My failed YouTube AI agent (don't laugh)

Learnings from the AI rabbit hole - use for your career

Author

Nelson R
June 22, 2025

Howdy my friend!

In today’s piece:

  • My experiments using AI to create YouTube content & lessons learned (use for your career in tech too)

  • Why “backend” processes might be better targets for AI automation than “frontend” outputs

  • Why the “buy” vs “build” decision is a false dichotomy with AI tools

  • The importance of “unified entity context” for AI in cybersecurity (credit: Daniel Miessler)

  • 5 big AI news stories, and 13 other handpicked things I think you’ll like

First a quick word from our partner…

Start learning AI in 2025

Keeping up with AI is hard – we get it!

That’s why over 1M professionals read Superhuman AI to stay ahead.

  • Get daily AI news, tools, and tutorials

  • Learn new AI skills you can use at work in 3 mins a day

  • Become 10X more productive

My failed YouTube AI agent (don't laugh)

These last few weekends, I’ve had extra time to disappear down the AI rabbit hole.

And disappear I did. I’d originally planned to spend each weekend getting an email out, but instead became obsessed with using AI for YouTube content.

I’ve been meaning to test YouTube content for my newsletter. I’m a YouTube addict, so it’s the perfect platform to diversify my traffic sources.

I’m running paid ads and this newsletter is costing me money.

The problem: YouTube content creation is time-consuming, and I’m already busy with my day job and this newsletter.

The solution? Use AI to repurpose my newsletter content into a video format.

Easy to say, less easy to do… in a vaguely watchable way.

I'd like to walk through my experience: what I tried and lessons learned.

Spoiler: it’s been a fail so far. But instructive and worth sharing.

Starting a YouTube channel may not be top of your career TODO list, but the lessons learned apply broadly to solving career problems with AI.

If you only care about AI or Cloud Security news updates, feel free to skip this section.

What I tried

Step one was checking for existing AI products that solved my problem.

I first looked at Invideo AI.

Their AI turns a prompt or script into a full video, including narration and visuals.

Promising, but too expensive. It also gave off “generic AI slop” vibes.

Next, I built an AI agent using the OpenAI Agents SDK. The overall workflow looked like this:

This is the little guy in action:

This video was my first attempt using my agent. It’s kinda janky and I wouldn’t recommend watching all of it. But it was an interesting first effort.

It’s highly generalisable. I tried a “hand-drawn sketch” style here, but there’s an unlimited range of content options.

A video where all the objects are made from ice cream? Possible.

Emulating the aesthetic of old video tapes with digital glitches? Also possible.

This agent was my first taste of what Andrej Karpathy calls “software 3.0”.

Instead of precisely specifying the sequence of tasks as you would in traditional software, I created an agent and gave it access to the key tools for video generation (i.e. crafting an image prompt, splitting a script, generating a video from an image, etc).

I gave loose instructions in the system prompt, and it figured out how to call these tools to get the job done.

So the heart of the software was encapsulated in this short block of code:

Why use an agentic implementation here, given this is a simple, linear workflow?

An agentic design enables truly creative, generalised video generation.

Imagine giving this agent access to advanced video editing tools, like editing videos or applying effects in Adobe Premiere Pro.

Already, there are experimental MCP servers exploring this sort of thing.

It could apply something like human judgement to make creative, interesting edits to the video.

And it could do so in ways that don’t conform to a simple deterministic workflow.

With more work, this approach could lead to something interesting.

However, it’s not there yet.

Next, I tried stitching together a few AI tools:

This was the result. Better than attempt #1 but still barely watchable.

I tried OpusClip to turn this long-form video into a series of shorts, which OpusClip assigns a “virality score” to.

These shorts brought in 1.5k+ views in 25 mins. I got excited until I realised viewers did NOT like the content. Fail.

The bar is high on YouTube. You’re competing with people who put in a LOT of effort and have truly mastered the medium.

Back to the drawing board. I’ve got another workflow to test — but that’s for next time!

What I learned

“Backend” processes might be better targets for AI automation than “frontend” outputs

Turns out, people hate content that looks like low-effort AI slop.

They like authentic content made by a real person. Shocking!

The most promising areas to automate with AI right now might therefore be processes or systems that aren’t directly visible to the end user.

For example, generative AI really shines for coding. Here, the actual code isn’t visible to the end user.

The user doesn’t care how the code was written. They care whether the application does what they want.

Another example is researching information for a blog post or work proposal. The reader doesn’t care how you found the information.

They care about the information and your analysis.

In this example, the research process is a “backend” process, and your final write-up is the “frontend” output.

I’m not arguing AI can’t be used for “frontend” things. It’s just harder, and more at the mercy of the “AI-human” gap.

