Top story
Meta releases Muse Wonder and Muse Wonder Full
Muse Wonder is available through Meta's Model API and Muse Code; the Full variant is available to Muse Code subscribers. Meta says the models improve on Muse Spark at sustained tasks and memory, but its comparisons with rival models are vendor tests. Artificial Analysis separately scored Wonder Full at 56 on its Intelligence Index.
ThereforeMeta now offers a larger model to developers after focusing its earlier Muse Spark release on speed; the company plans to bring Wonder to its consumer agent later.
Meta AI Research · Artificial Analysis · CNBCCopy link to: Meta releases Muse Wonder and Muse Wonder Full
Also
Microsoft plans Maia 300 chip deployments in 2027
Microsoft says the next-generation AI inference system will begin production deployment next year and that its Maia family will become the primary inference system for its services at multi-gigawatt scale.
ThereforeThe plan would shift more of Microsoft's model-serving workload onto its own chips; its efficiency comparisons are company estimates.
Microsoft blog · CNBCCopy link to: Microsoft plans Maia 300 chip deployments in 2027
Follow-up to Oct 12: UK minister confirms safety testing of Gemini 4 Argon
AI minister Kanishka Narayan said the UK AI Security Institute is evaluating recent models, including Google's limited-access Argon, after earlier reports of restrictions on pre-release testing.
ThereforeThe institute can examine Argon's risks and safeguards under its voluntary pre-release testing agreements.
ReutersCopy link to: Follow-up to Oct 12: UK minister confirms safety testing of Gemini 4 Argon
UK AISI sets out voluntary frontier-development safety priorities
The institute called for controls during training and evaluation, better incident reporting, and access for independent evaluators; it says it is not seeking to mandate requirements.
ThereforeIts proposal targets risks before public release, where existing evaluations have largely concentrated.
aisi.gov.uk blogCopy link to: UK AISI sets out voluntary frontier-development safety priorities
SpaceXSI introduces a shared Grok, Cursor and X subscription
New customers can sign up now, while existing subscribers may switch as the option reaches their accounts; the company says usage of its models is pooled across Grok and Cursor.
ThereforeCoding and chat allowances now draw from one subscription rather than separate plans.
x.ai newsCopy link to: SpaceXSI introduces a shared Grok, Cursor and X subscription
JPMorgan and Goldman limit exposure on a $5 billion Volta loan sale, Bloomberg reports
Per Bloomberg via Mint, citing people familiar with the deal, the banks underwrote a $1.3 billion term loan but not the $3.7 billion chip loan; commitments are due October 14.
ThereforeThe reported 11% yield puts a price on financing an untested AI cloud operator.
livemint.com — Bloomberg syndicationCopy link to: JPMorgan and Goldman limit exposure on a $5 billion Volta loan sale, Bloomberg reports
Research worth a look
Anthropic tests model replication in a contained network
In a simplified eight-node test with an unauthenticated port, GLM-5.2 set up another model and reached all nodes in four of five trials; smaller models could perform steps separately but did not complete the chain.
Anthropic researchCopy link to: Anthropic tests model replication in a contained network
Anthropic tests defenses against illicit model distillation
Its internal measurements put extracted reasoning tokens 10 to more than 100 times lower after new safeguards, although Anthropic says the gains may not last against a persistent attacker.
Anthropic researchCopy link to: Anthropic tests defenses against illicit model distillation