When Marc Andreessen wrote in 2011 that software was “eating the world”, he foresaw a worldwide substitution: innovative companies that used software replacing incumbents in industry after industry.

He saw early signs that companies that were providers or early adopters of Silicon Valley-developed software would outperform their peers. In his analysis, he extrapolated this trend to the world, and in 2026, it is obvious he was right.

Today, the question is no longer whether software would affect all industries (it did), but how AI will shape them this time around.

So let’s take the same approach of looking at the early trends and extrapolating from there.

Starting with the low-hanging fruit

When ChatGPT came out in late 2022, it quickly became clear to its early adopters that this was more than a toy, but rather another technology that had the potential to touch many industries.

In early 2023, I predicted that by the end of 2024 most knowledge workers would have a digital assistant. That happened: By October 2024 every Google search result included an AI summary, and ChatGPT reached 300 million weekly active users. Today, that number is well beyond 900 million.

While chatbot-like use of AI is already profoundly affecting private lives and the day-to-day tasks of people in many white-collar jobs, it is just the tip of the iceberg of a more profound change that is currently underway — so far mainly visible in a single industry.

What is happening in software engineering?

While pretty much every business software tool nowadays ships with a chatbot, most of them are just gadgets. You type a question, the chatbot answers, you close it. Nothing changes in your project. You get nothing you couldn’t get from ChatGPT. When people share “magical” workflows on LinkedIn, they are usually a concatenation of hopping between tools or automatic repetitive tasks.

However, since December 2025, there is one industry that has shifted far more profoundly than the other areas: Software Engineering.

So, what is changing in that industry that is not happening anywhere else?

Let me explain:

It was already clear early on that context-awareness, speed and computational abilities would be necessary properties of any useful “co-pilot” for real-world engineering.

Concretely, this means that the AI does not “wake up dumb” or need re-explaining every time it starts but can quickly pick up from existing work and that it can do work that is meaningful in a real-world work context instead of relying on the human to insert their knowledge and copy-paste data.

The distinction is simple: does the AI act on your data inside your work environment (Agent), or does it just talk about it (AI Chat)? When agents work and use tools on the user’s behalf, they turn question-answering machines like ChatGPT into actual collaborators.

The first place this happened at scale was coding agents: released only in February 2025 as “Claude Code”, and since then transforming the software world as adoption has gone through the roof.

Today, these agents code entire features, review and ship them while humans are asleep.

Software engineers are still in charge of deciding what to build, but more and more of the actual execution is shifting from humans to machines. These are not minor changes to the way people in that industry work. It has gone well beyond reformulating emails or writing memos; it has changed the entire way every software engineer works on Earth in the past nine months.

Why has all of this almost exclusively been happening in a single industry?

Why was coding the first frontier?

I believe there are three main reasons why software development has been the first (and for now arguably only) industry that the big AI labs focused on:

•The people building the AI models and the tooling around them are software engineers themselves. It’s the domain they know best, so they can iterate quickly and build great products for themselves first.

•Some AI labs believe that if they make coding better, they gain a major competitive advantage, since they can then build better AI faster using those same tools. Strong incentive.

•As Andrej Karpathy noted, when the correctness of a task is easy and cheap to verify (which is exactly what automated tests and compilers do for code), it is easy to optimize an AI to perform well at that task.

That last point especially drives home why software engineering has been such an easy target. Code is the only engineering artifact that is natively machine-addressable and lives in a single place; it is “only” text (which LLMs understand very well), but can be executed and tested, so its outcomes can be automatically verified.

What lessons and tools can now be reapplied across other industries?

While this explains why the first wave of agentic work was directed towards software engineering, the interesting question that arises now is in which areas the lessons learned and tools built for software engineering can best be reapplied in a second wave of transformation.

Here are some criteria:

•Most advances in the past three years in AI have been made in Large Language Models (LLMs) that can input and output text extremely well. So, the methods and tools developed in software engineering will be most applicable in tasks that rely heavily on complex text.

•Agentic workflows work best when the AI agent has direct access to a centralized repository that contains the context of what the user is trying to accomplish. In coding, this is available in specification files, comments or even the code itself.

•The AI agent needs to be able to actually interact with the data (traverse it, make suggestions, etc.) using the so-called “harness” to assess the potential impact of changes and propose them autonomously for the user to review and confirm.

While it is immediately obvious that industries such as marketing, legal, translation and others that fulfill those criteria will be disrupted next by reusing the methods and tooling invented for software development, I want to focus on what is next in the world of engineering.

Why hardware is next in engineering

The area in hardware engineering that comes closest to coding is Requirements Engineering (RE):

•Requirements describe in text form the product the engineers are trying to develop.

•Requirements Management Systems are central repositories that contain the logically connected context of the entire design of a product across disciplines.

•A well-connected AI agent can navigate requirements across multiple hierarchies, connect to their implementation and verification, and based on this, assess and propose meaningful changes.

This is the area in which an AI agent can assist engineers as a colleague. It allows hardware engineers to make the same leap that software engineers made in the past nine months:

•Software engineers went from using a “nice-to-have” chatbot that, with limited context, sometimes autocompleted some code, to defining high-level functions and seeing AI agents plan, code and test them end-to-end.

