AI frameworks. App developers will need to design for dynamic model switching based on task complexity and privacy requirements.
Analysis generated by deepseek-reasoner