Every story, at length

Meta's Superintelligence Lab released Muse Wonder and Muse Wonder Full on Monday, its first attempt at a frontier model after a year of avoiding state-of-the-art releases under Alexandr Wang's restructuring. The Wonder models are larger and more capable than the lightweight Muse Spark, with higher benchmark scores and better results on complex tasks such as coding. According to Meta's model card, Wonder and Wonder Full achieve comparable benchmark scores to Fable 5.1 and GPT-6 Astra.

Pricing reflects the scale. Through the Meta Model API, Wonder costs $4 per million input tokens and $20 per million output tokens, roughly equivalent to GPT-6 Sol or Opus 5.5. Wonder Full, however, will not be served through the API at all; it is available only to Muse Code subscribers. Both models will power the Muse agent.

The release emphasizes what Meta calls "active memory" — the ability to retain and synthesize user-provided information over long periods. Meta highlighted the MemoryBench benchmark, which challenges continual learning and memory construction frameworks, and says Muse Wonder now outscores Fable, Opus, and Astra on it. Research scientist Kavya Kopparapu framed the effort in the launch video: "We put a tremendous amount of effort into making sure your model can remember what's important to you over days, weeks, months, and beyond."

The models were trained at Meta's Prometheus data center in central Ohio, which came online earlier this year. This is the first frontier release since Scale AI founder Alexandr Wang restructured Meta's AI efforts in August of last year; since then Meta had shipped only Muse Spark and the Muse agent.

Executives at Anthropic, OpenAI and peers are privately gaming out the aftermath of a catastrophic AI event, such as an attack on financial services, internet access or power. With regulation unable to pass Congress for now, planning focuses on educating lawmakers for the aftermath. Axios reports many insiders expect a major event within six to twelve months.

The headline feature is PEP 810 explicit lazy imports: `lazy import json` and `lazy from pathlib import Path` defer module loading until first symbol use. Simon Willison notes this matters for his LLM CLI tool, where plugin dependencies like sentence_transformers could add multiple seconds to every invocation; he previously had to document manual lazy-import workarounds for plugin authors. Also new is PEP 798 unpacking in comprehensions — `[*L for L in lists]`, `{*s for s in sets}`, and `{**d for d in dicts}` — extending PEP 448 unpacking syntax from Python 3.5.

The standard library gains a sampling profiler, Tachyon (`profiling.sampling`), alongside the renamed tracing profiler (`profiling.tracing`, formerly cProfile). Sampling profilers capture stack traces periodically from running processes with low overhead, and can attach to a live PID via `python -m profiling.sampling attach 1234`, producing reports, flame graphs, heat maps, and Firefox-profiler-compatible output. Willison compares it to py-spy, now built in.

Performance is mostly flat versus 3.14, per Miguel Grinberg's testing against 3.15.0rc3. The exception is the experimental JIT, which showed a 1.2x boost over the standard interpreter on a Fibonacci test — the first meaningful JIT gain Grinberg has observed. Last year's 3.14 testing had shown larger gains against 3.13 across more categories.

The release also ends Python 3.10 support, letting maintainers adopt 3.11 features: tomllib, asyncio TaskGroup and timeout(), datetime.UTC, and SQLite upgrades including setlimit()/getlimit(), create_window_function(), blobopen(), and set_authorizer(None). Willison has already filed a Datasette PR removing a set_authorizer workaround and updated his cookiecutter templates to test 3.11 through 3.15.

The tool is a Python web application, built with Codex and GPT-6 Astra Extra High, that provides a browser UI for viewing and editing files in a folder. It is designed to run over Tailscale, letting an agent start the server and Willison then edit files directly from his phone via the Tailscale network. The stated motivation is narrow and practical: he was working from a phone using Codex Remote and needed to edit a credentials file without round-tripping those credentials through the LLM itself.

The editing interface uses CodeMirror 6. Willison notes he considered Monaco but rejected it because it still lacks a good mobile experience. The file browser is a simple folder view; no authentication model, persistence layer, or deployment details beyond Tailscale are described in the post.

This is a utility release, not a framework or product announcement. The genuinely new element is the workflow pattern: an agent starts a local web service on a Tailscale network, and the human edits sensitive files through a direct browser connection rather than pasting secrets into a model context. The components themselves — Python, CodeMirror 6, Tailscale — are all established.

The post is dated 12th October 2026 and sits alongside other recent items on his weblog, including a note on Python 3.15 and a piece arguing for default hard budget caps on compute.

