The Signal Beneath the Noise

Years ago, I wrote the principles of behind the "machine teaching era", an era in which we create software using examples rather than coding, the book data loops layer down the 3 basic principles this new coding method was based on:

1.⁠ ⁠Models are data reflectors, we don't train models , we develop data.

2.⁠ ⁠The data developers (often referred to as data labelers or annotators in the deep learning age) are the intelligence source, without them there is no "I" in AI.

3.⁠ ⁠Intelligence is about continues learning, hence all AI applications will forever run in data loops with their human trainers. This is the only way true learning can happen. I would argue that while EVERYTHING has changed, NOTHING has really changed in the underlying principles. Machines still learn from human curated data, this is still done in training cycles and there is an infinite feedback loop of data: collect, label, retrain but the question we try to let AI answer has gone up with complexity, and what used be a "Hot dog / not hot dog?", is now "Explain to me how quantum computers work" and the data validator needs to have a physics Phd. So leap in application, complexity, demand and usefulness - yet the underlying principles are the same.

One of the changes happening is that pre 2023 only few specialized AI expert companies were training deep learning models, today - all companies and individuals are training an agentic model.

History does not repeat itself but definitely rhymes. I have not used AI to write this article, while at the same time I use AI around the clock for coding and everything else. The writing muscle needs training, putting AI into it now doesn't feel right, it will require two cognitive loads at the same time and these days it's all about managing cognitive load, so going for quality over quantity here. AI will be used for proofreading, grammar and other at the end, but only for that.

AI Is Not an Intelligence Revolution. It's a Communication revolution.

I am a big fan of information theory and signal processing and my entire mental model around AI is that AI is an "information revolution"; communication shift, not intelligence. In my view, Intelligence, whatever that may be, is reserved for living things (with the potential for quantum computing to change that), allowing them to adapt to environment changes within the same generation acting as the software in the evolution process, bypassing the limitation of DNA-driven changes (the hardware in the evolution process) that adapt only on the next generation. Whether what we call AI is on the path for being real intelligence or not, the fact is that all today's AI can not adapt to change (without retraining), hence can not learn, hence is not qualified to be called intelligence. Naturally, these topics are broad so we'll narrow our scope to the pragmatic changes that impact most of us on our profession and workplace, focusing on knowledge work, a fancy name for the work we do using a computer.

Unless you are living under a rock, if most of your work is done using a computer, you are now constantly being told you are about to be replaced by AI, you have to level up your capabilities and your team's capabilities, and the organization is under existential threat (unless you are in the business of AI infrastructure) and you must become "AI Native". The change is real, the hype isn't. I have been (and still am) struggling to put clear lines and definitions on the limitations and boundaries of the technology, they are clear and blurred at the same time, PhD brilliance mixed with stupidity, deception, and laziness, all within the same context window, available 24/7.

As the models saturate and the ecosystem matures, noise levels are beginning to drop and clear signals start to emerge. We are all part of this mass, unprecedented exploration effort. We will be able to tell what is real once investment starts to demand returns. Right now we can divide the markets into three groups of companies:

Token producers: from Nvidia to OpenAI, Anthropic and Google.

Token consumers: the rest of us.

B2A & AI Apps: The newly born agent economy - we'll discuss in a separate article.

Token producers or generators are having their moment in the sun right now while driving existential fear into the rest of us to keep those tokens flowing. Yet at some point everybody will have to face the question of profits and ROI; this will also be the point in which we would be able to screen hype from reality, with some organizations experiencing severe token withdrawal symptoms along the way.

Adapting to the Chaos

How do we work and operate in such a chaotic environment as individuals? teams? organizations?

The saaspocalypse is real and we are going to change the way we work completely. There will be winners and losers and being AI Native is critical for both individuals and teams, yet, if you adapt to the change you will survive and thrive. Being able to adapt to change is a part of what it means to be intelligent. We must be pragmatic to adapt.

