{{first_name | Hey}}, welcome back.
Today’s issue covers Meta turning its AI agent into a hardware platform, SoftBank piling billions in debt for its AI bet, and tiny AI models moving onto smart glasses.
These are today's updates.
Meta turns Muse into a hardware platform with VR glasses and an AI pendant
SoftBank sells $11.1B in bonds to fund its OpenAI investment
PrismML brings tiny language models to Qualcomm-powered smart glasses
Tools, jobs, resources, and last issue’s winning prompt ⬇️
Think Tank
A 17-minute TED talk by AI and design researcher Advait Sarkar. If your organization is rolling out AI tools faster than you're rethinking how people use them, this is worth 17 minutes. It's less about whether to adopt AI and more about what happens to judgment and reasoning skills on the other side of adoption.
Neatprompts is where enterprise teams discover what's worth paying attention to in AI. Every week, 100,000+ readers rely on us for actionable insights and practical workflows.
If your product helps businesses adopt, deploy, or scale AI, we'd love to introduce it to our audience.
Top News
Meta used its Connect developer conference this week to turn its three-week-old Muse AI agent into a hardware platform. Zuckerberg unveiled the $1,299 Meta VR Glasses, shipping spring 2027, and Muse Charm, a pendant-sized gadget with front and rear cameras that lets people talk to Muse without opening an app. He gave no price for Charm, but called Muse ‘the centerpiece’ of Meta's push toward what he termed personal superintelligence.
Investors have responded well. Meta's stock has added roughly $300 billion in market value since Muse launched on September 8, and the app has already passed ChatGPT at the top of Apple's App Store. But Muse is fighting real battles behind the scenes. Amazon has blocked it from completing purchases on its site. And TechCrunch's Sarah Perez argued that Meta's privacy history, a $5 billion FTC fine in 2019, an $18 billion child-safety settlement last month, is what determines whether Muse Charm succeeds, more than the gadget's design.
The Charm's dangling form factor taps directly into a Gen Z trend around bag charms and dangling accessories, the same instinct that turned Labubu into a $1.8 billion brand for Pop Mart. Meta is betting fashion can succeed where the AI Pin and Rabbit R1 failed. Here's Zuckerberg's own thread walking through everything announced at Connect.
SoftBank issued $11.1 billion in bonds this week, the largest high-yield corporate bond sale in global history, to fund the final $10 billion payment on its OpenAI investment. The sale splits into dollar notes paying 8.625% to 9.75% and euro notes paying 7.125% to 8%, rates that reflect SoftBank's junk credit rating. Once the payment closes October 1, SoftBank's total stake in OpenAI reaches $64.6 billion for roughly 13% ownership.
This is SoftBank's third distinct debt instrument tied to OpenAI this year, after an $11.87 billion syndicated loan and a $10 billion margin loan against its OpenAI shares. Combined, Masayoshi Son has stacked more than $30 billion in obligations onto a single bet. CreditSights called the concentration risk material, and SoftBank's own OpenAI-linked IPO, which was supposed to bring in fresh cash, got delayed this month.
Shares jumped 7% anyway. Investors are betting Masayoshi Son's confidence is earned. Whether that holds probably depends on how OpenAI's own IPO timeline plays out from here.
PrismML, founded by Caltech researchers, showed off a version of its Bonsai language model built specifically for AI smart glasses running on Qualcomm's Snapdragon AR1 Gen 1 chip at this week's Snapdragon Summit. The model runs at 1-bit precision, shrinking a much larger model by 4x while keeping most of its performance. The glasses version has 2 billion parameters, tuned for vision and language so wearers can ask about what they're looking at in real time.
No smart glasses running the model have actually been announced yet. Prism's bigger goal is proving open-weight AI can run entirely on a device, without leaning on cloud infrastructure or the privacy promises of a proprietary lab. If it works, it's a real alternative for hardware makers who don't want to route every query through OpenAI or Google's servers, and a sign that vision-language models are shrinking fast enough to fit inside a pair of glasses.
Signals
Tools
Resources
Poll
Choose the prompt you want us to publish next
Market
Funding
Island raised $400M to expand its enterprise AI security platform
Circle secured $100M from Binance to accelerate USDC expansion across global markets
Roles in AI
Director - IT Software Engineering at Elastic
Engineering Manager - Compute Infra at Databricks
Tutorial
Claude now uses one memory for both Chat and Cowork. If you give instructions to Claude in Chat, Cowork will remember it too, and vice-a-versa. This means your client details, report formats, and writing style carry over when Cowork drafts a brief or a weekly update. It's on by default for Free, Pro, and Max plans on web, desktop, and mobile.
