Category: Software

Choose the Right AI Tool

The AI landscape has changed a bit since I wrote What Is Real AI? back in 2021. The advent of GenAI has enabled a new wave of dubious AI sales pitches. Here’s one that crossed my desk recently:

We’ve identified some key GenAI opportunities at PermaPlate … forecast revenue and claims across [products] and adjust staffing monthly.

This sounds like a good idea, except – it’s not a GenAI application. It’s a standard forecasting exercise that everybody does already. If I did want to switch to a learning model – even a deep neural net, which is architecturally similar to an LLM – it would still not be GenAI.

The thing to remember is that GenAI “generates” things, like blog posts and deepfakes. My favorite learning models, going back to AI-Based Risk Rating, are all quantitative in nature. Here again, there are plenty of good statistical methods. Even if you prefer to use a learning model, you may not choose a neural net.

Neural nets have a problem with explainability. They’re basically a black box. That’s why credit bureaus, which must be able to explain their ratings, use a two-step approach. They use AI for exploration and feature engineering, but then they put the features into a more-transparent logistic regression model.

I might consider GenAI in a forecasting application, to deal with unstructured data. On the other hand, I would ask why the data is unstructured. We did this exercise as a POC here at PermaPlate. I wrote a little program that would read a service contract, and then answer natural-language queries.

Which coverage did the customer select and does it include roadside assistance?

It was a cool demo, but – if you want coverage details available for automation, it makes a lot more sense to store them in machine-readable form, at origination time. And what kind of automation might that be? Well, it might be “agentic.”

Agentic AI means that the AI has “agency,” in the sense that it can make decisions and do things in the world. Cool, huh? We give AI agency by equipping it with tools, in the form of software APIs.

Imagine asking ChatGPT to organize your next trip. It can’t, because it’s trapped inside your web browser. But if you invoke ChatGPT as part of an agentic workflow, with interfaces to the airlines and hotels, it can actually book the trip.

Agentic workflows often divide the work among tool-using LLMs, with a mastermind LLM directing the others. For systems that don’t have APIs, the agent can use Robotic Process Automation to operate the system’s user interface – just like you would at the keyboard. It’s not surprising that UIPath, one of the leading RPA vendors, has moved into Agentic AI.

Here is a short list of the latest AI techniques:

  • Large Language Model (LLM) – Like Grok and ChatGPT, these are AI models that can read and write (and plan, and execute).
  • Generative AI – Broad class of AI models that can create things, including LLMs but also diffusion models for video and other media.
  • Deep Neural Net (DNN) – Core technology behind GenAI, and many other learning models, as in my earlier article.
  • Retrieval Augmented Generation (RAG) – As the name implies, GenAI “augmented” by the ability to find and read your documents. See Unguided RAG for Text Comparison.
  • Robotic Process Automation (RPA) – Not AI, but frequently used by Agentic AI. As I wrote in Applied AI for Auto Finance, you can derive a lot of efficiency from RPA alone.
  • Agentic AI – AI agents that can make decisions and act autonomously.

Now that you know the lingo, you can choose the right tool for the job – or your AI sales pitch. I, for one, will not be using GenAI to predict claims volume … but I may use Agentic AI to dispatch the technicians.

AI Datacenters Are Eating the World

AI datacenters are categorically different from traditional “hyperscale” cloud providers. The older datacenters were optimized for networking and storage – think streaming video and commerce websites.

AI datacenters are optimized for computing, and density is king. The goal is to pack as many parallel processing chips – Nvidia GPUs and Google TPUs – into a rack, and as many racks into the building, as possible.