The “AI-human” gap is the detectable difference in output between a human and an AI.

In some cases, this gap is small.

GPT-4.5 recently made history as the first model to pass the Turing test, meaning study participants couldn’t detect any difference interacting with the model versus humans.

In fact, GPT-4.5 was judged to be human 73% of the time. You read that right 🤯

Also, AI can now generate image outputs almost indistinguishable from human-created ones.

But in general, there’s a noticeable AI-human gap. The more this is true, the more “backend” processes will be an easier target.

The “buy” vs “build” decision is a false dichotomy

Should you pay for an existing AI solution or build your own?

I’m usually a fan of paying for a proven solution to your problem.

Likely built by experts, who do a much better job than whatever you cobble together yourself.

If there’s nothing on the market that does what you need (or nothing within your budget), that leaves the “build it yourself” option.

… or does it?

When thinking about AI, you can break your main problem into sub-problems.

This echoes Daniel Miessler’s concept of work as a series of intelligence tasks (a.k.a. an intelligence pipeline).

Let’s take the video generation example.

What’s involved in generating a video? You might break this down into:

  1. Researching content ideas for the video

  2. Creating a video script from the content ideas

  3. Creating nice audio from the script

  4. Creating engaging visuals to accompany the audio

  5. Creating a good video call-to-action

Maybe you struggle to automate all 5 tasks with AI, like me. But maybe AI could help with backend subproblems (1) and (2).

You could even use AI to streamline (3) with some manual tweaking.

This is a “buy AND build” paradigm. You’re paying for AI services that help you solve key subproblems.

Then you sort the other subproblems out yourself, either manually or with custom solutions.

Hybrid approaches combining AI and manual effort can work.

Don’t give up on pre-built AI offerings if you can’t solve the full problem right away.

OpenAI’s Agents SDK was easier to pick up than Google’s Agent Development Kit (ADK).

I initially tried Google’s ADK and found it surprisingly hard to work with.

Probably user error. I’m sure a smarter person would get it quicker.

The Google ADK has a steeper learning curve but provides a richer set of abstractions that could be useful for building complex, production-grade systems.

In contrast, the OpenAI SDK worked exactly like I expected. I had a working skeleton agent in 15 mins, then built my custom agent without major issues.

This tweet sums up my experience.

That said, it’s important to focus on capabilities over tools. Tools and frameworks change, but the fundamental concepts last longer.

AI subscription fatigue… RIP my bank balance

Shiny object syndrome is real with AI tools.

Each new tool expands the space of possible problems you can solve. Or improves your ability to solve existing problems.

Often both.

Historically, jumping around between tools was not a great career strategy.

But in the era of agentic AI, things are different.

Powerful tools are emerging at an unprecedented rate.

You don’t need months to master each tool. With a simple prompt, you can start getting value from these tools.

In this new paradigm, simply knowing about a tool’s existence confers a competitive advantage.

That gives you another “Lego block” to solve your problem.

The downside? These Lego blocks cost money. I did a self-audit recently and realised I was paying for 16 subscriptions… moment of silence for my bank account.

Like their traditional SaaS forebears, AI companies love the subscription pricing model.

I do wonder if we’ll see more innovation in the business model, not just the technology.

For example, Bret Taylor uses an outcome-based pricing model for his AI customer support startup Sierra. The idea is you pay for the job done, not a set amount per month.

For the sake of my bank balance, I hope more AI companies follow suit 🙏🏼

Winging aside, I like to positively reframe these costs as an “experimentation tax.” It’s the price to explore what’s out there.

I’m not a doofus spending too much on dumb things (keep repeating that to myself and it sounds true).

My favourite finds

Nelson's summary: The video explains that AI is heading towards a future of "Unified Entity Context," where a complete, dynamic understanding of an entity (like a person or a company) will be the foundation for all decision-making.

In other words, comprehensive information about an organisation that can be ingested by an AI will matter more than almost anything else.

Why it matters: This concept could revolutionize cybersecurity by turning it into a battle of who has the better contextual model of an organization, the attackers or the defenders. Fascinating concept with big implications for where to allocate cybersecurity spend.

Nelson’s summary: Meta’s swapping out humans for an AI that will auto-approve about 90 % of privacy and integrity checks before new Facebook, Instagram and WhatsApp features ship.

Why it matters: When AI is signing off on launches, security pros who can “audit the auditor” (i.e. the the AI) and bolt on guardrails will be in seriously hot demand.

Nelson's summary: OpenAI just dropped a 34-page playbook that walks builders through models, tools, orchestration patterns, and safety guardrails for shipping real-world LLM agents.