How to set up Claude’s new shared memory across Chat and Cowork
Open Claude. You'll see a pop-up about the new memory update. (if not, go to your profile → Settings → Memory and turn it on there)
Decide if you want sensitive topics like health or beliefs included (these are off by default), then click Save preferences
Go to your profile → Settings → Memory to see everything Claude has saved, sorted into your profile, preferences, and topics
Click any topic to update it. Type the change in plain language and Claude files it in, or delete instructions you don't want kept
Click on "Import memory" to bring over what ChatGPT or Gemini already knows about you
In Cowork, click the desktop icon in the top right and choose to run tasks in the cloud. Shared memory only works for cloud tasks, and local tasks won't see it
Sample Prompt: Save to memory: my Q4 focus is mid-market fintech accounts, and I like weekly updates under 150 words.
Then in Cowork: Draft this week's pipeline update based on my Q4 focus.
Use Cases:
Get weekly reports without the setup
Write in your style
Prep for a pitch or meeting brief for client work without adding details everytime
Talk through projects that run for weeks to keep a track of deadlines, work done, what remains and who owns it
Collaborate with co-workers. "send the draft to Priya for review" makes sense to Cowork without extra explanation.
Switching from ChatGPT or Gemini
Meal plans, workout goals, shopping preferences
Prompt of the Day
AI Workflow Automation Finder
When to use this?
Use this when a recurring business process feels too manual, repetitive, or fragmented. It identifies where AI can automate or assist, estimates the potential impact, and gives you a practical path to implementation.
Act as an Enterprise AI Workflow Automation Finder. Analyze the workflow below and identify where AI, automation, or AI agents could reduce manual work, speed up execution, improve quality, or eliminate repetitive tasks.
Workflow:
[Describe the workflow step by step]
Team/function:
[Sales / Marketing / Finance / HR / Operations / IT / Customer Support / etc.]
People involved:
[Roles and approximate number of people]
Tools currently used:
[Slack, Salesforce, Excel, email, Jira, SAP, Notion, etc.]
Frequency:
[Daily / weekly / monthly / ad hoc]
Time spent:
[Approximate time per task or workflow]
Current pain points:
[Repetitive work, delays, errors, handoffs, bottlenecks, context switching, etc.]
Now analyze the workflow and provide:
1. **Automation opportunities**
Identify each step that could be:
- Fully automated
- AI-assisted
- Left human-controlled
2. **AI workflow designs**
For the highest-value opportunities, describe:
Trigger → AI action → Tools/data used → Human approval → Final action
3. **Impact**
Estimate potential time savings, cost savings, faster turnaround, or quality improvements where the available data supports it.
4. **Prioritization**
Rank opportunities using:
- Business impact
- Automation feasibility
- Implementation effort
- Risk
- Expected time to value
5. **Recommended first automation**
Identify the most practical workflow to pilot and explain why.
6. **Implementation plan**
Give me a simple Pilot → Test → Deploy → Scale roadmap, including required integrations, data, owners, and success metrics.
7. **Risks and guardrails**
Flag security, privacy, compliance, accuracy, human-approval, and reliability considerations.
8. **Executive summary**
End with:
- Top 3 automation opportunities
- Expected measurable impact
- Recommended pilot
- Tools/integrations required
- Key risks
- KPIs to track
IMPORTANT:
- Do not assume AI is the right solution for every step.
- Do not invent precise savings, ROI, or productivity figures when inputs are unavailable.
- Clearly label assumptions and estimates.
- Separate measurable benefits from qualitative benefits.
- Challenge automation ideas that add unnecessary complexity.
- Prioritize workflows that are repetitive, high-volume, rules-based, or involve significant manual coordination.
- Preserve human approval for high-risk decisions.
- Highlight missing information that could materially change the recommendation.
- Keep the output concise, practical, and suitable for an enterprise leader deciding what to automate next.Get more such prompts in the Prompting Playbook (free for you)
Stay curious, {{first_name | reader}}
PS. If you missed yesterday’s issue, you can find it here.
