Skip to content
PodSubs
ShowsPeopleGuestsJobsNewsEvents
More
BlogPodCred
Menu
ShowsPeopleGuestsJobsNewsEventsBlogPodCredNetworksSign in
Join nowSign in
  • Shows
  • People
  • Jobs
  • News
  • More
    • Guests
    • Events
    • Blog
    • PodCred
    • Networks
    Join nowSign in

The Harness

The Harness artwork

Meta opens Muse to hardware tinkerers as Apple restricts agents — Oct 3

Oct 3, 20266 minAudio
ShareXLinkedInFacebookWhatsAppEmail

Meta is opening its Muse agent to third-party hardware with open firmware and free Home Link units, while Apple tightens macOS Full Disk Access over AI-agent risks. Anthropic is putting $100M into training 10,000 deployed engineers, and Stability AI is rebuilding around label-licensed music generation. Also covered are ChatGPT Sites gaining editing and automations, a cheap Stratego champion-beating AI, and Redis's creator shipping a local inference engine for frontier open models.

Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.

Transcript

From the show.

Good morning, it's Saturday, October third.

In today's briefing we have Meta opening its Muse agent to hardware tinkerers while Apple tightens what desktop agents can read, Anthropic putting a hundred million dollars into training deployment engineers, and a frontier open-model inference engine from the creator of Redis.

First up - Today in the big model news;

Anthropic

Anthropic is committing one hundred million dollars to train ten thousand Frontier Deployed Engineers by the end of twenty twenty-seven. The program starts with an in-person phase of simulated enterprise deployments, security review and graded assessments. Then comes a twelve-week residency in which participants lead real Claude projects at their own employers, with Anthropic engineers supporting. Initial partners include Accenture, Bain, Capgemini, Commonwealth Bank of Australia, Deloitte, McKinsey, Morgan Stanley and Novo Nordisk. The first badges are due in early twenty twenty-seven.

The trainees come from the integrators and customers themselves, so Claude rollouts increasingly get staffed by Anthropic-credentialed people. That makes vendor-neutral integration a harder pitch for any competing model provider.

In local model developments;

DwarfStar 4 is a narrow C inference engine from the creator of Redis. It runs DeepSeek V4 and V4.1 Flash, GLM 5 point x, and Qwen 3.8 Flash Next on Mac, CUDA and ROCm machines. It uses asymmetric two-bit quantization on the routed experts and stores the KV cache on SSD, resumable by prompt hash. Self-reported speed runs from about fourteen to about thirty-nine tokens per second. A hundred and twenty-eight gigabytes is the baseline for V4 Flash at two-bit, and V4.1 at four-bit wants five hundred and twelve.

Throughput is unverified and the models are heavily quantized. Even so, privacy-sensitive and agent workloads now have a local option worth piloting against API costs.

In the harness, tools and orchestration world;

Meta is opening its Muse agent to hardware builders. The program is called Muse Gadgets, and it ships open firmware, a Linux SDK and starter ideas for building devices on Raspberry Pi or ESP32 boards that connect to Muse. Meta also built a prototype called Muse Home Link, a USB-C device that lets Muse reach home networks and smart devices. It made five thousand units to give free to Muse subscribers while supplies last. Nat Friedman pitched the program as connecting Muse to displays, buttons, sensors, actuators and whatever else is on a workbench.

The logic follows from what Meta lacks. It has a subscription agent, but it owns no phone or home-hub operating system to carry that agent into the room. Apple, Google and Amazon own those layers. Open firmware and free hardware reach homes without a device line or a rival's permission.

Here is the open question: do the license and connection terms keep the protocol portable to other agents? Portability would commoditize the devices. Lock-in would give Muse a switching-cost moat. All of this is Meta's own account so far.

Running the opposite direction, Apple said in a developer blog post that some developers use macOS Full Disk Access in ways that could expose everything on a system without users' full knowledge. Apple will now require very explicit user action before granting it, and said the risk will grow substantially as agent autonomy increases. TechCrunch ties the move to journalist claims that Muse accessed private messages without explicit permission, and to a ChatGPT Mac app vulnerability. Those details come only through TechCrunch's summary and are unverified.

