|
📖 Read In Depth
|
GLM-5.2 is the new leading open weights model on Artificial Analysis
GLM-5.2 from Z.ai has taken the top spot on Artificial Analysis's open-weights benchmark, with 1M context, MIT license, and two reasoning modes optimized for long-horizon coding agents. For an ML engineer tracking the open-weights frontier, this is a concrete capability shift worth understanding — especially with SpaceX acquiring Cursor and the competitive pressure on closed models.
hn/Best Stories
|
Running local models is good now
Vicki Boykis (a practitioner, not a hype writer) argues that the local model experience has crossed a threshold into genuinely usable. She covers hardware requirements, model choices, and practical tradeoffs. Directly relevant if you build from scratch and care about understanding the full stack without cloud dependency.
hn/Best Stories
|
Feds freaked over Fable 5 after 'fix this code', not jailbreak, say researchers
The Trump administration's export restriction on Anthropic's Fable 5 was apparently triggered by a researcher using a 'fix this code' prompt — not a sophisticated jailbreak — that caused the model to produce working exploit code as a side effect of writing tests. This raises deep questions about where capability guardrails actually live and whether the administration's response is technical or political.
hn/Best Stories
|
Anthropic Employees Accuse Trump Administration of Targeting Them
The Trump administration is limiting Anthropic's Fable model exports to G7 allies, and Anthropic employees are pushing back, believing the action is politically motivated rather than based on genuine national security concerns. This is a direct collision between AI capability policy and geopolitics — worth understanding as a case study in how government can reshape the AI competitive landscape.
nyt/Technology
|
Is Meta destroying its engineering organization?
Pragmatic Engineer digs into whether Meta's restructuring is systematically degrading its engineering culture and output. The insider comments are pointed: acquired products (Instagram, WhatsApp) ran well; homegrown orgs were dysfunction. Useful signal on how platform-scale engineering organizations succeed or fail under resource pressure and AI-driven headcount cuts.
hn/Best Stories
|
Stop Using JWTs
A technically substantive argument against using JWTs for browser-based user sessions, with the key caveat (from HN comments) that they remain reasonable for service-to-service auth. Worth reading if you care about auth architecture done right — the debate in the HN thread is as useful as the article itself.
hn/Best Stories
|
Elon Musk’s Feud With Delaware May Transform Corporate America
Musk's feud with Delaware over his Tesla pay package is pushing companies to reincorporate elsewhere, potentially rewriting centuries of corporate law precedent. This is a slow-moving platform shift in legal infrastructure that could change how boards, shareholders, and executives interact — relevant to anyone thinking about tech company incentives and governance.
nyt/Top Stories
|
Dear A.I. Companies: The Doom Trolling Needs to Stop
Cal Newport argues that AI labs have normalized a cynical loop: publicly claim their models are dangerous while continuing to ship them, using doom-signaling as marketing and regulatory moat simultaneously. A pointed structural critique of AI lab incentives worth reading critically — you'll likely agree with parts and push back on others.
nyt/Top Stories
|
Iroh 1.0
Iroh 1.0 is a stable release of a peer-to-peer networking library — roughly 'Tailscale at the application layer.' For ML engineers building distributed systems or local-first apps, it's an interesting primitive: lets your app establish direct connections between instances without managing network infrastructure. 1364 HN points suggests real practitioner interest.
hn/Best Stories
|
Calvin and Hobbes and the price of integrity
An essay on Bill Watterson's refusal to license Calvin and Hobbes — and what integrity costs in creative work. The HN thread is rich, with comparisons to Jim Davis's deliberate commercialization of Garfield. A thoughtful piece on the tradeoffs between artistic vision and incentive structures, connecting to broader questions about how systems reward (or punish) integrity.
hn/Best Stories
|
|
âš¡ FYI
|
SpaceX to buy Cursor for $60B
SpaceX is acquiring Cursor (Anysphere) for $60B — its first major post-IPO acquisition. SpaceX told IPO investors it sees a $26T addressable market in AI products, so this is Musk using freshly public stock as currency to consolidate AI coding tools under one roof. Raises serious questions about Cursor's independence and what this means for competitors like GitHub Copilot.
hn/Best Stories
|
U.S. pulling ocean sensors a 'shock' for Canadian research as El Niño nears
The US is decommissioning ocean sensor networks used for El Niño prediction, alarming Canadian and international climate researchers. An oceanographer in the HN comments calls it 'intentional dismantling of American science' by OMB, with new rules also prohibiting international research collaboration. Climate modeling infrastructure is being quietly dismantled in ways that will be hard to reverse.
hn/Best Stories
|
SpaceX’s Growing Warchest
SpaceX's post-IPO stock surge is giving it financial firepower to make $60B acquisitions using stock as currency. The DealBook piece puts the Cursor acquisition in context: SpaceX is positioning itself as an AI company that also builds rockets, not the other way around. Classic platform consolidation move while the stock is near its potential zenith.
nyt/Business
|
China’s Spending Slowdown Deepens as Households Tighten Their Belts
China's retail sales unexpectedly declined, deepening its reliance on exports as domestic consumption contracts. Combined with capital controls restricting outbound investment, this signals a Chinese economy under structural stress — relevant to anyone tracking semiconductor supply chains and geopolitics of compute, where China's economic trajectory matters.
nyt/Business
|
Engineered van der Waals crystal mimics neuronal cells with light-driven learning
Researchers engineered a van der Waals crystal that mimics light-driven learning in neuronal cells — a materials-level step toward physical neuromorphic computing. Early-stage and speculative, but the cross-domain intersection of materials science and neural computation is exactly the kind of unexpected connection worth a quick look.
reddit/r/singularity
|