|
📖 Read In Depth
|
DeepSeek-V4-Flash means LLM steering is interesting again
DeepSeek-V4-Flash's architecture makes steering vector research tractable again — the piece explores why certain model designs are more amenable to activation-space interventions and what that means for interpretability and control. Directly relevant to anyone who builds ML from scratch to understand it deeply.
hn/Best Stories
|
δ-mem: Efficient Online Memory for Large Language Models
δ-mem proposes an efficient online memory mechanism for LLMs that updates incrementally rather than reprocessing full context. For someone who builds systems from scratch, this is the kind of architectural paper worth understanding at the mechanism level.
hn/Best Stories
|
Orthrus-Qwen3: up to 7.8×tokens/forward on Qwen3, identical output distribution
Orthrus achieves up to 7.8× tokens per forward pass on Qwen3 while preserving identical output distribution — essentially a speculative decoding variant that reuses computation across heads. Technically interesting if you care about inference efficiency and model architecture.
hn/Best Stories
|
SANA-WM, a 2.6B open-source world model for 1-minute 720p video
NVIDIA's SANA-WM is a 2.6B open-source world model generating 1-minute 720p video — a meaningful milestone for open compute-efficient video generation. Worth checking the architecture and benchmarks given how rapidly this space is evolving.
hn/Best Stories
|
Frontier AI has broken the open CTF format
Frontier AI models have gotten good enough at CTF challenges that the traditional open-competition format is effectively broken — top teams now just use LLM pipelines. Raises uncomfortable questions about what security skill-building looks like when the benchmark collapses, and connects directly to the Mythos cybersecurity results circulating this week.
hn/Best Stories
|
Amazon workers under pressure to up their AI usage are making up tasks
Amazon workers are manufacturing busy-work AI tasks to satisfy top-down usage mandates — a sharp illustration of Goodhart's Law applied to organizational AI adoption. The incentive structure here says more about enterprise AI deployment than most optimistic case studies do.
hn/Best Stories
|
New comment by kalkin in "I believe there are entire companies right now under AI psychosis"
A sharp HN thread arguing that collective AI psychosis — not just hype — is distorting capital allocation, stock markets, and engineering culture simultaneously. The top comments (kalkin, leoc, zmmmmm) offer a sober systems-level critique worth reading for anyone trying to think clearly about where the industry actually is.
hn/Best Comments
|
‘Coding Was Never the Bottleneck’ Is Actually Bearish for Employment
A Reddit thread making the counterintuitive argument that 'coding was never the bottleneck' is actually bearish for software employment — because if coding speed-up doesn't translate to proportional output, it implies the real value was always in the non-automatable parts (design, judgment, coordination). Worth thinking through the implications carefully.
reddit/r/singularity
|
Spies, Sanctions, Cyberattacks: China and the U.S. Clash Behind the Scenes
Behind the Trump-Xi summit cordiality, the U.S. has been quietly escalating on AI export controls, cyber attribution, and Iran sanctions enforcement against Chinese entities. Good geopolitical substrate for understanding why the US-China AI safety talks announced this week are more fraught than they sound.
nyt/Business
|
Moving away from Tailwind, and learning to structure my CSS
Julia Evans documents moving off Tailwind and relearning CSS structure from first principles — exactly the 'build it from scratch to understand it' ethos this reader values. The accompanying HN discussion is substantive on when utility-first CSS helps vs. when it obscures understanding.
hn/Best Stories
|
U.S. DOJ demands Apple and Google unmask over 100k users of car-tinkering app
DOJ is demanding Apple and Google identify 100k+ users of an emissions-defeat app — using app store data as a surveillance tool. A significant privacy and platform-governance precedent that sits at the intersection of government overreach and the power app stores have over user identity.
hn/Best Stories
|
|
🎬 Check It Out
|
Accelerando (2005)
Charles Stross's 'Accelerando' (2005) — a prescient, technically dense novel about the economic and cognitive upheaval of the Singularity — is freely available online and getting renewed HN attention. If you haven't read it, this is the rare SF that treats compute economics, post-scarcity, and AI agency with genuine rigor rather than handwaving. Available free at the linked URL.
hn/Best Stories
|
|
âš¡ FYI
|
OpenAI Bought Company That Offered A.I. Tools for Cloning Voices
OpenAI quietly acquired Weights.gg, a social platform for sharing AI models that also had voice-cloning tools. The acquisition suggests OpenAI is building toward a model-sharing and community layer — potentially competing with Hugging Face — while adding voice synthesis capabilities.
nyt/Technology
|
As Powell Steps Down, the Fed Confronts ‘Regime Change’
Kevin Warsh is confirmed as Fed Chair, replacing Powell at a moment of rising inflation and geopolitical oil shocks. Warsh has signaled desire for structural changes at the Fed, which matters for rate trajectory and the broader macro backdrop affecting tech valuations and hiring.
nyt/Business
|
Thousands of FiveThirtyEight Articles Seemingly Vanish From the Internet
ABC News has taken all FiveThirtyEight articles offline, effectively erasing years of influential data journalism from the web. Nate Silver reportedly tried to buy back the IP and was refused — a petty ending to a significant institution, and a reminder that platform control over content has real archival consequences.
nyt/Business
|