Published on June 16, 2026 at 18:01 CEST (UTC+2)
Running local models is good now (176 points by jfb)
Running local models is good now – Vicki Boykis argues that local AI models have finally become viable for serious use, citing her own experience with a 2022 M2 Mac. Early models were slow and inaccurate, but newer releases like GPT-OSS and Gemma 4 have closed the gap with cloud APIs. She now uses local models for fast, personalized coding tasks and even agentic workflows, achieving ~75% of frontier model performance. The post highlights a major shift: local inference is no longer a compromise but a practical alternative for many developers.
Mechanical Watch (2022) (423 points by razin)
Mechanical Watch (2022) – This is an interactive, in-depth technical exploration of how mechanical watches work, focusing on the movement (the internal mechanism). The author uses 3D visualizations and animations to explain each component, from the mainspring to the escapement, without relying on batteries or electronics. The article is educational and not related to AI/ML.
Subquadratic – Introducing SubQ 1.1 Small (36 points by EDM115)
Subquadratic – Introducing SubQ 1.1 Small – SubQ announces a new model based on Subquadratic Sparse Attention (SSA) that dramatically reduces the compute cost of long-context reasoning. The model handles up to 12M tokens with near-perfect retrieval, using 64.5x less compute than dense attention and running 56x faster than FlashAttention-2. It targets enterprise use cases like analyzing entire codebases or document collections, bypassing the need for chunking or RAG pipelines.
Fable ban was never about a jailbreak (29 points by amarant)
Fable ban was never about a jailbreak – The U.S. Commerce Department invoked an obscure export control directive to force Anthropic to take down its latest models (Fable 5 and Mythos 5), citing national security. Anthropic complied by shutting down the models to all customers, even though the letter lacked specifics. The article warns that this unilateral government action sets a precedent for the entire tech industry, showing that AI companies can be shut down without court approval.
Correlated randomness in Slay the Spire 2 (174 points by rdmuser)
Correlated randomness in Slay the Spire 2 – The author discovers that the game’s random number generators are unexpectedly correlated, making some outcomes predictable (e.g., a curse being ~54% likely to be “Debt”). This happens because the game’s RNGs share a common seed state, so knowledge of one generator’s output leaks information about others. The post is a deep dive into game design and probability, not directly related to AI/ML but relevant to understanding randomness in systems.
I admire Fabrice Bellard. He is almost certainly a better overall programmer (641 points by apitman)
I admire Fabrice Bellard – John Carmack tweets admiration for Fabrice Bellard, a French programmer who created foundational software like FFmpeg (used by YouTube, Netflix) and QEMU. The thread highlights how Bellard’s quiet, prolific work underpins much of the modern internet. The post is a tribute to exceptional programming skill, not an AI/ML article.
Google Chrome's Next Update Will Mark the End of Popular Ad Blockers (68 points by arnejenssen)
Google Chrome's Next Update Will Mark the End of Popular Ad Blockers – Chrome is removing the last flag that allowed Manifest V2 extensions, effectively killing legacy ad blockers like uBlock Origin. Google cites security risks and technical debt. This affects all Chromium-based browsers, though they may choose to keep support. The move is part of Google’s long-term plan to shift to Manifest V3, which limits ad-blocking capabilities.
A backdoor in a LinkedIn job offer (1451 points by lwhsiao)
A backdoor in a LinkedIn job offer – A developer receives a LinkedIn message asking him to review a public GitHub repo. Suspicious, he spins up a throwaway VPS and uses an AI agent to inspect the code. The agent discovers a hidden backdoor: test code that assembles a URL and executes arbitrary payloads. The story illustrates a sophisticated social engineering attack targeting developers, leveraging fake job offers to introduce malware.
The time the x86 emulator team found code so bad they fixed it during emulation (421 points by paulmooreparks)
The time the x86 emulator team found code so bad they fixed it during emulation – Raymond Chen recounts a story where a compiler generated 65,536 individual byte-write instructions to initialize a 64KB buffer (instead of a loop). The x86 emulator team optimized the emulated code on the fly by recognizing the pattern and replacing it with a single efficient operation. This tale highlights the ingenuity of JIT compilation and binary translation.
An interview with an Apple emoji designer (50 points by nate)
An interview with an Apple emoji designer – The author of “Face with Tears of Joy” interviews Ollie Wagner, one of Apple’s first emoji designers (2008 intern). Wagner discusses the design process, challenges, and the early days of emoji at Apple. The article is a historical piece about visual communication, not AI/ML.
Local models have crossed the viability threshold – The article “Running local models is good now” confirms that open-source models like Gemma 4 and GPT-OSS now perform well enough for serious coding and agentic tasks on consumer hardware (e.g., M2 Mac with 64GB RAM). This trend matters because it reduces reliance on API-based models, cutting costs, improving latency, and enabling offline or privacy-sensitive applications. Actionable takeaway: teams should reevaluate their inference stack to include local models for non-reasoning-heavy workflows.
Subquadratic attention is breaking the long-context barrier – SubQ 1.1 Small’s sparse attention mechanism achieves 1,000x compute reduction for long-context retrieval while maintaining strong general reasoning. This directly challenges the quadratic scaling of standard attention, enabling models to process entire codebases or document collections without chunking. For AI/ML development, this means future architectures will likely move toward sparse or linear attention, making RAG pipelines less necessary for many enterprise use cases. The implication is a shift in how we design reasoning agents and knowledge bases.
Government regulation is accelerating and can be opaque – The Anthropic model ban shows that U.S. export control laws (not just AI safety rules) can be used to forcibly remove AI products from the market with minimal due process. This creates uncertainty for AI labs: even after thorough safety testing, models can be abruptly shut down. For the industry, this trend demands proactive legal compliance and perhaps decentralization of model hosting. It also signals that geopolitical tensions will increasingly shape AI availability.
AI-powered social engineering is becoming a real threat – The LinkedIn backdoor story demonstrates how attackers are using AI agents (the developer used an AI tool to review code) as part of both the attack and defense. The recruiter specifically asked the victim to examine code, a common tactic that can now be weaponized with hidden payloads. This highlights the need for AI-aware security practices: treat all third-party code as hostile, use isolated environments, and employ AI for defensive code review. The trend is that AI both enables new attack vectors and provides better detection.
Ad-blocker extinction is a signal for browser-based AI conflicts – Chrome’s removal of Manifest V2 extensions is not directly about AI, but it parallels the ongoing battle between platform control and user autonomy. As AI agents become more common (e.g., browser-based assistants that rewrite content or block trackers), similar restrictions may emerge. The implication is that AI/ML applications relying on browser extension APIs may face platform-level limitations, pushing development toward native apps or alternative browsers.
AI model efficiency improvements are outpacing hardware gains – SubQ’s 56x speedup over FlashAttention-2 and local models running on 2022 hardware show that algorithmic advances (sparse attention, quantization, efficient architectures) are delivering larger performance leaps than raw hardware upgrades. For developers, this means optimizing model architecture and inference stack matters more than waiting for newer GPUs. It also suggests that edge deployment (local models) will become increasingly practical.
Trust and verification remain critical in AI supply chains – Several articles (backdoor, Anthropic ban, local model quality) underscore that trust in AI code, models, and data sources is fragile. The backdoor incident shows that even GitHub repos from recruiters can be malicious. The local model article implies that users must verify model outputs. The Anthropic ban shows that government trust can be revoked arbitrarily. A key takeaway is that the AI/ML community must invest in reproducible builds, provenance tracking, and independent audits to maintain reliability and safety.
Analysis generated by deepseek-reasoner