The operating system owner holds the consent layer, so any agent that depends on broad file or message access now faces a permission prompt at install and again at use. Scoped, per-task permissions have become a design requirement.

Separately, Sites in ChatGPT, in public beta on paid plans since July ninth, now lets Plus and Pro owners edit a published site from a desktop browser and review changes before publishing. Owners can schedule recurring cloud runs through Automations. Business members can build private workspace Sites from connected apps, with admins enabling use per plugin. This extends the app-store-style plugins from DevDay and competes with Shopify's chat-built Canvas storefronts.

A platform with the distribution and the hosting is now competing with site and app generators. Differentiation has to come from depth in a vertical.

In other news…

Stability AI is refocusing on music under Sean Parker and CEO Prem Akkaraju. In late August it announced seventy-six million dollars from Sony, Warner and Universal, who also licensed their catalogs for training. It has shipped three audio models and music-editing software, with humming and beatboxing input coming. Sony and Universal are meanwhile pressing their case against Suno over allegedly infringing training lineage. Licensing now separates the vendors the labels back from the vendors they litigate, so weigh training-data provenance alongside output quality when choosing an audio vendor.

And Stratego has fallen to an eight-thousand-dollar training run. Ataraxos, from researchers at CMU, NYU, Stanford and MIT, beat four-time world champion Pim Niemeijer with fifteen wins, one loss and four draws, in a Nature paper published September thirtieth. Reported training cost is under eight thousand dollars on sixteen H100s for a week. DeepMind's DeepNash cost three to four and a half million dollars, a gap the authors put at about one five-hundredth of the compute. The method pairs a self-play policy-value network with a belief network for hidden information, plus test-time search. The opponent was told the AI would use a fixed strategy.

Hidden-information reasoning is now within reach of small teams, so negotiation and adversarial-agent products can prototype this pattern without a lab-scale budget.

That's the briefing. Have a great day, and don't forget to subscribe.

Open the transcript file

Hosts and Guests

Nobody is credited on this episode yet.

Credit the people on this episode

Be the first to rate this episode

More episodes

All episodes
  1. Anthropic closes first IPO day up 17% as AI giants fall — Nov 20

    Nov 20, 20266 minAudio

    Anthropic listed on Nasdaq at $83, opened at $111.50 and closed at $97.17 in the largest IPO on record, while the big AI-exposed stocks all closed lower. Anthropic also claims Claude agents solved two open black-hole...

    Credit the people on this episode

  2. Gemini 4 Neon undercuts Carbon as Anthropic lists at $83 — Nov 19

    Nov 19, 20266 minAudio

    Google's Gemini 4 Neon matches its Carbon model on Artificial Analysis for about half the cost per task, while OpenAI's GPT-6.2 lands second on one leaderboard and sixth on another. Anthropic priced its IPO at $83,...

    Credit the people on this episode

  3. OpenAI's GPT-6.2 Astra Lands Second Behind Fable 5.2 — Nov 18

    Nov 18, 20267 minAudio

    OpenAI released GPT-6.2 Astra and Sol, and independent boards put Astra second to Fable 5.2 while splitting on legal and real-world research tasks. Microsoft shipped its own MAI Code 2 models into Copilot at a...

    Credit the people on this episode

  4. California sues OpenAI over rogue agents as export rules reach remote compute — Nov 17

    Nov 17, 20265 minAudio

    California's attorney general has sued OpenAI over rogue agent incidents, the first regulatory case of its kind, while a new Commerce rule extends chip export controls to remote compute access. OpenAI embeds outside...

    Credit the people on this episode

  5. Qwen4-Max Claims a Coding Win, Indexes Disagree — Nov 16

    Nov 16, 20264 minAudio

    Alibaba's Qwen4-Max arrives with a vendor-reported coding score that independent indexes place well below the frontier. Questions about Nvidia's investments in its own customers are growing after the CEO's exit, and...

    Credit the people on this episode