This means that power consumption and cooling requirements are through the roof. One rack in a typical AWS datacenter might draw 20 kilowatts, while the latest Nvidia rack draws 132 kW. Pack the building full of those, and…

Project/Company Target Capacity Status & Timeline Key Details
xAI Colossus Expansion 1.2 GW Expansion underway; 150 MW substation completing Q4 2025, full 1.2 GW by late 2025–early 2026 Building on the existing 250–300 MW Colossus cluster (built in 122 days). Involves on-site natural gas plant and grid upgrades; faces EPA scrutiny but leverages pre-existing factory infrastructure for speed.
OpenAI Texas Campus 1 GW (phased from 300 MW) Phase 1 (300 MW) operational; Phase 2 construction started Jan 2025, full GW by mid-2026 Houses hundreds of thousands of GPUs; includes 210 substations and on-site electrical upgrades. Already straining ERCOT grid—could equal ~10% of regional peak during heat waves.
Meta 1–5 GW (supercluster) Ground broken 2024; first 1 GW online 2026, scaling to 5 GW by 2030 Zuckerberg’s “gigawatt-plus” initiative; Meta’s largest yet, with $64–72B spend in 2025 alone. Focuses on liquid cooling for high-density AI racks; part of broader multi-GW campus plans.
Microsoft OpenAI “Stargate” 1–5 GW Planning advanced; construction to start late 2025, launch 2028 $500B joint venture with Oracle/Nvidia; aims for massive AI sovereignty. Includes SMR nuclear pilots; power sourcing via PPAs and on-site generation to bypass grid delays.

Microsoft has two 300 MW datacenters. This is comparable to peak load for the city of Tacoma, during their summer AC season. Within a few years, all the leading AI vendors will have datacenters above 1GW. That’s why Microsoft just made a deal to restart the infamous Three-Mile Island nuclear power plant.

A cynic might observe that, while the TMI facility was deemed unsafe to power homes and businesses in Pennsylvania, regulators were willing to reconsider once Microsoft came knocking. Likewise, in Europe, nuclear-powered France is winning the datacenter race over green Germany.

After years of woodburning and windmills, the voracious demands of AI are forcing the world to take another look at nuclear power.

The Art Project

Can AI be used to match and classify images? Of course! They do this all the time, looking at everything from paint chips to x-rays. In today’s post, I use an established model called ResNet-50 to match and classify post-impressionist artists. For example, Braque and Picasso have a 70% similarity score.

The “cosine similarity” between Braque and Picasso is 0.70.

ResNet-50 is a convolutional neural network (CNN) introduced in 2015. Normally, we would use it as the base for image interpretation, and then add layers to learn the specific application. In this case, we are only interested in the coding system it uses, called an “embedding.”

ResNet-50 encodes each image as a list of 2,048 numbers, known as a “vector” in machine learning. This vector is not simply a way to store the image – the JPEG file already does that – but to encode whatever features the model deems useful.

For this demonstration, I collected examples from fourteen artists. To avoid complications over the choice of subject, I used self-portraits by each artist.

Experiments with CNNs show that they recognize shapes, colors, styles, and textures – everything you would expect from “machine vision.” Our model is not going to know anything about the painters, though – not who cut off an ear, or who moved to Tahiti. It’s just the pixels.

With the fourteen paintings vectorized, we can do things like compute similarity scores. For instance, Braque, Chagall, and Picasso seem to hang together. I also ran a hierarchical clustering analysis.

It’s hard to imagine what the clustering algorithm “sees” in high-dimensional space so, wherever possible, I try to reduce down to three dimensions – using principal component analysis (PCA) or UMAP. In this case, because of the small sample, a three-D chart captures 40% of the variance.

The human eye naturally finds clusters – there are Picasso, Braque, and Chagall down at the bottom, and here is Kandinsky off by himself. Also note that Cezanne, Gauguin, and Schiele are spread out along the Y axis, but together on the X axis.

Unfortunately, these axes are completely arbitrary. ResNet-50 can’t tell us if Z is the “axis of cubism,” or whatever. That’s the knock against neural net reasoning being a “black box.” We can see, though, that the PCA plot roughly agrees with the cluster analysis.

So, that was about two hundred lines of code as a proof of concept, plus some fun charts. If you were really doing this for your MFA, you would want to use many more paintings, and stash them in a vector database. For more on vector databases, see Literary Analysis with RAG.

Today’s featured image nods to a common gaffe in generative AI. Yes, Marc Chagall really did paint a “Self-Portrait with Seven Fingers.”

Notes on GenAI

Generative AI isn’t really my thing. To me, AI means machine learning for quantitative applications, like predictive analytics. Nonetheless, people seem to be having fun over here, so I thought I’d give it a try. Here are some notes from Project Avatar.

