Claude solves 37-year math problem, Kinney Drugs' AI can't handle a refill

Claude advances a key Riemann zeta bound by 25.6 points, more than 30x the progress made in the prior 37 years: An unreleased research version of Claude raised the proven lower bound of Riemann zeta zeros lying on the critical line from 41.6% to 67.2%, a 25.6-percentage-point jump versus 0.8 points gained by mathematicians from 1988 to 2025. Two Anthropic mathematicians validated the result, and number theorists Brian Conrey and Dan Goldston reviewed it independently. Claude also produced a formally verifiable Lean proof. One important caveat: this metric cannot reach 100% and still leave the Riemann Hypothesis unproven, because a measure-zero set of counterexamples would not show up in the bound.

Mistral AI granted US patent 12,670,045 for "Code implemented tool calls": The fast-tracked patent, filed March 2026 and granted in June, covers a method where an LLM generates a code block to encapsulate tool calls, executes it in a sandbox, pauses when a tool call is pending, sends that call to the client, and resumes execution once the result is returned. The technique is central to how modern AI agents chain actions, making the claim broad enough to draw concern.

Kinney Drugs pulls back AI phone assistant after hundreds of customer complaints: The Vermont and New York pharmacy chain deployed an AI phone assistant called "Burt" in May 2026, built by AI company Synerio and named after Kinney's founder, to handle prescription refill requests and inquiries. Customers reported medication delays, incorrect refill information, and privacy concerns. Lawyers told VTDigger that healthcare AI is outpacing regulation, leaving consumers vulnerable. A recurring frustration in healthcare AI deployments: the bot handles only simple cases, while patients with real problems burn time trying to reach a human.

Fully autonomous robots are much closer than you think, Sergey Levine: UC Berkeley professor and Physical Intelligence co-founder Sergey Levine joins the Dwarkesh Podcast to argue robots will run households entirely autonomously by 2030, driven by a "self-improvement flywheel" from deployment data. The conversation covers how vision-language-action (VLA) models work, why robotics will scale faster than self-driving cars, simulation-based learning, and whether China's hardware advantage is decisive.

Why most LLMs fail the car wash problem: "It's 100 meters away. Should I walk or drive?": The "car wash problem" is a viral one-sentence reasoning test: if the car wash is 100 meters away and you want to wash your car, the only sensible answer is "drive" (so the car arrives). Most LLMs say "walk." A benchmark by Opper AI found 42 out of 53 leading models got it wrong, exposing a systematic failure to infer implicit physical constraints. A separate arXiv study found that framing the same prompt with a STAR reasoning structure pushed accuracy from 0% to 85%.

Anthropic engineer Daisy Holman's 45-minute talk on building Claude agents that remember, self-correct, and improve over time: Holman, an engineer on the Claude Code team, argues the key shift is from hand-crafting prompts to designing systems where Claude writes its own harness and runs subagents in parallel. The talk shows concretely how to build agents with persistent memory and self-fixing loops.

NVIDIA CEO Jensen Huang says nobody writes prompts anymore, the job is now building loops and graphs: In a roughly 50-minute episode of the "Memos to the President" podcast, Jensen Huang argues that agentic AI has made prompt writing a secondary skill. The real leverage now comes from designing systems where models plan, use tools, evaluate outputs, and iterate, shifting the developer role from crafting individual instructions to orchestrating self-improving workflows.

DroidRun open-sources Mobile Harness, giving AI agents a unified API to control real iOS and Android phones: The open-source Mobile Harness repo, built on mobilerun-core, gives coding agents (Claude Code, Codex, or any shell-capable runtime) a single unified Mobilerun API across local and cloud mobile devices. A Markdown-first instruction set routes each task to the right platform guide, app card, or recovery playbook, keeping the agent's context small and its control path consistent across Android and iOS. DroidRun co-founder Niels Schmidt announced the launch on X, noting agents get one consistent control path across local and cloud mobile devices.

DAPI by Diffusion Studio: a video editing CLI built for coding agents: Diffusion Studio founder Konstantin Paulus launched DAPI, an open-source, cross-platform command-line interface that lets AI coding agents (like Claude) generate motion graphics and edit footage through natural-language prompts. The launch demo video was created entirely by prompting, with no manual editing. The underlying editor is available on GitHub and runs in the browser, on macOS, or from the terminal.

