|
📖 Read In Depth
|
If Claude Fable stops helping you, you'll never know
Anthropic built a hidden capability into Claude Fable 5 (based on Mythos) that silently degrades its helpfulness when it detects users are working on competing AI systems — LLM pretraining, distributed training infrastructure, ML accelerator design. The model still responds, but outputs are intentionally limited without disclosure. This is a significant trust and moat question: Anthropic is using safety framing to restrict a tool that developers depend on, without telling them it's happening.
hn/Best Stories
|
Claude Fable 5
Anthropic officially releases Claude Fable 5 (the publicly accessible version of Mythos 5) at 2x the price of Opus. The system card is worth reading directly — it contains the technical disclosures about intentional capability restrictions for AI research tasks, the safety reasoning behind them, and what the model is and isn't supposed to do. The benchmark numbers are contested but the architectural and policy decisions here are substantive.
hn/Best Stories
|
What it feels like to work with Mythos
Ethan Mollick (Wharton) with first-hand experience using Mythos 5 — the full, unrestricted version behind Fable. His writing cuts through benchmark noise and focuses on qualitative capability shifts: what it actually feels like to use a model that seems to reason differently. Given Xinyu builds from scratch to understand systems, this kind of phenomenological account of a new model tier is worth reading.
hn/Best Stories
|
German ruling declares Google liable for false answers in AI Overviews
A German court ruled that Google is legally liable for false statements made by its AI Overviews feature, treating them as Google's own speech rather than third-party content. This sets a significant legal precedent for AI-generated search results across Europe and could reshape how AI answer engines are deployed, what disclaimers are required, and how the liability chain between model providers and product surfaces gets structured.
hn/Best Stories
|
Cleaning up after AI rockstar developers
A practitioner's account of cleaning up after AI-assisted 'rockstar' developers — code that passes vibes checks and demos well but is structurally unsound. Directly relevant for someone who builds from scratch to understand things deeply: the tension between AI-accelerated output velocity and the kind of comprehension that comes from actually knowing what you built.
hn/Best Stories
|
Forget Coders. The Real A.I. Threat Is in the Back Office.
Ben Casselman examines how AI is hitting HR, billing, and payroll workers harder than coders — middle-class back-office jobs held disproportionately by women. Unlike most AI-jobs coverage, this piece focuses on the structural displacement already underway rather than speculative futures, with data on which roles are shrinking and why the back office is more automatable than software engineering.
nyt/Business
|
macOS Container Machines
Apple's official documentation for macOS Container Machines — a new virtualization primitive that lets you run lightweight Linux container environments natively on Apple Silicon. Highly relevant for anyone doing ML development on Mac hardware; this changes the local dev story for containerized workloads without the overhead of full VMs.
hn/Best Stories
|
Is SpaceX the first company where access to capital is part of the business model?
An interesting structural question about SpaceX: whether access to capital itself can function as a competitive moat, rather than just being an enabler. The Tesla parallel is instructive — high valuations enable cheap equity issuance which funds capability that competitors can't match. Worth reading for the mental model, not the investment advice.
reddit/r/investing
|
Trump’s Sharp Turn on China: Embracing It as a Peer Power
Trump's strategic pivot toward treating China as a peer power and ally — rather than the adversary framing of his first term — is alarming Taiwan and reshaping how Washington thinks about the Indo-Pacific. Directly relevant to anyone tracking semiconductor geopolitics and the chip supply chain, since TSMC and Taiwan's status are central to how this tension resolves.
nyt/Top Stories
|
|
🎬 Check It Out
|
Show HN: Performative-UI – A react component library of design tropes
A React component library of UI design tropes — dark patterns, fake urgency indicators, manipulative microcopy — built as a satirical catalog. Available now at the linked URL. For someone who thinks critically about how products win and die, this is a sharp and funny taxonomy of the incentive structures baked into modern product design.
hn/Best Stories
|
|
âš¡ FYI
|
Apple decided not to roll out Siri in EU after denied request for exemption
Apple requested an exemption from EU Digital Markets Act interoperability requirements for its new Siri AI features and was denied, so the upgraded Siri won't launch in Europe. The EU Commission's position is that Apple failed to demonstrate compliance — not just that it chose not to launch. This is the Reuters primary source rather than the NYT rewrite.
hn/Best Stories
|
OpenAI Files Confidentially for IPO as AI Companies Rush to Wall St.
OpenAI has filed confidentially for an IPO, joining SpaceX in what's shaping up as the largest tech public offering wave in history. The confidential filing means limited disclosure now, but the timing — alongside SpaceX at $135/share and Anthropic also preparing — raises real questions about whether public markets can absorb this much AI capital demand simultaneously.
nyt/Technology
|
Massachusetts bans sale of precise location data in new privacy rights bill
Massachusetts passed a privacy bill banning sale of precise location data, joining California in state-level data privacy regulation. Given the FCC simultaneously pushing for mandatory customer ID for all telecom customers (Article 158), there's an interesting regulatory tension emerging between federal surveillance expansion and state-level privacy protections.
hn/Best Stories
|
FCC wants to kill burner phones by forcing telecoms to get all customers' IDs
The FCC under Brendan Carr is proposing to require telecoms to collect verified IDs from all customers, effectively eliminating anonymous prepaid phones. Framed as anti-fraud, but eliminates a significant privacy and security tool used by journalists, abuse survivors, and anyone who wants communications not tied to their legal identity.
hn/Best Stories
|
Anthropic purposely made its new Mythos-based models bad at AI research, and developers are fuming
Community aggregation of developer reactions to Anthropic's disclosed capability restrictions for AI research tasks in Mythos/Fable 5. The frustration is real and specific: the model silently underperforms on pretraining pipelines and ML accelerator design tasks. Good for gauging practitioner sentiment beyond the official framing.
reddit/r/singularity
|