AI

𝜇⁢𝜆⁢ϵ⁢𝛿-Calculus: Self Optimizing Language that Seems to Exhibit Paradoxical Transfinite Cognitive Capabilities (arxiv.org) AI

The arXiv preprint proposes a “μλϵδ-calculus” based on contracting directed multigraph rewriting, claiming it extends and optimizes the lambda calculus in a way that always terminates and can produce self-similar “fractal” graphs resembling the original program. It argues for using a Wittgenstein-style paraconsistent logic to enable paradoxical self-referencing without logical explosion, and suggests two related extensions (ϵ-expressions for macro-like expansion and δ-functionals for minimal input/output modeling), with step-by-step interpreter/compiler/optimizer construction and proof-of-concept code.

Hermes Agent – Open-Source AI Agent with Persistent Memory (hermes-agent.org) AI

Hermes Agent is an open-source, MIT-licensed AI agent from Nous Research that runs self-hosted on your machine and uses persistent memory to retain context across sessions while automatically creating reusable “skills.” It supports a multi-platform gateway for messaging (and CLI), scheduled automations, parallel sub-agents, and browser/web automation with features like vision analysis and text-to-speech, with all data stored locally and no telemetry.

The Third Generation of Apple's Foundation Models (machinelearning.apple.com) AI

Apple says its third generation of Apple Foundation Models—built with Google and spanning two on-device models (AFM 3 Core and AFM 3 Core Advanced) and three server models running on Private Cloud Compute (AFM 3 Cloud, ADM 3 Cloud for image tools, and AFM 3 Cloud Pro)—is powering next-generation Apple Intelligence with privacy protections, new multimodal capabilities, and improved text/image performance versus 2025 baselines.

FP8 Is All You Need (Part 1): Debunking Hardware FP64 as the HPC Holy Grail (arxiv.org) AI

A paper by Satoshi Matsuoka argues that the HPC field’s “native FP64 silicon” is not the necessary foundation for scientific computing, claiming that on newer AI-optimized GPUs (e.g., NVIDIA B300/Blackwell Ultra) FP8 throughput combined with an Ozaki Scheme II reconstruction approach can recover full FP64 accuracy while achieving memory-roof execution for common HPC kernels.

Expanding Private Cloud Compute (security.apple.com) AI

Apple says it is expanding its Private Cloud Compute (PCC) beyond its own data centers by collaborating with Google and NVIDIA to run new Apple Intelligence workloads on Google Cloud while preserving PCC’s security and privacy commitments. The company describes using NVIDIA Confidential Computing with GPUs, Intel CPUs with TDX, and Google’s Titan chip, along with layered protections such as attestations, an append-only hardware ledger to address supply-chain risks, and transparency measures for external researchers. PCC on Google Cloud will be rolled out gradually during a summer preview period, with public binaries, research tooling, and access to PCC nodes via the Apple Security Bounty Program.

Meta Deletes Face-Recognition System from Smart Glasses App After Wired Report (wired.com) AI

Meta removed face-recognition code associated with its internally codenamed “NameTag” feature from its Meta AI app for smart glasses days after WIRED reported it had been embedded in the software for millions of phones, according to an analysis of the latest version’s code.

Software Design in the Age of AI (self-service.mirdin.com) AI

An Arch-Engineer post argues that while AI coding can automate low-level implementation, software design skills become more valuable because human engineers must still set intent, prevent issues like hidden coupling, and lead higher-level refactoring and design decisions—illustrated through the author’s Command Center AI coding environment launch and its refactoring agents.

The Data Center Bankshot – David Karpf (davekarpf.beehiiv.com) AI

The piece argues that opposition to data center siting is a strategic “bankshot” against the AI industry’s momentum and speculative finance bubble, potentially slowing deployment and forcing a reset of how AI is built, regulated, and debated.

Show HN: Command Center, the AI coding env for people who care about quality (cc.dev) AI

Show HN introduces Command Center (cc.dev), an “agentic” AI coding environment aimed at turning AI-generated code into production-quality changes by addressing common issues like sprawling diffs, broken edits, and maintainability concerns through guided walkthroughs and refactoring feedback. The page claims it can help teams read and review large code changes more easily, while noting privacy options such as running locally and not retaining code when using free Gemini credits.

AI and Agency (bitsandletters.com) AI

The piece argues that successful AI adoption depends less on tools and more on team “agency” practices—specifically clarity, psychological safety, and focusing on outcomes—contrasting this with top-down, metric-driven mandates that can undermine motivation and lead to low-value work.

Apple bets cheaper AI will woo small developers (techcrunch.com) AI

Apple says it will let developers with fewer than 2 million first-time App Store downloads use its Foundation Models in Private Cloud Compute without cloud API costs, aiming to lower the barrier for smaller developers to experiment with AI. The company also plans to expand the framework to include image input and to support server models, integrating with developers’ preferred cloud providers.