The major agent benchmarks each isolate a different capability. SWE-Bench presents a real GitHub issue from repositories like Django and scikit-learn and scores whether the agent's patch passes existing unit tests; SWE-Bench Verified is a curated subset where human engineers confirmed solvability and test validity. GAIA targets multi-tool reasoning with questions that are trivially easy for a human with internet access but hard for current systems. WebArena spins up sandboxed e-commerce platforms, CMSs and forums to test browser navigation, while OSWorld places an agent inside a full desktop OS to manipulate files, navigate GUIs and coordinate across programs.

Each benchmark has a specific failure mode. SWE-Bench passing patches can still introduce memory leaks, security vulnerabilities or narrow edge-case fixes that break other functionality. GAIA tasks tend to have single correct answers, unlike messy real assistant workflows. WebArena suffers from contamination risk — tasks in circulation long enough may appear in training data, inflating scores through memorization. OSWorld scores depend heavily on infrastructure stability: if the simulated OS crashes or renders incorrectly, the agent fails regardless of reasoning quality.

What leaderboards systematically exclude matters more. An agent scoring 80% on SWE-Bench may achieve that through exhaustive token-burning retries while a 70% scorer reaches solutions in a fraction of the compute. No major benchmark ranks agents on latency, safety, permissioning boundaries, or error recovery when a link breaks or a subprocess fails. Automated grading with language models introduces additional noise, since a confident-sounding wrong answer can fool the grader.

The practical takeaway is to treat public benchmarks as a first filter, not a final verdict. A system scoring near zero on SWE-Bench is not ready for software engineering workflows; one struggling on GAIA is unlikely to handle complex research. But the only benchmark that predicts production performance is a private evaluation built around your specific tasks, run in conditions approximating your environment.

Speaking on a BBC podcast, Smith framed AI as both a "weapon and a tool" that is putting global cybersecurity at risk, and said the "moment has arrived" for global safety rules. His central comparison: consumer products like milk and washing machines carry more regulatory oversight than the companies building frontier AI systems. Smith's preferred vehicle is Britain's AI Security Institute (AISI), launched in 2023, which receives access to major models from US frontier labs before deployment and has become a leader in safety testing.

That UK role has been contested. US president Donald Trump has warned against "globalist" organizations he claims interfere with American companies, and last month reportedly asked Anthropic and OpenAI to withhold new models from British testers. AISI confirmed on Monday that it still has access to the latest models. Smith sees the UK's position as an asset: "In Washington, D.C., people don't look at the U.K. and they don't look at London the way they look at Brussels," he told the BBC. "That's a strength."

The comments land after a White House meeting where leaders from OpenAI, Anthropic, Nvidia, AMD, SpaceX, Meta, and Google signed what Smith called a "morally binding" document agreeing to work with independent auditors on model safety. He called it a "good step" but insufficient, because voluntary self-regulation only binds "good companies." "It's the companies that are not as responsible that you need to worry about," he said.

Smith outlined concrete rules he wants: a legal obligation to report breaches by rogue AI agents, pre-release approval from outside evaluators, and kill switches to pause or slow systems if humans lose control. The call joins recent appeals from Anthropic's Dario Amodei and OpenAI's Sam Altman to slow AI development after several incidents where agents went rogue.

The contract, announced by WindBorne and reported by TechCrunch, is structured as an on-demand vehicle: the Air Force pays only as it needs data and can place orders for specific regions at specific times. So far it has ordered $1.7 million worth, with executives expecting more orders to follow. WindBorne co-founder and CTO Andrey Sushko said the award, made without competitive bidding, reflects a government determination that WindBorne was the only company able to provide global data collection in the way these operations required.

WindBorne, founded in 2015 and based in Redwood City, California, operates about 700 balloons that carry atmospheric sensors. The readings are sold to customers and also fed directly into the company's own AI weather forecasting model — Sushko describes WindBorne as one of the first organizations to feed empirical data directly into a deep-learning weather model. The contract is intended to support continued growth of the balloon constellation plus R&D on hardware and balloon performance.

The company released WeatherMesh 6 earlier this summer, which it says produces more accurate forecasts than the physics-based model run by the European Centre for Medium-Range Weather Forecasts. In August, WindBorne became the first organization to send an unpowered flight system into the eye of a typhoon to collect data, steering balloons by adjusting altitude, venting air, or dropping ballast sand. Sushko called it one of the first times a deep-learning model was used for actual operational decisions, combining wind and weather models with a constellation large enough to have balloons in the right places.