Why it matters: This gives us a plug-and-play blueprint to design and lock down agent-driven workflows before attackers do. These guides from frontier labs distill a lot of wisdom about agent engineering into fairly digestible patterns.

Required reading as security practitioners. You can’t secure something without an understanding of how it’s built.

Nelson’s summary: OpenAI slashed o3 pricing 80% and introduced a heavyweight o3-pro tier for advanced reasoning and strategic planning.

Why it matters: cost is a key factor driving model choice for real-world use cases. This sudden price drop for o3 makes OpenAI competitive with Google on cost, an area they were previously falling behind.

o3-pro (only available on the API or the ChatGPT Pro plan currently) works well for challenging questions where reliability matters more than speed. Some have called it a master strategist. Though for most users, most of the time, the additional response time won’t be worth it - o3 will be the better choice.

Nelson’s summary: Meta has acquired 49% of AI unicorn Scale AI for $14.3 billion, and lured Scale AI’s Alexandr Wang to help build their new “Superintelligence Lab”.

Why it matters: This will have a really interesting impact on Meta and Scale AI's positions in the competitive landscape.

Both companies have strong “data moats” — Meta with their insane amounts of consumer data, and Scale AI with it’s data labelling capabilities. There is a natural synergy between them; a partnership makes sense. Watch this space.

Nelson's summary: Legendary investor Mary Meeker’s latest 340-page deck on AI reveals the industry’s explosive growth, with ChatGPT reaching 800M weekly users and $10B in annual revenue by 2025.

Why it matters: AI's rapid evolution will make specialized skills in AI security and risk management critical going forward.

The slides at the start of this deck paint the clearest picture of how unprecedented this level of technological change is. They’re all very “up-and-to-the-right”-y.

The security landscape (and the world in general) is transforming at a vertigo-inducing speed. Strap in, it’s gonna be a wild ride.

Other Handpicked Things

State of the art prompting from the Lightcone Podcast: Great high-level crash course in prompt engineering from startup incubator Y Combinator. Includes a great breakdown of the rising trend of metaprompting, as well as some cool prompt examples (including a massive six-page prompt used by Parahelp).

YC-backed startup Tropir, the first autonomous LLM-Ops engineer: early example of a totally new product category - “AIs that help build better AIs”.

chai.new: vibe code AI agents with text-to-agent capabilities. Very much like bolt.new but optimised for AI agent creation. I gave it a go and it worked great for a simple “text in / text out” agent use case. But not so great when I wanted to do more complex workflows involving images or videos etc. Cool concept, but jury's out on this one.

Runner H: new agentic AI start up that raised over $220 M is now available for free to the public. I gave it a brief whirl & it looks interesting. Let me know what you think if you try it.

OpenAI adds connectors to Gmail, Drive, Outlook and Teams: ChatGPT can now connect to your Google and Microsoft applications for a more personalized experience. A natural next step for them.

Raaz Herzberg (CMO and VP Product Strategy at Wiz) interview: Cool interview from the CMO at Wiz, the iconic cloud security startup that Google has just acquired for $32B.

Shoto & Trent on Dwarkesh Patel: must watch for any AI enthusiast. These two guys are big AI nerds, but fascinating listening to them talk.

Jeff Dean on the coming transformations in AI: Jeff Dean's been a key figure behind-the-scenes with AI at Google. Always worth a listen.

AI playlist on Spotify: new feature on Spotify where you can use text prompts to create your own custom playlists. For example: “create a playlist similar to my top tracks, but that I haven't heard before”. This is a nifty feature, I've been enjoying it.

Tracing the thoughts of a large language model: Anthropic produces some of the best mechanistic interpretability research out there. Fascinating read and real achievement in the AI safety space.

On the Biology of a Large Language Model: Another mechanistic interpretability paper from Anthropic. This research really deserves headline spots in my newsletter. I'd like to do a deeper write-up in the future.

AI diplomacy: a dozen of the top AIs were pitted against each other in a game of Diplomacy, which involves trying to take over the world in a simulated game. Fascinating to see the different Machiavellian strategies that emerged and which models performed the best. The models that learned to deceive and backstab effectively did the best… tad worrying but also kinda funny.

Nvidia’s Blackwell latest MLPerf benchmark: this benchmark measures how long it takes the platform to train models. TLDR; Nvidia's newest Blackwell GPU architecture kicks butt, performing 2.2 times faster than the previous Hopper generation.

Before you go, I’d love to know what you thought of today’s newsletter. That way I can improve the experience for you and minimise suckiness.

What'd you think of this email?

Login or Subscribe to participate in polls.

Take care.

Nelson, fellow AI enthusiast and cloud security dude.