In partnership with
A Better Way to Deploy Voice AI at Scale
Most Voice AI deployments fail for the same reasons: unclear logic, limited testing tools, unpredictable latency, and no systematic way to improve after launch.
The BELL Framework solves this with a repeatable lifecycle — Build, Evaluate, Launch, Learn — built for enterprise-grade call environments.
See how leading teams are using BELL to deploy faster and operate with confidence.
Beginners in AI
Good morning and thank you for joining us again!
Welcome to this daily edition of Beginners in AI, where we explore the latest trends, tools, and news in the world of AI and the tech that surrounds it. Like all editions, this is human curated, and published with the intention of making AI news and technology more accessible to everyone.
THE FRONT PAGE
The $0.70 Model That Matches GPT-5: What DeepSeek's V3.2 Release Means
TLDR: Chinese AI lab DeepSeek released two new open-source models claiming to match GPT-5 and Gemini 3.0 Pro on reasoning benchmarks—at a fraction of the cost—reigniting questions about who's leading the open-source AI race.
The Story:
DeepSeek dropped V3.2 and its high-compute variant V3.2-Speciale on December 1st, both available on Hugging Face under an MIT license with full open weights. The Speciale variant scored gold-medal level on four elite competitions: 35 out of 42 points at the 2025 International Mathematical Olympiad, 492/600 at the International Olympiad in Informatics (ranking 10th overall), and second place at the ICPC World Finals—all without internet access or tools during testing. The key technical advance is DeepSeek Sparse Attention (DSA), which cuts inference costs by 70%: processing a 300-page document now costs $0.70 per million tokens versus $2.40 for its predecessor. Input pricing starts at $0.028 per million tokens—roughly 10 to 30 times cheaper than OpenAI or Anthropic.
Its Significance:
Open-source AI just got a lot more competitive. When any developer can download, modify, and deploy a model that matches proprietary offerings from OpenAI and Google—for free—the economics of the entire industry shift. Startups, researchers, and enterprises that couldn't afford API fees now have access to frontier-level reasoning. That matters because open weights let developers inspect how models work, fine-tune them for specific tasks, and run them on their own infrastructure without sending data to third parties. It's also not just DeepSeek pushing this direction: Arcee AI released its Trinity models the same day—the first open-weight models trained end-to-end in the United States under an Apache 2.0 license, with a 420B-parameter version coming in January. Chinese labs like Qwen, DeepSeek, and Moonshot have set the pace on open-source AI for much of 2025. Whether American labs can catch up—or whether it matters when the models are freely available anyway—is shaping up to be one of the defining questions of 2026. What DeepSeek's motivation is for giving away frontier-level AI also remains to be seen.
QUICK TAKES
The story: OpenAGI released Lux, a foundation model designed to operate computers autonomously by interpreting screenshots and executing actions across desktop applications. The company says Lux achieves an 83.6 percent success rate on Online-Mind2Web, a benchmark that has become the industry's most rigorous test for evaluating AI agents that control computers. That score is a significant leap over leading models—OpenAI's Operator scores 61.3 percent and Anthropic's Claude Computer Use achieves 56.3 percent. Lux also operates at roughly one-tenth the cost while executing tasks faster.
Your takeaway: A small MIT-founded startup is claiming to beat the major AI labs at computer control—and at a fraction of the cost. If these benchmarks hold up in real-world use, it could pressure the industry to move faster on autonomous agents.
The story: Visa and AWS have teamed up to enable AI agents to transact securely and autonomously on behalf of users. The companies will publish blueprints to the public Bedrock AgentCore repository for multi-network agentic retail shopping, travel booking and B2B payment reconciliation agents. The blueprints are being developed in coordination with Expedia Group, Intuit and the Eurostars Hotel company.
Your takeaway: This creates the foundation for AI agents to actually complete purchases on your behalf. When you tell an agent to "buy me tickets if the price drops below $150," this infrastructure makes it possible for that transaction to happen securely.
The story: Agent memory remains a problem that enterprises want to fix, as agents forget some instructions or conversations the longer they run. Anthropic believes it has solved this issue for its Claude Agent SDK, developing a two-fold solution that allows an agent to work across different context windows. The approach uses an initializer agent to set up the environment, and a coding agent to make incremental progress in each session and leave artifacts for the next.
Your takeaway: AI agents often lose track of what they're doing on complex, multi-day tasks. This solution mimics how human engineers hand off projects between shifts—leaving clear notes and status updates for the next session.
The story: Amazon sessions that resulted in a purchase surged 100% on Black Friday compared with the trailing 30 days, while sessions without Rufus that resulted in a purchase increased by only 20%. AI traffic to U.S. retail sites increased by 805% year-over-year on Black Friday. Shoppers who came to a retail site from an AI service were 38% more likely to buy, compared with non-AI traffic sources.
Your takeaway: This is hard data showing AI shopping assistants aren't just novelties—they're driving real purchases. Nearly half of consumers surveyed said they've used or plan to use AI for holiday shopping.
The story: Arcee AI announced the release of Trinity Mini and Trinity Nano Preview, the first two models in its new "Trinity" family—an open-weight model suite fully trained in the United States. Trinity Mini is a 26B parameter model with 3B active per token, designed for high-throughput reasoning, function calling, and tool use. A 420B-parameter model, Trinity Large, is now in training and set for release in January 2026.
Your takeaway: Chinese labs like DeepSeek and Qwen have dominated open source AI development this year. Arcee is one of the first U.S. companies to release competitive open models with full data provenance—important for enterprises worried about where their AI comes from.
