Watchman Privacy Podcast - Episode 213

NanoGPT: Pay-Per-Prompt AI Service (Feb 9, 2026)
Host: Gabriel Custodiet
Guest: Milan Darid, Co-Creator of NanoGPT

Episode Overview

This episode features a returning conversation with Milan Darid, co-creator of NanoGPT—a privacy-focused AI hub offering anonymous, pay-per-prompt access to hundreds of text, image, and video AI models. Host Gabriel Custodiet and Milan discuss major updates since their last talk, dive into privacy and technical nuances, highlight creative user workflows, and assess both the promise and pitfalls of anonymous AI deployment.

Key Discussion Points & Insights

1. Major Updates to NanoGPT

• Improved Account Privacy – NanoGPT upgraded from cookie-based anonymous sessions to a system using random UUIDs (like some VPNs), making multi-device use seamless while preserving anonymity.
  • "You can log in from other devices...even being fully anonymous." (03:10-03:30)
• Optional Subscription – An $8/month plan grants nearly unlimited access (60,000 queries/month) to open-source, no-log models. The traditional pay-as-you-go still exists.
  • "A subscription for all of the open source models...no log, no training providers." (04:00)
• Expanded Model Offerings & Usability – Hundreds of AI models (text, image, video) have been added.
• Conversation Sync – Users can sync conversations across devices with either NanoGPT’s storage or by connecting their own (e.g. AWS S3).
  • "You can add in your own storage, so you can be fully in control of it yourself." (05:04)
• Cryptocurrency Support – In addition to Monero, now supports Zcash and Litecoin’s Mimblewimble extension for private payments.

2. Privacy Layers: How Private Is NanoGPT?

• Self-Hosting Models – Remains the gold standard for privacy.
• TEE Models (Trusted Encryption Environment) – Second-best: code and “no logging” promises can be verified through hardware enclaves.
• Open-Source Providers – Chosen based on no-logging assurances, but Milan is clear:
  • "We are relying on their word, right? … It’s just a sort of promise from them." (07:50)
• NanoGPT Itself Does Not Host Most Models – Instead, the service carefully selects reputable providers.

3. Trends and Tips: Video AI

• State of Video AI
  • "It's gotten insanely good. It's like scarily good...you can essentially create anything you want to and have it look just completely lifelike." (09:35)
• The “Sora 2” model allows production of up to 12-second, nearly photo-realistic clips, now available on NanoGPT.
• Pricing – Short clips can cost up to $0.70-$0.80. Experimentation is pricier than with images.
• Economy Models – For budget experimentation:
  • "The Wan 2.2 Turbo model...quite cheap...5 cents for about five seconds." (11:55)
• Prompt Engineering for Video
  • "The perfect prompt has gotten less important over time...For most, just describe what you would write down for a filmmaker." (12:50)
  • Users can examine prompt examples on the website to understand input-output dynamics.

4. Integrations in Crypto Wallets

• Access via Wallets: Cake Wallet, NanSwap, and bitcoin.com wallet now feature integrated NanoGPT access for seamless, private, and account-free AI usage.
  • "It's kind of a way for people to also see the power of crypto...not just investment." (14:35)
• Crypto-Powered Workflows: Some users enable AIs to manage wallets and make payments for further prompts, opening up new automation and business models.

5. Image Model Recommendations & Editing

• Model Selection: When you prompt image generation, NanoGPT recommends the most suitable model based on leaderboard data and your input’s characteristics.
  • "We try to classify it...then those leaderboards have what is the best image model for those categories." (16:51)
• Top Models
  • NanoBanana: Editing/furniture visualization.
  • C Dream 4.0: Versatile, cost-effective generation.
  • Quen (Alibaba): Powerful open-source choice.
• Technical Controls: Most platform-native controls (e.g., upscaling, editing) are available; some, like MidJourney’s unique API-only features, are exceptions.

6. Guidance for Terminology & Parameters

• Inference Steps: Represents refinement effort—more steps, better (to a point).
• Guidance Scale: Adjusts prompt adherence vs. creative freedom.
  • "The higher you make the guidance skill, the more exact the model sticks to what you said..." (23:30)
• Negative Prompts: Specify what to exclude from results.

7. Popular Text Models & Use-Cases

• Grok 4: Available, not widely used.
• Top Text Models:
  • Claude Sonnet 4.5: Favored for programming and role-playing due to developer focus and large context size.
  • GPT5 Pro: Expensive but extremely powerful for code and reasoning.
  • GLM 4.6 (Open Source): Versatile and cost-effective.
  • "Claude is really good for programming and roleplaying...but I'm mostly using GPT5 Pro now." (27:00)

Notable Quotes & Memorable Moments

• On Privacy Guarantees
  "We are relying on their word, right? We have no insight into their codebase." — Milan, (07:50)

• On AI Video Quality
  "Some of the commercials you'll be seeing come out in the next weeks. Probably they're using AI just because it's gotten so incredibly good." — Milan, (10:35)

• On Most Expensive Prompts
  "We've had people who spend like $20 on a single prompt...they would message us like, this has gotten really expensive. I'm like, yeah, you threw in essentially the entire Lord of the Rings into...one of our most expensive models." — Milan, (28:06)

• On Creative AI Use
  "Kids love it...they enter this simple prompt of like, Spiderman fighting Superman and they get this entire video and they're like, I created this." — Milan, (31:00)

• On AI’s Real-World Medical Use
  "I fed my test results in ... it was essentially saying...depending on what medication you use...maybe try this...day after I go to the doctor, and he essentially says exactly the same..." — Milan, (34:32)

• On the Internet as AI Training Data
  "If it can...sort through the bad answers and just see what gets upvoted most and what gets...thankful replies...that could actually be very useful." — Milan, (38:10)

Standout Segments & Timestamps

• NanoGPT Privacy Upgrades & Subscription Introduction (02:52–05:42)
• Open Source/TEE Models & Privacy Caveats (06:52–08:33)
• State of Video AI & Prompt Crafting Advice (09:25–14:18)
• Crypto Wallet Integration Details (14:35–16:43)
• Image Model Recommendation Logic (16:51–18:41)
• Favorite Image Models & Open Source Debate (18:47–20:30)
• Technical API Controls & Limitations (20:48–22:28)
• Image Generation Parameters Explained (22:43–24:40)
• Grok & Leading Text Model Usage (24:46–26:48)
• Expensive Prompt Anecdotes (28:06–29:37)
• Most Creative & Unusual AI Workflows (30:03–31:58)
• NanoGPT as a Business Backend & AI Automation (32:04–33:46)
• AI Hallucinations and Medical AI Experience (34:05–36:24)
• Reflections on AI Training Data Sources (36:24–38:45)

Final Thoughts

Milan reflects on the rapid pace of change:
"You sort of realize how much is happening in such a short time." (39:21)

Gabriel closes by reaffirming NanoGPT as a valuable privacy-respecting entry point for AI experimentation and encourages listeners to try the service via the website or through integrated crypto wallets.

Resources Mentioned

• NanoGPT.com
• EscapeTheTechnocracy.com – Gabriel’s course "AI Resistance"
• Cake Wallet
• bitcoin.com Wallet
• Previous Episode with Milan — Full NanoGPT walkthrough (see show notes)

For those seeking privacy-friendly, anonymous, and highly customizable access to the AI landscape—with everything from image and video tools, seamless crypto payments, and unique workflow automations—NanoGPT offers an intriguing, rapidly evolving playground.