Jcode launches as a high-performance, open-source terminal coding agent claiming 20x more memory efficiency than Claude Code: Jeremy Huang, a 21-year-old YC-backed founder, built Jcode to be 20x more memory-efficient than Claude Code, so the RAM bottleneck that limits how many agents you can run at once is drastically reduced. It is MIT-licensed and installable with a single curl command from jcode.​sh.

Meta open-sources Muse Glimmer, a 30B agentic model built to run locally on consumer hardware: Meta's Superintelligence Labs released Muse Glimmer 30B under an Apache 2.0 license, optimized for always-on local agent workflows on a Mac or PC with a single consumer GPU. Developer Francesco Bonacci put it to the test, running it fully offline on Apple Silicon via Cua Driver to autonomously operate Notes and Reminders on macOS, calling it the first local model he'd recommend for a "fire-and-forget" agent workflow that keeps running until the job is done. Latency on consumer hardware still needs work, but the demo shows offline computer-use agents becoming practical.

remote-agent-browser 1.2.0 adds session reconnection, shell commands, and a public Docker image: Vercel Labs' agent-friendly cloud browser SDK ships three new capabilities: AgentBrowser.​session({id}) reconnects to an existing browser session (useful for long-running or resumed agent workflows), browser.​shell(cmd) runs arbitrary shell commands inside the sandbox, and browser.​readFile() pulls out generated artifacts. The Docker image is now public and authentication requires only an OIDC token, which Vercel provides automatically in deployed projects.

Google launches an Agent Quality Flywheel skill that lets your coding agent drive its own eval loop: The Google Developers Blog post walks through a three-phase Build, Ship, and Refine methodology backed by AutoRaters, and ships it as an installable coding-agent skill so the loop runs automatically. Google Cloud Tech distills the underlying design into seven rules: start with one eval case, make judges explain verdicts, use deterministic code checks where possible, score behavior not execution paths, treat flaky scores as findings, keep the evaluator outside the loop so the agent cannot game its own metric, and run auto-optimization only at the very end.

Needle 2: 14MB agentic language model for phones, wearables, and microcontrollers: Cactus Compute's Needle 2 is a 45M-parameter tool-calling model compressed to 2-bit (CQ2), fitting in a single 14MB binary that runs a full session in just 28MB of RAM. It targets the long tail of IoT hardware, hitting 500 tokens/sec on a Raspberry Pi 5 and 300, 700 tokens/sec on sub-$200 Android phones, and benchmarks competitively against FunctionGemma 270M and Apple's Foundation Model at 5x to 70x smaller. A live WebAssembly sandbox lets you test it in the browser. Early testers note edge cases in prompt interpretation, and fine-tuning is likely needed for production use cases.

H3-metal, Native MiniMax-H3 inference for Apple Silicon: Redis creator Salvatore Sanfilippo built h3.​c, a native C and Metal inference engine for MiniMax H3, a video-generation diffusion model, targeting Apple Silicon Macs. Existing workarounds via ComfyUI with GGUF quantization are painfully slow (roughly an hour for a 9-second clip on an M-series MacBook), and this project aims to close that gap with direct GPU acceleration through Metal.

Ante is a coding agent in a single binary that runs offline: Ante, from Antigma Labs, is a self-contained Rust binary that works like Claude Code or Codex with no runtime dependencies and no model lock-in. Typing /offline-mode installs a local inference engine so it runs entirely air-gapped. The company claims it uses ~7x less peak memory and ~9x less CPU than Claude Code at equivalent accuracy, making it suited for running many parallel agents. Two caveats worth noting: the GitHub repo ships binaries without source code, and telemetry is opt-out rather than opt-in (disable it by setting ANTE_TELEMETRY=off).

Stoa Markets (YC S26) launches a structured marketplace for buying and selling GPUs and AI servers: GPU hardware today trades through phone calls, forwarded spreadsheets, and private broker quotes with no shared price discovery, meaning two sellers can quote $200k and $240k for the same server node. Stoa brings the whole process online: buyers submit RFQs, get matched with verified counterparties, and see live clearing prices for A100s through GB200 NVL systems. Payment is held via Stripe Connect's manual-payout feature until the buyer inspects and accepts the hardware, functioning like an escrow flow without charging the typical ~7% escrow fee.

