|
📖 Read In Depth
|
I believe there are entire companies right now under AI psychosis
Mitchell Hashimoto's viral thread on companies falling into 'AI psychosis' — outsourcing decision-making and thinking entirely to AI rather than using it as a tool. The HN comments are equally worth reading: one commenter predicts 'AI rescue consulting' will become a major industry as purely AI-written codebases hit complexity walls where defect rates outpace fixes. Direct relevance to someone who builds things from scratch to understand them.
hn/Best Stories
|
Elite researchers teamed up with Anthropic’s Mythos AI to smash Apple’s multi-billion dollar M5 security and build a kernel exploit in just 5 days.
Researchers used Anthropic's Mythos Preview to find the first public macOS kernel memory corruption exploit on Apple's M5 chip in just 5 days — cracking security Apple spent years and billions building. A second startup (Depthfirst) claims its own AI found additional critical vulnerabilities that Mythos missed at 1/10th the cost. This is a concrete data point on AI's real-world offensive security capability, not hype.
reddit/r/singularity
|
Bun Rust rewrite: "codebase fails basic miri checks, allows for UB in safe rust"
Bun's Rust rewrite is facing serious criticism: the codebase allegedly fails basic Miri checks and allows undefined behavior in nominally safe Rust. Combined with the HN thread on the rewrite being done in one week with detailed Zig-to-Rust mapping instructions, this raises real questions about AI-assisted rewrites trading speed for correctness. Exactly the kind of systems craft issue worth examining.
hn/Best Stories
|
Nous Research Releases Token Superposition Training to Speed Up LLM Pre-Training by Up to 2.5x Across 270M to 10B Parameter Models
Nous Research releases Token Superposition Training (TST), a pre-training efficiency method that speeds up LLM training by up to 2.5x across 270M–10B parameter models. Pre-training efficiency improvements at this scale have real economic implications for who can afford to train frontier models. Worth reading the arxiv paper linked.
reddit/r/singularity
|
We are retiring our bug bounty program
Turso is retiring their bug bounty program — the stated reason connects to AI-generated code flooding their codebase with complexity that makes security review untenable. The HN comments draw sharp parallels to 'tactical tornado' engineers who produce massive PRs that create more debt than value. A concrete case study in how AI tooling changes security economics.
hn/Best Stories
|
A 0-click exploit chain for the Pixel 10
Google Project Zero documents a 0-click exploit chain for the Pixel 10. The HN comment thread notes that AI-powered on-device features (like message search/understanding) are meaningfully expanding the 0-click attack surface — an underappreciated consequence of adding ML inference to the trusted computing base on mobile devices.
hn/Best Stories
|
Mullvad exit IPs are surprisingly identifying
Technical finding: Mullvad VPN's exit IPs can be used as a fingerprinting vector in ways that are surprisingly identifying, partially by design and partially not. The Mullvad co-founder engaged directly in the HN thread confirming some behavior is unintended and patches are being tested. Good systems-level thinking about where privacy assumptions break down.
hn/Best Stories
|
|
âš¡ FYI
|
Mistral AI founder to French Parliament: "Engineers at Mistral no longer write a single line of code
Mistral's founder told the French Parliament that engineers at Mistral no longer write a single line of code. He frames it as a shift away from craft toward something else entirely — a striking admission from a frontier lab founder, and a signal worth tracking for anyone thinking about where the ML engineering role is heading.
reddit/r/singularity
|
Nvidia C.E.O. Jensen Huang Hitches Ride With Trump to China After Last-Minute Invite
Jensen Huang joined Trump's China delegation on Air Force One for the Beijing summit, flying with 17+ CEOs including Tim Cook. The geopolitics of compute is playing out in real time — Nvidia's ability to sell into China is directly on the table in trade negotiations, which shapes everything from GPU supply chains to who trains frontier models.
nyt/Technology
|
Trump Uses Taiwan Arms Sales as Bargaining Chip With China, in a Risky Move
Trump explicitly called Taiwan arms sales a 'negotiating chip' with Beijing, which rattled allies and raises serious questions about U.S. commitments to Taiwan — directly relevant to TSMC's geopolitical risk profile and the stability of the semiconductor supply chain that underpins AI infrastructure.
nyt/Top Stories
|
OpenAI Considers Legal Action Against Apple in Strained Relationship
OpenAI is considering legal action against Apple over how ChatGPT has been integrated into Apple's ecosystem — apparently unhappy with the terms or execution. This is a classic platform power struggle: Apple controls the distribution layer and can dictate integration terms, which is a real moat question for AI companies dependent on mobile reach.
nyt/Technology
|
Bill to block publishers from killing online games advances in California
California legislation advancing that would require publishers to keep online games playable (or release server code) before shutting them down. The HN thread has a fascinating inside account of how legally complex open-sourcing a product actually is at a public company. Has broader implications for software-as-a-service and digital ownership debates.
hn/Best Stories
|
'Touch dreaming' helps humanoid robots handle five tricky tasks with 90.9% higher success
CMU and Bosch researchers developed 'touch dreaming' — an AI training method that dramatically improves humanoid robots' dexterous task performance (90.9% higher success on five tricky tasks). Tactile sensing has long been a bottleneck for manipulation; this is a meaningful step toward robots that can reliably handle real-world physical complexity.
reddit/r/singularity
|