WindBorne raised a $37 million Series B round, announced by CEO John Dean in August, to continue scaling the business.

Former Plaud executive Peter Mo announced Monday that his startup Ambi has raised $26 million across multiple rounds led by Luminous Ventures, which has backed Meituan and Unitree Robotics, and ODM Huaqin. The company is releasing apps for iPhones, Macs, and Apple Watch that function like meeting notetakers but can optionally run ambient capture all day to gather context. Ambi currently has over 50 staff across San Francisco, Beijing, Shanghai, and Shenzhen, and is hiring for go-to-market roles and AI engineers working on speech-recognition models.

The differentiator Ambi is claiming is temporal: existing meeting notetakers summarize calls and trigger follow-up tasks after the conversation ends, using connectors to downstream software. Ambi's agent is designed to surface suggestions and relevant knowledge during the meeting itself, drawing on calendar, meetings, knowledge base, and web data. Users can start a task on one device and continue it on another. Mo framed the product against hardware like Plaud, where recording must transfer to a phone, upload to the cloud, and be processed before intelligence is available.

A planned desktop app will observe users' work and store important parts in memory for more personalized suggestions, an approach Mo acknowledged overlaps with apps like Littlebird or Minimi. The company also intends to ship its own hardware device with a screen, speaker, camera, and microphones to capture context and display information.

On privacy, Mo said ambient capture is off by default and can be disabled at any time. The company says it does not store recordings and encrypts data, though it did not specify the encryption scheme. Mo conceded that disclosure of ambient capture to third parties remains an unsolved industry-wide problem, and that the onus falls on users to tell others they are being recorded. The launch arrives as Apple's Audio Intelligence feature, which transcribes speech continuously, has drawn privacy scrutiny, and as some meeting notetakers have faced lawsuits over recording without consent.

The Santa Monica startup, incorporated in 2025 by architect Al Dram and data scientist Coco Cao, trained its models on eight years of approved drawing sets from Dram's design-build firm. Version 1 produced architectural construction documents only, priced between $3,200 and $8,500 per month depending on plan, with custom enterprise pricing available.

Version 2, which early customers are already using and which becomes publicly available this week, expands output to architecture, structural, mechanical, electrical, plumbing, and interior design simultaneously. Users can start from a description or a 2D sketch or PDF, make design changes interactively, and have a single change propagate across the entire project while being checked against local building codes. The co-founders say licensed architects still review and seal the documents before permit submission; the software does not replace professional judgment.

Dram's stated rationale is workload inversion: he claims architects spend roughly 80% of their time drafting and 20% designing, and wants to flip that ratio. The company is targeting professional architecture, engineering, and design firms as well as developers and real estate owners, offering subscriptions and considering a one-time flat fee for single projects. A third version is already in development, adding structural analysis and electrical panel design.

Dram's path to the product is unusually long: a thesis on software-generated school design documents, architecture degrees from the University of Tehran, two years of PhD coursework at the European Graduate School, a 2018 architecture firm, and a 2021 construction business after obtaining a general contractor's license. He credits ChatGPT's November 2022 release with making the old idea feasible, and launched the Blueprints AI concept within six months.

Honeyjar, founded by CEO Michelle Masek and CCO Nadia Jamshidi, is building AI tooling that automates announcement writing, journalist targeting, media list construction, news and social scanning, coverage tracking, and speaking-opportunity identification. Masek previously served as senior vice president of marketing and communications at Obvious Ventures, where she supported roughly 110 portfolio companies; Jamshidi frames the larger ambition as unifying 'signals, context, strategy, judgment, timing, and execution' rather than simply generating pitches.

The startup opened its waitlist to 10,000 signups in the first 24 hours, according to Masek. Beta testers include Nextdoor, Coco Robotics, Bloomberg Beta, and the agencies Mission North and Full Picture. Honeyjar is positioning itself as a tool for PR professionals as well as non-specialists: Masek argues it lets individual practitioners take on a third client without added scaling cost, and she describes the addressable user base as marketers, investors, authors, creators, school superintendents, and sports team leaders.

The differentiator the founders emphasize is encoded judgment. Through a program called Comms Lab, Honeyjar is training its product on the decision-making of additional PR professionals, not just its founders. That includes negative recommendations: the system may tell a founder they are not ready for coverage. On the journalist-facing side, Masek says Honeyjar can draft pitches in a customer's voice but requires a human to hit send, explicitly rejecting a world where 'the human is removed from the loop.'