TabLOID

Meta’s Muse for Small Business wants to help the owner doing six jobs because there are only eight hours in the day. The announcement says it connects tools including Shopify, QuickBooks, Stripe, Klaviyo, Canva and Slack, plus Facebook and Instagram business accounts. Nothing publishes, sends or spends without approval, according to Meta.

The company’s examples include finding useful sales patterns, drafting content and deciding what deserves attention. It says Muse is available in the US and Canada, with free access for most needs and paid plans for heavier use. Custom connectors are supported.

Why it matters: the fight is moving from who answers questions to who holds your business context. Start with an approved draft that saves work before giving a robot the register. The approval promise is worth testing, not treating as a guarantee.

Sources: The Neuron, The Rundown AI and Headlines by Abe Burns, Sep 30; Meta's Sep 29 announcement. Go deeper: Meta's small-business announcement - OpenAI's competing Dots approach

Anthropic says Z.ai’s open-weight GLM-5.3 approaches Claude Mythos Preview on some cyber-exploit tests. The important difference is availability: a restricted model and a downloadable model can have similar skills but different exposure.

On a V8 benchmark, Anthropic reports 50 successful attempts out of 410 for GLM-5.3, versus 56 for Mythos Preview. Its separate binary-exploitation test reports full control-flow hijacks in 4% and 6% of trials, respectively. These were isolated targets, not real-world victims.

Why it matters: capabilities do not stay scarce forever. The same tools can help defenders find and fix flaws. This is Anthropic’s evaluation of a competitor, and the lab conditions matter. It is a reason to keep software updated, not a prediction that every website is now vulnerable.

Sources: The Code, Sep 30; Anthropic research, Sep 29. Go deeper: Anthropic's full evaluation and methods - Anthropic's research index

ElevenLabs says v4 Turbo reaches first speech in a median of roughly 150 milliseconds, with streaming output starting before the language model finishes a sentence. That tackles dead air. Hume’s research asks the other half of the question: did the agent hear how the human said it?

ElevenLabs lists more than 90 languages and Professional Voice Clones for v4 and Turbo. Its latency is a vendor measurement, not a guarantee for an entire call. Hume’s Real World VoiceEQ framework evaluates recognition, understanding, speech quality and live agent performance, among other dimensions.

Why it matters: sounding natural, conveying emotion and staying reliable are different skills. Hume reports that access to audio does not guarantee an agent uses vocal emotion. Test the noisy, interrupted, ten-minute call, not just the polished first sentence.

Sources: The Code and The Neuron podcast email, Sep 30; ElevenLabs and Hume. Go deeper: ElevenLabs v4 and Turbo - ElevenLabs API documentation - Hume's real-world evaluation framework

Dyna’s Dyna-2.1 demonstration follows Taku through an uncut hour of hotel laundry. Not one cinematic fold. A sequence of moving, washing, drying, folding and recovering. The glamorous future contains a lot of towels.

The company describes Taku as a semi-humanoid robot with two arms and a wheeled base. Its agent coordinates changing tasks and remembers the workflow as machines finish at different times. Small failures matter when a job has hundreds of steps.

Why it matters: the useful unit is a workflow completed without rescue. A longer demo is more informative than a highlight reel, but it is still the vendor’s demonstration. Dyna says its next milestone is testing this system at customer sites.

Sources: The Neuron, Sep 30; Dyna's Sep 29 technical post. Go deeper: Dyna-2.1 technical post and uncut demo - Dyna's product and company overview