AI Engineering 2025 PLAN: Max out AI COMPUTE for o1 Preview, Realtime API, and AI Assistants

in this video I want to share my current vision of the future of engineering with you this is my plan for how to win in 2025 in this Vision you and I wield generative ai ai agents and AI assistants to gather data synthesize it and build at rates that are completely absurd by today's standard there are four key advancements in AI tooling that occurred this year that's unlocked the path to Next Generation AI engineering Sonic 3.5 structured outputs 01 reasoning models and of course the real-time API before we discuss the future of engineering let me

share version three of ADA the personal AI assistant for engineering we've been building up on the channel hey adah hi there how can I help go ahead and pull in our SQL tables into your active memory tables loaded into memory create a markdown document and inside of it just lay out our table definitions sure I'll get started on the markdown document markdown document created n shaba go ahead and select all of our products uh order descending by Price SQL query save to file great AA go ahead and execute that and uh generate a CSV file

for US query executed results saved as CSV a generate the SQL for our top 10 products with the highest average rating create a CSV file for US top 10 products with highest average rating save the CSV wonderful Ada can you go ahead and generate a python let's do a bar chart and on the x-axis use our product name on the Y show the average rating please python code for bar chart generated it's saved in top 10 highest average rating products bar chart Pat ay to go ahead and delete that bar chart.png file and the python

generation code Force delete that both files deleted get our top 10 products by order volume sort descending create a CSV file for us top 10 products by order volume saved as CSV wonderful Ada go ahead and create a python chart let's do a bar chart of the top 10 products here based on our CSV file python code for bar chart generated it's saved in top 10 products by order volume bar chart. py wonderful thanks ad up you're welcome so the personal AI assistant you just saw is powered by the real-time API structured outputs and the

01 reasing model it showcases a glimpse of How High performing Engineers will maximize their output with compute in 2025 and Beyond Ada and the personal a assistant are only possible du due to key AI tooling breakthroughs that happened this year in 2024 the word on the street is we're going to see one additional release occur as well likely from anthropic but we'll see about that I think whatever comes out we've already had the key breakthroughs we need to redesign and rethink how we engineer in 2025 so why are these tools important what do they unlock

they allow us to continue upleveling our generative AI composes the pieces that when you stack them together lead to the next level of software engineering right we have prompt design we have ai agents we now have the AI assistant which is effectively our orchestration layer it allows us to control our AI agents eventually that will lead us to fullon agentic where we'll build an own self operating software over the past two videos we've been digging into the AI assistant thanks to the real-time API this is a massive unlock for in parallel engineering where your AI

assistant can work for you on your beh behalf while you're accomplishing other tasks it's a work in progress there are many things to improve but it absolutely showcases what the future of engineer can and will look like right now you myself and other Engineers that watch this channel are likely few of many that realize when you put these breakthroughs together and when you start composing your generative AI pieces from prompts to AI agents to assistants you start building software in a brand new way as an orchestrator as a manager and Commander of compute this is

going to be a big theme of software engineering in 2025 if you're looking for those outsize returns and if you're looking for Next Level productivity as an engineer The Prompt is the new fundamental unit of knowledge work never forget this and yes you heard that here first let's start from scratch what does it mean to be an engineer right this is you the engineer and the center what do we do as Engineers we take data in we ingest data and then we synthesize it into some form of output right some artifacts this is software engineering

in its most fundamental form I've stripped out all the details here this is what we do we take an information and we synthesize outputs let's dig a Little Deeper right let's go a layer into this so data for engineers is typically databases code bases documentation blogs Trends news research tools it's your job to ingest this information and for your personal work for your tools for your career for your job is your job to then synthesize everything coming at you into useful output let's break down output this is going to be code information and research media

content and products these are the top four of course there are many other things engineers can build and synthesize there is some overlap here but you get the idea right your job as an engineer is to ingest these types of content and then synthesize code information research media and products if we simplify what it means to be an engineer this is what we do but now things have changed so as you know there are hidden layers here in this process for how this ingestion and how this synthesizing actually happens now we have ai tooling so