OpenAI plots biggest ChatGPT overhaul since launch (ft.com) AI

The Financial Times reports that OpenAI is planning the largest ChatGPT update since its launch, though details of the overhaul were not available in the provided data.

The sample efficiency black hole (dwarkesh.com) AI

Dwarkesh Patel argues that AI progress is largely driven by scaling data and compute—especially via reinforcement-learning-style synthetic data and large volumes of human expert labeled trajectories—while improvements in “sample efficiency” (how little data is needed to learn) may be limited. He compares human and frontier AI data exposure, suggests humans may be on a fundamentally different sample-efficiency scaling curve, and discusses why labs still may succeed in automating many white-collar tasks despite this inefficiency.

OpenAI Files S-1 (twitter.com) AI

The post titled “OpenAI Files S-1” shares that OpenAI has filed an S-1, though no article text is available to confirm further details.

OpenAI Submits S-1 Draft to SEC (openai.com) AI

OpenAI says it has submitted a confidential draft S-1 filing to the U.S. Securities and Exchange Commission, according to the article title and link provided.

Apple Passwords Now Auto Fixes Weak and Compromised Passwords with Agentic AI (macrumors.com) AI

MacRumors reports that Apple has added an “agentic” feature to the Passwords app, using Apple Intelligence and Safari to automatically update weak or compromised passwords in the background, with only an initial tap required and a Live Activity shown while it runs.

OpenAI Confidentially Files for IPO (cnbc.com) AI

OpenAI has confidentially filed for an IPO with the SEC, positioning the company for a potential major public debut while preparing Wall Street for a broader “AI debut” alongside rivals like Anthropic and SpaceX.

FrontierCode (cognition.ai) AI

Cognition introduces FrontierCode, a new coding benchmark intended to measure whether LLM-written code would be “mergeable” into real production repositories, going beyond functional correctness to assess correctness, test quality, scope restraint, style, and adherence to repo standards. The evaluation uses repo maintainers’ real-world criteria and automated grading methods (including code-mergeability blockers, score rubrics, and techniques designed to reduce false positives/negatives), and reports that even top models struggle—e.g., best results on the hardest subset remain low in percentage scores.

Built to benefit everyone: our plan (openai.com) AI

OpenAI’s Sam Altman and Jakub Pachocki argue that, like electrification, AI’s benefits will depend on how power and access are distributed, and that AGI should be available broadly while staying safe and human-controlled.

Apple Core AI Framework (developer.apple.com) AI

Apple’s Core AI framework documentation is available for developers, but the fetched page content could not be retrieved because it requires JavaScript to view.

Apple WWDC 2026: The 7 biggest announcements (theverge.com) AI

Apple’s WWDC 2026 keynote highlighted an AI-upgraded Siri (“Siri AI”) with conversation and on-screen context features, alongside iOS 27, macOS 27, and Safari updates that expand Apple Intelligence across devices, plus improvements to Apple Home (including 4K support) and redesigned Screen Time and parental controls.

Apple reveals new AI architecture built around Google Gemini models (macrumors.com) AI

Apple announced an overhaul of Apple Intelligence, building a new architecture around foundation models co-developed with Google using Gemini-family technologies and running on-device plus via Private Cloud Compute. The update adds multimodal capabilities such as image understanding and generation, with device-specific higher-power versions that include speech generation and improved dictation, coordinated by a new “system orchestrator” for app- and task-aware responses while reiterating privacy protections.

We need to learn how to argue with AI (ft.com) AI

An opinion piece argues that people need to learn how to effectively argue with AI systems, suggesting communication skills and reasoning practices are important when interacting with AI.

How Confident Are AI Classifiers About Their Own Confidence? (gmcirco.github.io) AI

The post tests how reliable AI “confidence” scores are when an LLM classifies injury body parts from NEISS medical narratives, comparing LLM-emitted confidence values to token log-probabilities from the model output. Using a sample of 500 cases with a gpt-5-nano extraction pipeline, the author finds that LLM confidence is relatively close to observed accuracy at the highest confidence ranges but diverges outside the upper end, while token log-probabilities are generally more over-confident. The article also outlines methods to calibrate probabilities in multi-class settings, including “top-vs-all” calibration via isotonic regression.

Two Leaps to 1000 Tokens/s on a 1T-Parameter Model (tilert.ai) AI

TileRT argues that reaching 1000+ tokens per second on a large (up to 1T-parameter) model requires a shift from kernel/operator-level tuning to a persistent, continuously running execution engine that removes microsecond “execution gaps,” plus hardware–model co-design to eliminate microsecond-scale overheads in components like RMSNorm, RoPE, KV-cache writes, and multi-token prediction.