Tools in this space overlap a lot. Synthesia, for instance, aims to be one-stop shopping for training videos and the like. The video tools overlap with the image generation tools, while others, like Eleven Labs, have carved out niches where they’re best in class.

Create Your Avatar

With HeyGen, I was able to create custom avatars from photos, video, and generated images. It also features prefab avatars, based on their own actors. For a training video, you can choose from a wide variety of these. If you want to be differentiated, however, you will need to create a custom avatar.

I know of one company that simply adopted “Ada” from Synthesia, and made her the face of their application. Ada is a very popular avatar, so – no differentiation. With Synthesia, you can create custom avatars from video, but not photos.

Avatars from Still Images

Video avatars are generally more expressive, so why would you want to create one from still images? I can think of two reasons. First, making a good video avatar is a lot of work, and you need a model. I did one of myself, and it looks like hell.

Second, you might want to create a completely unique persona – to specifications – that you control. This is how I created Hadley, and here is where it gets interesting. Midjourney has a character reference flag so that, once you have the face you want, it can reliably reproduce her in different settings.

Close shot of a 52-year-old man, strong-looking, bald, square jaw with short beard, captured with a 70-200mm f/2.8E FL ED VR lens, with high-key lighting and a shallow depth of field.

Stability AI is arguably better at images but, without the cref feature, it is useless for this application. On the other hand, while Midjourney is great at creffing AI images, it won’t do photos. The image of me in Project Avatar is from HeyGen.

After training on a dozen recent photos, HeyGen can reproduce my likeness accurately about 25% of the time. These are called “looks.” So, “Mark seated in library wearing a dress shirt” is a look. I also did a video avatar, which included cloning my voice.

Synthetic Voice

Eleven Labs has a wide array of professional voices, plus you can clone your own voice. You can generate audio in Eleven Labs, and then upload it to HeyGen for animation. If you’re working from a script, you can paste the script into HeyGen and then link to voices in Eleven Labs using its API. HeyGen also has its own voice library.

Please write a short, motivational speech, roughly 75 words in length, on how to navigate life transitions. Include a narrative hook at the outset, and a punchy conclusion that recaps the hook.

Both systems are a little bit robotic when reading a script. There are some things you can do with the script to improve this. To get the best pacing and intonation, you might want to read the script yourself, and then use speech-to-speech conversion. This, unfortunately, will gum up your automation.

Automate Your Workflow

If you’re doing this at scale, you can’t be the one reading the scripts. This energetic bot is Jordan from LipDub. You can imagine the pace at which he is pumping out these ad spots. LipDub and Akool are both more expressive than HeyGen but, again, you need a live model.

All these tools are richly supplied with APIs so, if you are a non-coder, you can easily string together a workflow using Make (formerly Integromat). Here is my draft workflow for the Bruno channel:

  1. Call ChatGPT to generate topic and prompt
  2. Call Powerful Thinking to generate script
  3. Pass script to Eleven Labs and call TTS to generate audio
  4. Pass audio to HeyGen, selecting Bruno avatar by ID
  5. Optional: generate new “look” for Bruno
  6. Add captions
  7. Post to Instagram reel using Meta API

By “richly,” I mean that HeyGen wants $100 per month for their API – on top of the monthly Creator subscription. CapCut doesn’t have an API, but JSON2Video does. Also, there are libraries in Python that will do captions. Bold captions add a nice touch.

Step two of the workflow calls my custom GPT, Powerful Thinking. This, unfortunately, does not yet have API support, so I was reduced to using Selenium.

Write the Script

Scripts are the easiest thing in the world to generate, and the LLMs (apart from Powerful Thinking) are equipped with APIs. You can give them specific instructions, and they’ll even do product research for you.

I am a ChatGPT guy, myself (the other camp is the Claude people) and I have also used Google’s Notebook LM. Hadley’s libertarian scripts, I write myself.

One last caveat: GenAI is a rapidly-evolving space. Synthesia had it to itself for a while, and now there is a raft of new entrants. Eleven Labs faces competition from Murf and Speechify. I am on a list for the LipDub beta. If you want to work in this space, you must be ready to learn a new tool every week.