The Communication Shift

Before we can talk about the adaptation, we should talk about the shift. As I mentioned before, understanding the impact of "cheap and abundant intelligence" is a much harder mental model than understanding it as a communication shift. While both are hard, the latter is an easier mental model to reflect and create useful analogies with.

We (I assume you are one as well, if you've reached this point) are called knowledge workers since "all we do" is transferring pieces of information back and forth between ourselves and computers using text and clicks, and what is SaaS if not a fancy way to fill rows in a database, logging a piece of information for later usage and recall.

Think about it for a moment: we have optimized the enterprise world with buttons, tickets and forms, tables and dashboards, all organized with menus and submenus (and you should multiply it with 300, the average number of tools a company has) all that so we could just encode (type) or decode (read) a piece of information from a database. Many systems and services, fragmented and misaligned, leaving us the role of "reality alignment" through endless meetings. We don't like it, chasing tickets was never big fun whether it's CRM entry or bug reproduction instructions. These activities took so much time (for some reason you always need a meeting to make sure tickets are updated) that we actually consider this to be THE JOB instead of the actual value we deliver. In reality, the job value is to close that deal and fix that bug.

The information flow optimization does not end with our tools and environments, our entire organizational structure, management, hiring, training, compensation, all is built around the need to effectively communicate goals down and report status back up, from the 1x1s to board meetings, consolidate status up, correct course down.

AI changes both of these in a fundamental way, leaving us to focus on WHAT information we prioritize to pass and WHY. For too long we have focused on the "WHO" we deliver the information to and "HOW" we deliver it. In simple terms, we need to start thinking about value and outcomes first, starting as individuals and teams. We can let the agents handle the rest.

Welcome to the Jungle

Focusing on value is actually much easier than it sounds, given that we no longer need to handle the effort of carrying the information manually, translating it to machines using a mouse and keyboard and doing it fast enough to keep up with the daily world changes, we can now focus on INFORMATION QUALITY (and cost, but that's for later):

1. Fast - can we make this customer call complaint end up in a released bug fix within an hour?
2. Accurate - Is this the right thing to do now?
3. Up to date - Is it the latest information we have on the matter?
4. Complete - Do we have all the information impacting the matter?
5. Valid - Is this information still relevant given the context (change)?

The biggest enemy of information (signal) is noise, and boy do we have plenty of it these days. The basic information feedback loop every business has is:

The biggest source of noise is the market. The constant, rapid, and meaningful changes in technology, competition, regulations, geo-politics, and finance are overwhelming, making management's job extremely hard, which is to FOCUS the organization on winning the biggest winnable bet. In such situations management tends to hedge itself with several options, sending mixed and interleaving priorities (signals) into the org, to the mid-level management, responsible for translating the strategy to plans and deliverables.

The mid-level management faces three types of noise sources: the market (they have room and are expected to perform local adjustments), the management strategy and guidelines and the teams' "what's working for us?" feedback, which itself changes on a weekly basis.

Finally, the teams and individuals that are constantly experimenting, overwhelmed (and excited) all at the same time from the opportunities, challenges and threats this new technology brings.

With so much noise around us, the best approach is "keep it simple" and stick to first principles. So keep it simple. Whether you are an individual or an organization, define value, track it, see it grows. Ignore the Tokenmaxxing, Tokens are how, not why.

Clearing the Chaos: First Principles

The first thing you should have in mind is that you have much more time than it seems. Pause and think. Token producers are constantly bombarding you with fear and pressure and with stock prices P/E multiples, investments and mega disruption risk they face a different kind of pressure that motivates the cycle, yet there is a simple fact, your cost is their revenue so take that into account. Ask yourself if your revenue got 3X this quarter, because your token consumption likely has, and it is expected to do so again next quarter. Most of these tokens are burned with no returns, their biggest return is actually agentic training for your team, allowing them to transform and adjust to these new tools and technologies. Expect, budget (and even demand) token burn rates, but sooner rather than later, make sure ROI comes with this cost.

