|
📖 Read In Depth
|
AI is slowing down
A substantive contrarian argument that AI capability progress is genuinely decelerating — not just hype normalization. Worth reading critically given how much of the current IPO frenzy (OpenAI, Anthropic, SpaceX compute deals) is predicated on continued exponential improvement. The framing and evidence matter here, not just the headline.
hn/Best Stories
|
Apple reveals new AI architecture built around Google Gemini models
Apple's WWDC reveal shows they're building their AI stack on top of Google Gemini models — a significant architectural and strategic bet. This raises real questions about Apple's AI moat, their private compute infrastructure claims, and how deeply they're ceding model development to Google while retaining the UX layer.
hn/Best Stories
|
xAI is looking more like a datacentre REIT than a frontier lab
Argues that xAI is structurally behaving more like a datacenter REIT than a frontier AI lab — renting compute rather than doing leading-edge research. If accurate, this reframes the competitive dynamics around Musk's AI play and has implications for how to evaluate the xAI/SpaceX compute relationships ahead of the IPO wave.
hn/Best Stories
|
MiMo-v2.5-Pro-UltraSpeed: 1T model with 1000 tokens per second
Xiaomi's MiMo-v2.5-Pro achieves 1000+ tokens/sec on a 1T parameter model running on 8x commodity GPUs — no Groq, no Cerebras. They combined FP4 quantization with QAT, DFlash speculative decoding, and TileRT kernels. This is the kind of systems-level inference engineering that matters for understanding where the cost curve is heading.
hn/Best Stories
|
Apple Core AI Framework
Apple's new Core AI framework documentation — the actual API surface for on-device ML on Apple Silicon. For someone who builds algorithms from scratch, this is the right level of detail to understand what Apple is exposing to developers vs. keeping proprietary, and how it layers over their existing CoreML stack.
hn/Best Stories
|
Surveillance is not safety: A statement on the UK's latest threat to privacy [pdf]
Signal's formal statement against the UK's latest push for backdoor access — framed as a technical and philosophical argument, not just advocacy. The core claim is that surveillance infrastructure cannot be selectively safe, which is directly relevant to how AI systems handling private data (like Apple's privacy-preserving claims) get evaluated.
hn/Best Stories
|
The Smallest Brain You Can Build: A Perceptron in Python
A from-scratch perceptron implementation in Python — directly aligned with Xinyu's philosophy of building to understand. Worth scanning to see if the implementation choices and explanations are genuinely illuminating or just another tutorial; the HN traction (300+ points) suggests it resonates with practitioners.
hn/Best Stories
|
Making Graphics Like it's 1993
A deep dive into recreating 1993-era 3D graphics from scratch — raycasting, software rendering, the whole stack. This sits at the intersection of systems programming as craft and understanding how graphics pipelines actually work before hardware abstraction layers. The kind of project that rewards the build-to-understand mindset.
hn/Best Stories
|
|
âš¡ FYI
|
Confidential submission of draft S-1 to the SEC
OpenAI has filed a confidential S-1 with the SEC, formally kicking off the IPO process. Combined with SpaceX's pricing at $135/share and Anthropic also reportedly in the pipeline, the market is being asked to absorb multiple trillion-dollar AI valuations simultaneously — a stress test for investor appetite that has real implications for capital allocation across the sector.
hn/Best Stories
|
Microsoft's open source tools were hacked to steal passwords of AI developers
Microsoft's open-source developer tooling was compromised in a supply chain attack specifically targeting AI developers' credentials. This is a meaningful supply-chain security event for anyone in ML infrastructure — the attack surface isn't just models, it's the tooling ecosystem around them.
hn/Best Stories
|
S&P 500 will not be fast tracking SpaceX entry into its index and it won't waive its rule for unprofitable AI companies
S&P 500 declined to fast-track SpaceX into its index, refusing to waive profitability rules — which would have also opened the door for OpenAI and Anthropic. Nasdaq and Russell are still modifying their rules. The asymmetry here is interesting: passive investors in Nasdaq/Russell funds will get exposure whether they want it or not, while S&P 500 holders won't.
reddit/r/investing
|