|
📖 Read In Depth
|
Claude Design
Anthropic launched Claude Design, a new design initiative out of Anthropic Labs. This is getting significant HN traction (1095 points) and the comments debate whether AI-generated UI homogenizes design or is actually fine for most use cases — a real tension worth thinking through if you're building products on top of these models.
hn/Best Stories
|
Measuring Claude 4.7's tokenizer costs
Empirical breakdown of Claude 4.7's updated tokenizer and what it actually costs in practice — the new tokenizer maps the same input to 1.0–1.35× more tokens depending on content type. If you're building on the Claude API, this directly affects your cost model and budget planning.
hn/Best Stories
|
Qwen3.6-35B-A3B on my laptop drew me a better pelican than Claude Opus 4.7
Simon Willison runs Qwen3.6-35B-A3B (a 35B MoE model, only 3B active parameters) locally via LM Studio and finds it outperforms Claude Opus 4.7 on a visual generation task. This is a meaningful data point on open-weight MoE models catching up to frontier closed models on a laptop — directly relevant to anyone tracking the build-vs-buy question for inference.
hn/Best Stories
|
Ban the sale of precise geolocation
Lawfare piece arguing for a hard ban on selling precise geolocation data, not just 'anonymized' aggregates. The core argument — that sleep/work location pairs make anonymization a fiction — is technically sound and has real policy momentum. Worth reading for the legal and technical framing, especially as this could reshape data pipelines that many ML products depend on.
hn/Best Stories
|
How Do You Measure an A.I. Boom?
METR's AI task complexity chart has become the industry's go-to metric for tracking AI capability progress — this piece examines what it actually measures, its limitations, and why everyone has latched onto it. Good for anyone who wants to think critically about how 'progress' is being defined and potentially gamed.
nyt/Technology
|
"cat readme.txt" is not safe if you use iTerm2
Security vulnerability in iTerm2 where simply running `cat` on a crafted file can trigger unintended behavior via terminal escape sequences. A sharp reminder that trust boundaries in terminal emulators are more porous than most engineers assume — practical and technically interesting.
hn/Best Stories
|
Show HN: Smol machines – subsecond coldstart, portable virtual machines
A Show HN for subsecond cold-start portable virtual machines — the kind of infrastructure primitive that matters for serverless and edge compute. If you're interested in how fast VM startup actually gets, this is worth looking at the implementation approach.
hn/Best Stories
|
That Meeting You Hate May Keep A.I. From Stealing Your Job
As AI automates individual tasks, the work that remains is the social coordination layer — cajoling, persuading, managing ambiguity between humans. The piece makes a structural argument about which labor is actually complementary to AI, not just 'soft skills are safe' fluff.
nyt/Business
|
New unsealed records reveal Amazon's price-fixing tactics, California AG claims
Newly unsealed records from California's AG lawsuit detail Amazon's alleged price-fixing tactics — specifically how their algorithm enforces pricing parity across the market. Strong on the mechanics of platform power and how marketplaces extract rent from sellers and consumers simultaneously.
hn/Best Stories
|
Everything we like is a psyop?
Examines how the proliferation of AI-generated fake social media influencers (like the pro-Trump avatar wave) is collapsing people's ability to distinguish organic preference from manufactured consensus. Relevant both as a technology story and as a challenge to assumption-challenging about how culture and opinion actually form online.
hn/Best Stories
|
Spending 3 months coding by hand
A developer deliberately codes without AI assistance for three months to understand what they actually know versus what they've been offloading. The HN comments thread (especially the 6502 assembly teaching angle) is worth reading alongside — directly relevant to someone who builds algorithms from scratch to understand them deeply.
hn/Best Stories
|
Ada, its design, and the language that built the languages
A deep dive into Ada's design philosophy and how it influenced subsequent language design — strong vs. static typing, contract programming, and the emphasis on correctness over expressiveness. For someone who builds from scratch to understand things, this is the kind of foundational language history that connects fields.
hn/Best Stories
|
|
âš¡ FYI
|
Hyperscalers have already outspent most famous US megaprojects
A chart showing hyperscaler AI capex has already exceeded most famous US megaprojects in nominal dollars. The follow-up context (% of GDP comparison) makes it less dramatic but still remarkable — useful for calibrating how serious the infrastructure build-out actually is.
hn/Best Stories
|
Cerebras, an A.I. Chip Maker, Files to Go Public as Tech Offerings Ramp Up
Cerebras, the wafer-scale AI chip company, has filed to go public, joining a wave that includes SpaceX, Anthropic, and OpenAI. Cerebras's architecture (a single giant chip vs. GPU clusters) is a real architectural bet worth watching — an IPO filing will force more transparency on their unit economics and customer traction.
nyt/Technology
|