This experimentation cost is critical and important, but it is not well spent if goals are not defined and results are not measured and tracked. It is common nowadays to hear "this used to take days and now minutes" and immediately think of all the value we have just created. In reality, there are two ways to create (shareholder) value, increase revenue and cut costs. Both can be summarized as producing more with less, whatever you are producing, you should be measuring it and focusing on the AI impact of this production - more units with fewer people and resources. Ask yourself, What am I producing? features, leads, deals, closed issues, meeting regulations and tax reports are all valid answers, as long as you can trace it to cost or revenue. Unused features are liability, not assets and poor quality leads are time waster, not revenue opportunity. Quality matters, more than ever when AI slop is free.

The experimentation is not bound to a team or role, given we want to do more with less, we expect roles to merge. In reality, it's rare for a role to disappear completely, but tasks within roles are disappearing, whether it's allowing product managers to code or allowing sales to practice marketing, flexibility is the way forward. Management's job is to open and allow maximum information flow, encouraging employees to expand their impact and reward winners - people, tools and methodologies. Naturally, this comes with potential friction, which is another focus area that management should pay attention to with the average mid level manager AI transformation priorities is:

1. Facilitate zero to one wins - successful experiments.
2. Scale success horizontally to more teams and more units.
3. Manage the resources along the way - allocate more to success on the expense of failures.

Lastly, keep things super simple. The amount of tools and technologies is exploding; most are not relevant next week or month. With a clear mind, some common agent CLI or UI, markdown files and file storage can take you very far - more tools usually means more complexity and more noise. Have no worries; you will be amazed with the amount of complexity you can reach with markdown, git and terminals. Yes, you (your agent) should be using git. Have no worries; you will not care most of the time. Git is part of HOW information flows, and most of us don't care about the HOW anymore.

Data & Context: "Easy" Win

Naturally, I am biased since our product revolves around efficient scaling of success: How to supply high quality data to agents, a problem that is essentially incomplete and can never be solved 100% (more on that in future articles).

You will see many data context and memory management solutions out there with all having the same blindspot - Agents alone can not make progress, Agents are amplifiers, not generators of outcomes.

The key in agentic data management is how do you integrate the human (expert) in the loop, the human touch contains magic and your question is not how to automate but how to amplify, the human generates the information (create), the AI distributes it (copy and scale).

Human in the loop was and still is one of the top challenges of AI and now every one of us becomes a human in some AI data loops, this brings another major data context challenge - Privacy in the workplace needs to go for effective human-agent work.

Over the last 50 years we have had some privacy in our work (some might call it rightfully the illusion of privacy on our work), especially in corporations. For example, most employees will assume their manager will never read their emails and probably in many organizations this is correct unless an extreme or illegal condition took place.

In the old world this did not matter, since we carried information manually, manually communicating your inbox over calendar events (meetings) was the norm and a piece of information could travel for days, weeks and months until absorbed in all relevant places. AI makes information move at the speed of a token, often million times faster than the old ways. The agents should not wait for you to open the inbox to respond. This is a complex matter and requires changes in culture, tools, and methodologies, yet comes with guaranteed AI progress. Let the information flow, work with the team to identify the boundaries that match your needs, and constantly revisit your org's information speed. This is a good way to measure AI impact and "AI nativeness" of your team, once you are happy with token burn rates (and start working on cutting them down).

What's Ahead

As we and everything around us changes, navigating the change is what matters. The number of challenges ahead of us is significant and unknown, from HR to security, the impact is all around, with most of it being revealed as we peel the AI onion both personally and as teams and organizations. The organizations that treat AI like a communication problem, not the intelligence problem, will adapt the fastest and win the race.

If you want to learn how to make this adaptation and transform your team to be AI-native we'll be happy to talk and assist you. Feel free to drop me a personal note: eran AT langware DOT ai (we used to do that before agents to hide emails from spam :) or book a demo here.