let's break down the AI tooling we can now use to ingest and synthesize information okay we now have prompts this is the new fundamental unit of knowledge work if you master the prompt you will Master knowledge work full stop what's the next level at the next level we have AI agents AI agents are prompts and logic combined with data to solve a specific problem if you've been using the latest generation AI coding tools you using one or more AI agents right that is a tool that wraps logic code and UI on top of what the

base level right the llm the entire Revolution that's happening right now is powered entirely by large language models these are our text models these are our vision models this is a generative AI power house if the llm did not happen nothing else happens nothing else comes without the llm nothing else comes without the prompt the AI agents compose one or more prompts to solve a specific problem in a specific system at the next level we now have your orchestration layer this is what we've been working on on the channel over the previous two videos right

this is Ada this is our personal AI assistant it doesn't need to have speech to speech capabilities it's just a killer fast interface for our orchestration layer the the key differentiation here is that your AI assistant allows you to orchestrate many AI agents across use cases right and that's what separates this layer from previous tools a lot of the tools you're going to see are built with specific use cases built to solve a specific problem that's great that's still a wrapper around several AI agents the orchestration layer or your personal AI assistant is all about

helping you do what you need to do across all the domains that you operate in and then at the last level we have a g so these are fullon what I like to call living pieces of software that operate on your behalf while you sleep a while back I put out a video called the two-way prompt and you know at the end I described this new piece of software agentic software where um you're no longer prompting your AI it is now prompting you for what it needs to do next right it's gotten that good so

this is the last stage this is where we're going on the channel this is our North Star let me be totally clear this this is you know a 3 5 10 year goal and journey this isn't happening anytime soon this is insurmountably nontrivial okay so how does that change things right how does that tie to my plan for engineering in 2025 uh it's really this simple now you should be asking yourself given all of my tasks where I'm ingesting and synthesizing how can I use generative AI to help me do these things faster better or

cheaper or all three because that's where the real productivity gains really come in and just to throw some rough estimates on here utilizing tools that give you access to language models is you know relative to each other you know that's the first step right that's your initial 2x when you start using specialized agents to solve problems rapidly over and over that's going to get you to your you know your next huge incremental bump right and I don't actually know what the scale is this is just the way I've felt it over time so far right

this is your 5x now this 10x is something I'm still working on right I literally just showed you a version of ADA that gives us an orchestration layer on top of any set of AI agents okay so this is our orchestration layer I'm really excited about this the real- time speech to speech plus reasoning models plus structured outputs really enables this right for me the vision is complete for the orchestration layer and then there's the last level and and you know I don't even know if uh 10x from here is the right amount if it's

100x um it's probably much higher than that when you have this fully autonomous being this fully autonomous piece of software operating on your behalf it's probably much higher than this we're saving that for later this is a more faded outline for a reason um that's the North Star it's always to keep the North Star in your mind but realistically we're focused here 2 5 10 these are all attainable productivity gains you can get right now it's all about asking the question during my ingestion tasks when I'm looking through databases writing queries when I'm reading code

trying to determine how to modify this feature or add this new feature to this code base or looking through documentation or gathering information for an architectural decision you need to make right it's now about asking yourself the question how can I best use one of these layers of AI tooling to help me ingest and synthesize that's really what it's about let's look at a concrete example right the most obvious example is code generation this is the lowest possible hanging fruit in the age of generative AI for software engineer it's definitely one of the most important

but it is so overhyped it's so insanely overhyped right now and what are examples of this right now everyone's using cursor everyone's using ader continue Zed it's a really hot space right now and you know this is the you know 5x so if you're using AI code tooling right now you're in a great spot right because you know what that feels like to be able to write a prompt and generate massive amounts of code so that is a concrete example of using AI tooling to both ingest and synthesize information code generation is incredible but it's

only a small piece of the puzzle as a software engineer you just saw me working on a mock doc database instance with Ada we able to quickly read ingest and write SQL statements with artifacts in real time very very quickly um you know you saw that all happen just now right and and that's that's me tapping into this you know orchestration layer Advantage right I was still only effectively using these closely related agents all closely related to um SQL generation right like if we uh crank open that code base you can see I've got the

tools here separated by use case and so you know we used all the SQL database operations we also used the AI chat history management tools and a few other ones here right we use create an update file of course you know what we're doing here is we're starting to tap into these these cross domain actions where it's not just about writing an SQL statement it's about reading tables it's about writing documentation on the tables it's about understanding the structure of the database it's about generating multiple versions it's about generating diagrams and documentation right software engineering

