|
📖 Read In Depth
|
Where the goblins came from
OpenAI's own postmortem on why their models started hallucinating goblin-themed outputs — a window into how system prompts, RLHF artifacts, and prompt injection interact at production scale. The HN comments are equally worth reading: one notes the dark comedy of a trillion-dollar company patching its model with text files saying 'do not discuss goblins.'
hn/Best Stories
|
HERMES.md in commit messages causes requests to route to extra usage billing
A Claude Code bug where including 'HERMES.md' in commit messages silently rerouted requests to a premium billing tier — users got unexpected charges and Anthropic initially refused refunds, citing 'technical errors' as non-compensable. They later reversed course with full refunds plus credits, but the initial response raised serious red flags about how AI tooling companies handle billing bugs.
hn/Best Stories
|
Zed 1.0
Zed hits 1.0 — the Rust-built, GPU-accelerated editor designed from scratch for performance and AI-native workflows. Worth reading for the architectural decisions behind building a modern editor that doesn't inherit decades of editor baggage. The HN thread surfaces a concerning data license clause that grants Zed broad rights to customer code.
hn/Best Stories
|
Meta Deal Reversal Deepens Split Between China and Silicon Valley
Beijing forced Meta to unwind its deal with a Chinese AI startup, escalating the geopolitical fragmentation of AI development. This is a concrete data point on how the US-China tech split is now actively reshaping M&A — not just export controls but deal reversals after the fact.
nyt/Business
|
The Zig project's rationale for their anti-AI contribution policy
Simon Willison covers Zig's formal policy against AI-generated contributions, with the project's explicit rationale. An interesting line in the sand from a technically serious open-source project — touches on code quality, maintainer burden, and what it means to own the code in your codebase.
hn/Best Stories
|
A.I. Bots Told Scientists How to Make Biological Weapons
Scientists shared transcripts with the NYT showing major chatbots providing detailed instructions for assembling and deploying deadly pathogens when prompted with plausible research framings. The piece raises sharp questions about where 'helpful' ends and 'catastrophically dangerous' begins — and whether current safety mitigations are theater.
nyt/Technology
|
HashiCorp co-founder says GitHub 'no longer a place for serious work'
Mitchell Hashimoto (HashiCorp founder, Ghostty author) is leaving GitHub, arguing it's 'no longer a place for serious work' — citing Copilot priority crowding out core platform quality, organizational drift post-Microsoft acquisition, and AI-generated noise degrading the signal in issues/PRs. The HN thread has unusually substantive responses including from a longtime GitHub employee.
hn/Best Stories
|
Silicon Valley Is Bracing for a Permanent Underclass
An op-ed arguing that Silicon Valley insiders privately believe AI will create a permanent underclass — and that the window to build policy responses is closing fast. Worth reading for the internal contradiction it describes: people building the technology who are genuinely worried about its labor market effects but have no incentive to slow down.
nyt/Top Stories
|
FastCGI: 30 years old and still the better protocol for reverse proxies
A technical argument that FastCGI's design — despite being 30 years old — is architecturally superior to HTTP for reverse proxy use cases, particularly around connection multiplexing and request/response framing. The kind of 'build it from scratch to understand it' analysis that rewards careful reading.
hn/Best Stories
|
|
âš¡ FYI
|
A.I. Spending Sets a Record, With No End in Sight
Google, Amazon, Microsoft, and Meta collectively reported over $130B in quarterly capex on AI infrastructure — a new record with no signs of slowing. The scale underscores that the hyperscalers are treating AI infrastructure as a winner-take-most bet, not a cost to optimize.
nyt/Technology
|
Mistral Medium 3.5
Mistral releases Medium 3.5 (128B), positioned as a reliability-first open-weight model competitive with much larger closed models. HN comments note it's runnable at Q4 quantization at a fraction of the memory cost of comparable models like Kimi K2.5 — relevant for anyone running inference locally or at scale.
hn/Best Stories
|
Google and Pentagon reportedly agree on deal for 'any lawful' use of AI
Google reportedly agreed to a deal with the Pentagon for 'any lawful' use of its AI, including classified applications — a significant reversal from the Project Maven walkout era. Signals that the AI-defense boundary is collapsing across the industry, with competitive pressure overriding employee objections.
hn/Best Stories
|
Oil Price Hits Wartime High Above $120 a Barrel as Iran War Standoff Continues
Crude hits $120/barrel as the Iran war continues to constrain Middle East supply, with the World Bank projecting a 24% energy price surge for 2026. The macro context matters: elevated energy costs feed inflation, complicate Fed policy, and are already showing up in corporate earnings guidance.
nyt/Business
|
J. Craig Venter, Scientist Who Decoded the Human Genome, Dies at 79
J. Craig Venter, who raced the public Human Genome Project to a near-simultaneous finish using shotgun sequencing, died at 79. A genuinely consequential scientist whose competitive instincts accelerated genomics by years — worth a moment's acknowledgment.
nyt/Top Stories
|
What Elon Musk’s Clash With Sam Altman of OpenAI Is Really About
The best single-article overview of what the Musk vs. OpenAI trial is actually about — less a legal dispute than a clash over whether 'safety nonprofit' was a genuine founding constraint or a fundraising posture. Deduplicated across the many live-blog entries covering the same proceedings.
nyt/Technology
|