What AI actually is so you don't have to fake it
A guide for CEOs and Founders for AI implementation
CEOs ask us consistently about AI strategy. Often we speak different languages.
Atmosphere (us) will do a multi-part series over the next 4 weeks to give you the language and use cases to make sense of it all, as well as some tutorials for implementing AI at your organization.
We will not cover everything. Niche experts go deeper in all of this. There is a plethora of click-bait content of ‘Be an AI expert in 30 days!’. We’ll do something different. A practical guide, direct from the operator’s seat. Tips you can actually use.
What is AI anyway?
It’s not a stupid question. Executives are all well informed, smart, read the AI news, but it’s still confusing. Today’s AI has morphed from the traditional textbooks of Machine Learning, Natural Language Processing, Computer Vision, and Generative AI. You’ll still hear these terms. But today’s AI in practice has coalesced around a smaller number of concepts.
Machine Learning
Traditional Machine Learning (ML) is very much alive. These are systems that learn from data to make predictions and decisions. They have wide adoption and are best when you have lots of historical data. For structured tasks, purpose-built ML models can be more accurate and consistent than throwing the problem at a general purpose LLM.
Dumping your data into a Frontier Model, like ChatGPT, and hoping for the best is a common (and bad) mistake. The systems, people, and techniques are different. There are other tools you should be using like R and Jupyter notebooks.
The operator question is whether the problem is better solved with ML or Frontier Models. Even that line is blurring because the models can run real analysis and code. The mistake is pasting data and trusting the answer rather than directing the model to run a specific ML analysis.
Frontier Models (LLMs)
All of the attention has shifted here. ChatGPT, Claude, Gemini. They generate new data, reason, produce text, create images and videos. They are increasingly bundled. When people say ‘AI’ in a meeting, 99% of the time they are referring to these models.
They are probabilistic, which is why they will confidently tell you the US was founded 357 years ago even as we limp toward our 250-year celebration. Then instantly backtrack when you point out the error. That’s hallucination.
They’ll also give you a different answer for the exact same question because they are non-deterministic (you might also hear the word stochastic). Again don’t worry so much about the definitions. We just want you to be aware of the vocabulary.
It’s why high-risk situations (legal review, accounting audits) almost always have a human in the loop. You shouldn’t trust AI to close your books autonomously or give you a precise legal opinion. I could make the same argument around software architecture and data, which is why people hate on vibe-coding.
Two concepts that come hand in hand with LLMs are training tokens and parameters. Training tokens are a measure of how much data was used to train the model. Order of magnitude trillions. Parameters are the resulting weights that give the model intelligence. A bunch of numbers organized into matrices. Order of magnitude hundreds of billions. Smaller models can be in the single (7b) or double digit billions (15b) if tuned to a specific domain.
Parameter count is still a good predictor of intelligence but it’s no longer the whole story. You’ll hear terms like Reinforcement Learning from Human Feedback (RLHF), Fine-Tuning, and Retrieval Augmented Generation (RAG). These are techniques that can help a well-trained model beat a bigger raw one.
Specialized Models
These models are trained to do a specific thing extremely well. Examples here are text to speech models (ElevenLabs), speech to text models (Deepgram) or OCR models that can parse a PDF or scan a receipt. Frontier models get the press. These models do half the actual work.
AI Inference
This is the act of actually serving the model to a user. Processing a prompt, generating a response. The labs combine this all. What you see on your monthly bill is the tokens consumed and the aggregate cost. Typically in cost per 1m tokens, segmented by output and input tokens (Google, OpenAI, Anthropic etc.). But if you’re building your own with an open-source model you need to think carefully.
Inference can be a huge driver of cost and latency. More simply put, you actually have to pay for the model and compute. If you forget everything else, just remember this is where your AI bill will actually sting.
Agents
Where the industry momentum is heading. It’s a lengthy topic that we’ll cover in a separate post. You will hear these words in conversations.
Harness: the wrapping infrastructure and evaluations
Skills: the specific capabilities and guidelines for what the agent can do
Tools: functions the agent can call (data fetch or CRM record updates)
Guardrails: the instructions of what an agent can/can’t do (no processing payment)
Model Context Protocol (MCP): the common standard to connect AI to data/tools
Agents can do things instead of just generate content which is why the industry is leaning this way. The terminology can vary a bit across providers so treat this as directional.
The alphabet soup decoded
You may hear all of these terms spat out in a meeting. You don’t need to memorize any of this. Each has a different underlying technology. Don’t spend the time trying to understand the difference between transformers, neural networks, or diffusion models unless you’re a tech leader. You won’t need it in normal operating decisions. Worse, when you use it incorrectly you instantly lose credibility.
The state of the art changes every day so by the time you learn it you’ll already be behind.
Large Language Models (LLMs): reasoning, code/content generation, ChatGPT
Text to Speech (TTS): converting text into speech, ElevenLabs
Speech to Text (STT/ASR): converting live speech into text, Deepgram
Text to image (T2I): take a prompt and synthesize a digital asset, Midjourney
Text to video (T2V): take a prompt and create a video, Veo
Image to video (I2V): take a reference image and generate a video, Runway
Video to video (V2V): adjust a video for lighting, character etc., Runway
Multi-modal LLMs: unified models across text, audio, images, and video, Gemini
Graphics Processing Unit (GPU): massively parallel chips for training/inference
Tensor Processing Unit (TPU): Google’s chips optimized for neural network math
Pre-training: the initial phase of training a model on tons of data
Post-training: everything done to the model afterward to make it useful and safe
Next up: how to start on your AI implementation. Subscribe so the next one hits your inbox!