is so much more than just writing code this whole anyone can write code narrative is so overblown it's very true anyone can generate code with AI but can you build and maintain software that actually produces value for a user that's a whole different story right that's a whole different story story coding isn't everything and that's why it's important for you the engineer to understand that you need to match up your ingestion tasks with the highest composition level of AI that you possibly can and then do the same thing when you're synthesizing results right and actually

outputting you know your artifacts and your output and your raw content so what is my plan in 2025 given this setup my plan is the same as I just described I am looking through all of the cases and all the situations where I'm ingesting and synthes exing and then I'm asking myself how often do I do this and how important is it that I have this problem solved at record speeds record time faster than ever with generative Ai and the more important it is the more you need to push the problem up this composition chain

if you generate documentation once a month you're probably cool to just open up chat GPT run a random ad hoc prompt and just get the job done but if you're building new features and your architecting systems on a weekly or daily basis you probably want a specialized AI agent and more likely an AI assistant to help you accomplish that task but also closely related task right next to it right so you probably want to build or find an entire AI assistant with the right AI agents inside of it to help you accomplish those tasks at

a rapid Pace right and that's what it's really all about what operations consume the most of your time and based on that answer you you should be pushing that up this generative AI composition chain if it's a one-off task open up Claud open up Gemini open up chat GPT fire a prompt off a CLI right it doesn't really matter but as soon as that use case comes back right it's just like typical rules of automation as soon as that happens I like three times as a pattern then you should be building a reusable prompt after

that happens build an AI agent it can be a script it can be a small tool it can be an entire application but when that problem comes again and again and again you should be thinking about automating it right and I've listed the most common operations here for software Engineers so it's a good place to start here we all have to interact with databases so you should have your SQL your om your data access layers streamlined by AI tooling right same thing with code bases you should be able to read find fly through code bases

with generative AI tooling you should be able to also consume documentation and understand and Gro documentation with AI tooling it's the same deal here right blogs Trends news research you should be able to filter out the things that don't matter and while there are so many things that don't matter there's so much noise in the world this section here is becoming really really important for engineering uh there's so much just garbage coming out right now and um you know just useless Echo chambering type information is really important to you know filter out you know get

good reliable channels of information so so cut up each one of those ingestion use cases see how often you do them the more often the more important they are you should be delegating these items out to your slew of prompts AI agents and soon hopefully AI assistant same thing on the synthesizing step right the most common things we do as Engineers we code we create information and research we generate media and we output full-on products right code and products is probably the most popular one for most Engineers Okay so these things need to be super

super locked in I probably should have explicitly added documentation but it kind of fits under information as well but you get the point right like these these are the first things to go after when you're trying to figure out what do I focus on what do I automate right and this is my plan for 2025 is to take all of my ingestion tasks and all of my synthesizing tasks and make sure that I can lean on generative AI tooling to help me accomplish that goal so I have three big things things queued up to share

before the year ends that can accelerate your engineering we're going to dive into meta prompting Concepts right we're going to start dipping our toes into what it looks like to get this 100x uh keep in mind that you don't jump between these productivity improvements you it takes a long time to transition and sometimes you're operating right out of chat gbt right getting the lower uh productivity gains and then other times you're going to use a specific tool or slew of Agents um I am going to be working really hard to get Ada off the ground

and make her as useful as possible for real software engineering because the gains here are just incredible so we're going to be digging into meta prompting I'll be releasing the AI coding course I'm really really excited about that I've been working really really hard for honestly over half a year now on this course this is going to make you a master of AI coding tools the tools of today and tomorrow I want you to win not just today but tomorrow next year the year after so I'm really excited to share that that'll be closer toward

the end of the year likely December and then finally we'll be continuing to push on the capabilities of Next Generation AI tooling with a focus on of course shipping software from AI coding to crafting powerful prompts to deploying AI agents and building out our personal AI assistant all the way up to the fully agentic systems right this is our Northstar that's where we're going on the channel if that interests you hit the like hit the sub join the journey we are going to build living software that works for us while we sleep stay focused and

keep building