Your next desktop AI might need a second desktop. Meta’s Muse Spark Open can run across two M5 Ultra Mac Studios with a combined terabyte of unified memory, according to the company. That is a meaningful shift in where a large model can operate—not a claim that everyone’s laptop has suddenly become a data center. It leads AI news today, October 16, because control over deployment is becoming part of the product.

Thursday’s other developments test that idea from several directions. Amazon wants a wearable or phone attachment to turn conversations into useful memory. Cohere is pursuing a cybersecurity acquisition, while Ramp has raised another large financing round as it expands AI-related financial controls. In a Boston courtroom-policy meeting, experts confronted the harder question of how to trust what a recording appears to show. If you are choosing tools, the useful thread is control: over compute, captured conversations, software vulnerabilities, costs and evidence.

AI news today: Meta releases Muse Spark Open—with a hardware and license reality check

Meta released Muse Spark Open’s weights on October 15. The multimodal model has 816 billion total parameters, 42 billion active parameters and a one-million-token context window. It targets coding and agentic work, with hosted API access through several inference providers as well as downloadable weights.

Meta describes local execution using two M5 Ultra Mac Studios, each with up to 512GB of unified memory, through Apple’s MLX framework. The Verge reports that the normally rival companies collaborated on the launch. The practical news is another deployment option for organizations that want to operate the model themselves, not a universal replacement for a hosted service.

Meta’s “most capable US open-weight” description is a company claim. The weights carry the Muse Model License with limited commercial terms; “open-weight” should not be silently upgraded to unrestricted open source. Before a deployment, examine the license, supported configuration and your own task results. Our Maia 300 brief provides context on the parallel race to control AI infrastructure.

Amazon’s new Bee devices put conversation memory behind a physical button

Amazon introduced Bee Frontier and Bee Eclipse: a wearable available for $74.99 and a magnetic phone attachment coming soon from $99.99. Basic summaries, recaps and 14-day memory come with the purchase. Bee Pro costs $11.99 monthly and adds features including app integrations, suggested to-dos and the desktop application.

The new connections include Alexa+ and Amazon.com. Amazon says conversations are processed into insights in real time and the audio is deleted afterward; transcripts and summaries are distinct records that users can delete. Those are the company’s privacy representations, not an independent audit of the system.

The product question is therefore bigger than whether the summary is good. Decide who can consent, which environments should remain off-limits and what information should reach connected applications. A visible indicator and an off button help communicate capture, but they do not substitute for an organization’s rules. Compare the ongoing subscription cost and permissions with the actual remembering you need done.

Cohere proposes an AISLE acquisition rather than a completed security integration

Cohere’s October 15 announcement describes a proposed acquisition of AISLE, which builds autonomous systems to discover and help remediate previously unknown software vulnerabilities. The company places the transaction within its push for a sovereign AI platform combining models, applications, deployment infrastructure and security capabilities.

The status deserves careful wording: this is a proposed transaction, not confirmation that an acquisition has closed or that every customer can use an integrated product today. Cohere says AISLE’s approach combines foundation models, security expertise, large-scale code analysis, validation and automated remediation.

For buyers, the appeal is a shorter path from finding a flaw to verifying a fix. The risk is collapsing those steps into a single opaque score. Ask who authorizes testing, how false alarms are checked, what evidence a patch produces and how it is rolled back. Watch transaction completion, product availability and independently reproducible results before treating the strategic ambition as an operational capability.

Ramp raises $1.85 billion as AI spending becomes a finance workflow

Ramp announced a $1.85 billion primary financing round led by Dragoneer Investment Group at a $60 billion valuation. The company reports more than $2 billion in annualized revenue and positive cash flow. Those are company-reported operating figures; annualized revenue is not the same as revenue already earned over a completed financial year.

Its release points to a broader financial-operations platform, including AI Token Spend Management and a model-routing service. These are described as recent launches, not necessarily products first released on October 15. The day’s new event is the financing.

For leaders, the useful signal is growing attention to the cost of inference across providers, teams and workloads. Do not adopt a vendor’s savings claim as your budget assumption. Compare total spend, completed tasks and review effort before and after any change. A large funding round can finance distribution and development; it does not independently prove that a particular automation pays for itself.

A judicial panel hears why spotting a deepfake is becoming harder

Reuters reports that OpenAI and Meta representatives warned the U.S. Judicial Conference’s Advisory Committee on Evidence Rules in Boston on October 15 about increasingly convincing synthetic audio and video. The committee appointed a subcommittee for further examination. Members also cautioned against rushing to AI-specific rules when existing rules already address fabricated evidence.

This was a discussion about authentication, not the adoption of a new nationwide evidence rule. For teams handling important recordings, the practical response is to preserve originals, document how they were obtained and keep an editing history. A confident human inspection or a detector’s score should not become the entire verification process. The next thing to watch is the committee’s actual proposals, not predictions about what courts might eventually require.

Watch & Learn

Watch “Should You Use Open Source Large Language Models?” — IBM Technology. Martin Keen explains trade-offs involving transparency, customization, community contributions and risks shared with proprietary models. Set aside about ten minutes for the lesson and notes. It suits product owners assessing Meta’s release without wanting a deployment tutorial. IBM’s companion page confirms the teaching content. Treat its older model examples as background, and inspect today’s licenses separately. A text link avoids an unverified embed.

AI, Translated: content provenance

Content provenance is the recorded history of a digital file: where it came from and what happened to it. A photograph might carry signed information identifying the capture tool and later edits. C2PA Content Credentials provide a tamper-evident way to represent such history. That can support verification, but it does not prove that the scene or claim depicted is true. You should care because tracking an asset’s origin answers a different question from deciding whether its message is accurate.

Try This Today: rehearse a recording-consent notice in ChatGPT

Goal: draft and stress-test a short, plain-language notice for an invented AI note-taking pilot in ten minutes—not obtain legal approval.

1. In the ChatGPT desktop app, start a new temporary chat with Command–Shift–N on macOS or Ctrl–Shift–N on Windows. If you customized shortcuts, check Settings → Keyboard Shortcuts.
2. Provide fictional facts: what is captured, what is retained, who can access it and how someone opts out. Include no real meeting content.
3. Ask for a notice and a skeptical participant’s objections. Resolve missing facts before having your organization review the wording.

Example prompt: “Using only these fictional facts, draft a 60-word notice. Then identify three ambiguities about consent, retention or access. Do not invent legal requirements or promise confidentiality.”

The official OpenAI documentation lists these desktop commands. App versions and customized bindings can differ. A temporary chat does not authorize uploading restricted information.

One thing to remember

More control over where AI runs is useful only if you also control what it captures, what it can change and what evidence it leaves behind. Ownership and trust are related—not interchangeable.