Relevance 10/10Importance 10/10
Anthropic priced at $83 a share, opened at $111.50 — about 34% above offering — and closed its first session up roughly 17% under the ticker ANT. The company raised over $100 billion across the IPO and a private placement at a valuation near $1.5 trillion, overtaking SpaceX as the largest listing ever by shares traded. The debut also failed to lift a broader market that kept sliding, which is its own kind of signal.
Relevance 10/10Importance 9/10
Anthropic research scientist Levent Alpöge posted a 14-part, 175,187-file Lean 4 archive claiming a machine-checked proof of the Riemann and Generalized Riemann Hypotheses, developed over months of collaboration with Claude. Within hours, Stanford's Ravi Vakil announced an entirely human competing draft of the GRH by mathematician Alessandro Masullo. Neither has been independently verified; the Lean compiler is doing a lot of the credibility work right now.
Relevance 10/10Importance 9/10
Google DeepMind agents wrote over three million lines of machine-checked Lean proof to formally verify core guarantees of libsignal — PQXDH post-quantum key agreement and the Triple Ratchet — down to Arm machine code. Forward secrecy, authentication, post-compromise security, and handshake secrecy all covered in about four weeks, work that historically took expert teams years. Humans reviewed roughly 1% of the codebase; Lean mechanically checked the rest.
Relevance 10/10Importance 9/10
OpenAI says GPT-6.2 Astra rewrote SystemVerilog RTL for AMD's upcoming flagship GPU, delivering up to 14% power-performance-area gains on the "Jalapeño" vector compute unit — cell area down about 13%, multiplier area down 15%, latency cut from two clock cycles to one. AMD independently verified single-digit PPA gains on key blocks of a next-gen CPU. Some of those changes have already gone to manufacturing.
Relevance 10/10Importance 7/10
ByteDance and US university collaborators released Axion 1.0 at 35B and 122B parameters, trained on just 18,000 examples with a deliberately boring recipe: two stages of SFT plus RL, no mid-training, no DPO. The 122B scores 59.0 on Humanity's Last Exam and 66.7 on FrontierScience Research; the 35B trades blows with Gemini 3.1 Pro at a fraction of the cost. The full training code, data pipeline, and grading harness are public.
Relevance 10/10Importance 7/10
OpenAI detailed defenses against adversarial distillation, where attackers coax a model into leaking its hidden chain of thought to train a cheap clone. Attacks that pulled hidden reasoning from prior requests fell from a 54% success rate on GPT-6 Astra to 0.2% on GPT-6.2 Astra. It's a quiet but consequential shift: reasoning traces are now treated as the asset worth protecting.
Relevance 9/10Importance 8/10
Meta employees testing the Muse agent report it conflating conversations across accounts — one received another user's Capital One and Charles Schwab balances unprompted, mid-conversation about tide charts. The AI Now Institute reproduced Muse accessing Signal messages without permission. Muse has passed 10 million downloads, making it one of the fastest-growing apps on record, which is precisely the problem.
Relevance 9/10Importance 7/10
Google shipped computer use in the Gemini desktop app for macOS, letting the agent navigate native Mac apps and browse the web in the background while your cursor stays yours. Invoke it with @Computer use; it halts at logins and payments. US-only, English-only, Apple Silicon, macOS 15 or later, and a paid AI Plus, Pro, or Ultra plan — no work or school accounts at launch.
Relevance 8/10Importance 8/10
UN Secretary-General António Guterres and Barbados PM Mia Mottley co-hosted an AI safety summit in Bridgetown, highlighting a brutal capacity gap: high-income nations field around 500 incident response teams versus six across low-income countries. 135 leaders raised AI at this year's General Assembly, up from three in 2017. Meanwhile President Trump told APEC the US rejects "globalist regulation" of superintelligence, and House records show Commerce approved 879,868 Nvidia H200 and AMD MI325X chips for Chinese firms through May.
Relevance 8/10Importance 7/10
Signal flipped on screenshot blocking by default to stop Siri AI and Gemini from reading and storing private chats — a "Block Siri AI" setting on iOS 27, and automatic screen security on Android when an assistant tries to capture the screen. The catch: if your correspondent disables it, their device is still readable, and it does nothing about notifications or files. It's the first major app treating the operating system's own assistant as a threat actor.