TOOLS ON OUR RADAR
📝 Haxiom
Free: Keep your team's Markdown docs organized with AI that detects duplicates, prevents stale content, and syncs directly with your GitHub workflow.✏️ Sketch
Freemium: Design beautiful interfaces on Mac with powerful vector tools, real-time collaboration, and instant prototyping—built by designers, for designers.📬 Inbox Agents
Freemium: Merge email, LinkedIn, Instagram, WhatsApp, and Slack into one unified inbox that surfaces money-making messages and hides the noise.🖼️ Loomoz
Freemium: Capture full-page screenshots, add threaded annotations, and get client approvals faster—with Slack, Asana, and Trello integrations built in.
TRENDING
Artificial Tendons Give Muscle-Powered Robots a Boost – MIT engineers developed artificial tendons made from tough and flexible hydrogel that they attached to lab-grown muscle tissue. When they stimulated the muscle to contract, the tendons pulled a robotic gripper's fingers together three times faster, and with 30 times greater force, compared with designs without tendons.
Google's AI Advantage: What It Already Knows About You – A Google Search executive said that one of the company's biggest opportunities in AI lies in its ability to get to know the user better and personalize its responses. The company is connecting Gmail, Calendar, and other apps to make AI answers more personal—but raises questions about where helpful ends and invasive begins.
Space Force Seeks AI and Cybersecurity for Satellite Command – Project Kronos seeks to consolidate and upgrade legacy systems into a unified software suite that provides satellite operators with intelligence tools and command-and-control capabilities. Companies can respond to the opportunity by December 11.
OpenAI Partners with Accenture to Accelerate Enterprise AI – Accenture is rolling out 40,000 ChatGPT Enterprise licenses and naming OpenAI its primary intelligence partner. Accenture will use OpenAI's AgentKit to help clients rapidly design, test and deploy custom AI agents.
TRY THIS PROMPT (copy and paste into Claude or Gemini(click build first))
Meeting Cost Calculator: See what meetings actually cost your company—with live time tracking and ROI analysis
Build me an interactive Meeting Cost Calculator as a React artifact that reveals the true financial impact of meetings and helps me run leaner, more valuable ones.
The console should include these tabs/sections:
1. **Meeting Setup** - Quick input section:
• Meeting name/topic (text input)
• Duration slider (15 min → 2+ hours)
• Frequency selector: One-time | Weekly | Bi-weekly | Monthly
• Attendee builder:
- Add attendees by role/title
- Estimate hourly rate (or use defaults: Junior: $30, Mid: $60, Senior: $100, Director: $150, VP+: $250)
- Visual count and total hourly rate display
2. **Live Cost Tracker** - Real-time meeting meter:
• Big, ticking cost counter showing current meeting spend
• "Start Meeting" button that begins live timer
• Per-person cost breakdown (visual bars)
• Cumulative annual cost (based on frequency)
• Cost comparison: "This meeting costs the same as [relatable item]"
• Pause/Stop buttons
3. **Value Assessment** - ROI calculator:
• Rate the meeting outcome (slider: Waste of time → Decent → Extremely valuable)
• What was decided/accomplished? (text input)
• Could this have been an email? (Yes/No toggle)
• Efficiency score based on cost vs. rated value
• Suggestion engine: "Next time, try [shorter duration/fewer people/async update]"
4. **Meeting Audit Dashboard** - All meetings overview:
• List of all saved meetings with costs
• Total monthly/yearly meeting spend
• Most expensive meetings ranked
• Average meeting efficiency score
• Visual breakdown by meeting type
• "Kill Switch" button to remove recurring meetings
5. **Optimization Toolkit** - Best practices section:
• Meeting necessity checklist (5 questions before scheduling)
• Attendee trimmer: "Who actually needs to be here?"
• Async alternatives generator
• Meeting agenda template
• One-touch "Meeting Charter" (purpose, agenda, decisions needed)
Make it professional with clear data visualizations, smooth interactions, and a slight edge of "holy crap, we're spending HOW much?"What this does: Creates an eye-opening calculator that shows exactly what meetings cost in real dollars—making it painfully clear when a meeting should be shorter, smaller, or eliminated entirely. Perfect for making the business case to cut meeting bloat.
What it looks like:
WHERE WE STAND (based on today’s stories)
✅ AI Can Now: Control computers by looking at screenshots and clicking through apps—one new model succeeds 83% of the time on complex tasks.
❌ Still Can't: Remember what it was doing yesterday. Long-running AI agents still lose track of instructions when tasks stretch across multiple sessions.
✅ AI Can Now: Help shoppers find and compare products so effectively that AI-assisted customers are 38% more likely to complete a purchase.
❌ Still Can't: Make purchases on your behalf without new infrastructure. Secure, autonomous transactions require payment networks and AI platforms to build trust layers that don't fully exist yet.
✅ AI Can Now: Power robots with real muscle tissue, using artificial tendons to move 3x faster and with 30x more force than muscle alone.
❌ Still Can't: Operate those bio-robots outside the lab. The living tissue still needs controlled environments, nutrients, and protection to survive.
RECOMMENDED LISTENING/READING/WATCHING
In 15 minutes, CGP Grey explains why automation is different this time. He walks through how machines are taking not just factory jobs but white-collar work, creative work, and jobs we assumed only humans could do. His stick-figure animations make complex ideas clear without talking down to you.
The video came out in 2014, and it's only gotten more relevant. When he compares unemployed humans to unemployed horses after the car was invented, it hits hard. This is the video to send someone who thinks AI is just a tool like any other tool.
Thank you for reading. We’re all beginners in something. With that in mind, your questions and feedback are always welcome and I read every single email!
-James
By the way, this is the link if you liked the content and want to share with a friend.