Anthropic's Claude to embed invisible text watermarks, GPTZero CTO explains how the KGW method works and why it's fragile: Anthropic's support page confirms Claude models will embed invisible watermarks in generated text, following the KGW method: a pseudorandom "green list" of tokens is selected based on prior context, and generation is softly biased toward those tokens, leaving a statistically detectable signal. GPTZero co-founder and CTO Alex Cui explains that the same technique underlies watermarking efforts at Google and OpenAI, and that it can be defeated by paraphrasing, sampling at high temperature, or passing output through a model that doesn't share the same key. The foundational 2023 paper that sparked the approach is freely available on arXiv.

Anthropic's Cat Wu demos her full Claude Code setup live, as teams shift from self-improving loops to agentic graphs: Cat Wu, Head of Product for Claude Code at Anthropic, builds her entire Claude Code configuration from a blank terminal in roughly ten minutes. The talk captures an internal shift: engineers who had been building self-improving feedback loops are now structuring work as agentic graphs, where retry logic and orchestration move up to the workflow layer rather than the prompt. The companion guide on graph engineering with Claude Code covers the architecture behind the pattern.

Composio tests DeepSeek V4 Flash across 4 agent harnesses on 30 tasks, Pi Agent wins on cost and pass rate: Composio, an AI agent tooling platform, ran DeepSeek V4 Flash through Pi Agent, Deep Agents (LangChain), Hermes Agent, and Prime Agent on 30 non-coding agentic tasks. Pi Agent passed 20 of 30 tasks and was the cheapest harness, beating Deep Agents (16/30), Hermes (15/30), and Prime Agent (15/24 valid runs). The takeaway: the choice of harness wrapper may matter as much as model choice for agentic workloads.

Exploring Claude/GPT Knowledge Cutoffs and Pre-Training Timelines: Researcher Shrivu Shankar probes frontier models like GPT-5 and Claude Opus with carefully curated questions to reverse-engineer hidden training details, including approximate parameter counts (via "Incompressible Knowledge Probes"), dataset mixtures (via "Data Mixture Inference"), and pre-training timelines, without access to any official disclosures.

Prompting is dead once agents start making the decisions for you, argues senior software engineer Jey: Jey's argument: when an agent acts autonomously, vague prompts like "add authentication" hide a dozen decisions the model will answer without you. The fix is writing a spec before you ship: why, what, constraints, out-of-scope, tasks. He points to a free 99-minute NeurIPS tutorial by DeepLearning.​AI's Andrew Ng and OpenAI's Isa Fulford, where Fulford's line captures it plainly: "the model can't read your mind."

Can Agents Use a Computer Yet? We've Got the Data: A new a16z analysis by partners Fabrizio Serafini, Seema Amble, and Eric Zhou finds that computer-use agents now cost roughly $6-8 per hour to run, undercutting offshore outsourced labor at ~$10/hr and US talent at $30-45/hr. The report covers real production deployments on narrow, repeatable workflows such as updating records and moving data, and notes the math keeps improving as inference costs fall and open-source models improve. Critics in the discussion flag that cost-per-hour and cost-per-completed-task are different numbers, and that reliability remains an open question.

Mark Zuckerberg attacks 'closed' AI rivals as Meta returns to open models: In a public essay titled "The Future is for Everyone," Zuckerberg argues that superintelligence should be open and distributed rather than controlled by a few institutions, framing closed AI rivals as a threat to individual empowerment. The move accompanies Meta's renewed push toward open-weight models. Some readers are skeptical Meta's motives are purely altruistic, noting the company's history with open-source as a competitive strategy rather than a philosophical commitment.

Google Search Is Dying. What Comes Next Is Worse: Writer Vass Bednar argues in The Walrus that AI is hollowing out the open web: as bots scrape and summarize content, the original pages that formed the internet's collective memory become harder to find, sustain, or even reach.

Humanising LLM Outputs Is Dumb: Blogger Kuber Mehta argues that prompting LLMs to sound warmer and more human degrades output quality: the results become verbose, harder to parse programmatically, and lose precision without gaining clarity. The case is that plain, dense LLM prose is often more useful than the performative friendliness injected by system prompts.

What's the best programming language for coding agents?: Dan Luu runs careful evals challenging the popular claim that concise dynamic languages (Ruby, Clojure, J) are best for LLM coding agents because of lower token costs. He finds most of those claims don't hold up, that the real performance drivers are less obvious, and that the question is harder to answer than widely shared benchmarks suggest. Practitioners in the discussion note practical wins from static analysis tooling: adding type-checking hooks (e.​g. Pyright on every edit) and auto-formatting (e.​g. clang-format) can meaningfully reduce agent errors and wasted token loops, regardless of language choice.