•Hardware engineers will go from checking and reformulating individual requirements to discovering and aligning the detailed design intent that can then be developed and manufactured in much quicker iterations.

In requirements engineering, the requirements are the context, and the correct setup makes them available to agents to work with.

Experienced engineers can then intuitively judge whether a product is well defined, and in the near future, simulation might even become the hardware equivalent of compiling. You can simulate with significant accuracy whether electronic circuits work, how mechanical loads are distributed, how heat and fluid flow behave. Then, formalized design rules can automatically verify whether a design meets its specs.

One can imagine a future where most hardware design is verified automatically and only higher-level functions are checked by hand; much like manual QA still exists in the software world.

What I predict

Three things changed in the last twelve months, and together they make agentic hardware engineering real.

First, models got good enough at structured editing that they can be trusted to propose changes inside an engineering tool, not just talk about one. Tool use is no longer a novelty.

Second, hardware engineering tools are exposing their content as structured data. This is boring infrastructure work that’s mostly invisible from the outside, but it is the precondition for providing context to agents that can then build on top of it.

Third, we now know what coding agents are capable of once you connect a tool-using AI model to the right harness: developing entire systems under the guidance of experienced engineers.

So my prediction:

By the end of 2026, the engineering tools for hardware product development will ship the first useful agentic capabilities that let engineers make high-level decisions on requirements and have agents break them down for engineers to approve.

By the end of 2027, the race we have seen in coding will have broken out in hardware too. Tool vendors will be competing to offer solutions that let engineers oversee the entire development cycle from early requirements engineering to assisted implementation and automated verification.

By the end of 2030, the most innovative companies will be closing the loop between theoretical engineering and iterative production, making sure the designs that the AI agents help develop actually translate quickly into physical prototypes that can then be tested automatically and from which the lessons learned can be funneled back into the AI-assisted development cycle.

A morning in 2027

Picture two engineers at a mid-sized hardware company developing, say, a new industrial sensor module.

One is a requirements owner — a senior engineer accountable for the sub-system requirements of the sensor. The other is a design owner — an electronics engineer responsible for developing the board that makes it do those things. Today, these two people spend a depressing fraction of their week translating between each other’s artifacts. In 2027, they won’t.

The requirements owner sits down on Monday morning. Overnight, a colleague two levels up adjusted a handful of top-level system requirements — tightened the operating temperature range, added an EMC clause, and relaxed a power budget. Her queue is already waiting…

The agent has compared the new parent requirements against her sub-system architecture tree, identified eleven sub-requirements that no longer line up, drafted concrete edits for each, and grouped them into three coherent review batches for software, thermal and PCB aspects. Alongside each batch, the agent has flagged the test procedures that would need updating and the verification-matrix entries that would shift as a result. She approves two batches in fifteen minutes. The third she rejects in one sentence, because the agent missed a regulatory constraint; the agent redrafts on rejection. Nothing changes in the system until she clicks approve.

Changes that used to take weeks to ripple through her sub-system tree had already been resolved before she sat down.

Across the floor, her electrical design owner opens his ECAD environment. Alongside his schematic, the live requirements his board has to satisfy are right there. Not a PDF export from three weeks ago, the live ones. When the temperature range tightened overnight, his agent already flagged which components are now out of spec, drafted a note explaining which alternatives would close the gap, and listed the bench tests that would need rerunning.

He spends his morning on the two changes that actually need his engineering judgment, instead of on the twenty that don’t. For the first time in months, the afternoon is for design.

Neither engineer has been replaced. Both have been promoted from individual contributor to lead of their own team of AI agents. The busywork, the cross-referencing, the bookkeeping, the polite chasing of stakeholders — all delegated to agents that never get tired of it. What’s left is the part of the job that made them want to be engineers in the first place.

Software ate the world. Agents will redesign it.

Andreessen’s 2011 thesis was about who builds the world’s products. Software companies, he said, would build them, or buy the ones that did.

The agentic shift is about how those products get built. The next decade of hardware will not be defined by which company has the smartest engineers — every company has smart engineers. It will be defined by which company’s engineers have the most leverage.

And leverage, in 2027, means a team of agents working on your data, in your tools, surfacing the right suggestions at the right time.

Personal plug:

This is the world my team and I have been building towards at Altium’s Requirements Portal. And we have just launched Agentic Requirements Engineering.

Obviously not as a chatbot pinned to the side of the app, but as agents that work directly on your live requirements: reading the current tree, proposing concrete edits, applying your team-defined skills, and surfacing every change as a suggestion you approve.

The interesting part isn’t the product. The interesting part is how the job of an engineer changes.

Both are wrong. The interesting question was never “how much AI?” It is where and when.

Is AI good or bad for my team?

It’s clear that AI can help you move fast: At Google, 75% of Google’s new code is now LLM generated before being reviewed by an engineer. The ICIJ’s Pandora Papers surfaced thousands of passports from huge amounts of leaked documents using machine learning. And in Kenya 10,000 health related questions by pregnant women are answered daily by SMS. Most of this would have been science fiction in 2022.

However, it is also clear that AI can be stupid enough to delete entire production databases, make people publish summer book reading lists with made up titles, create fake citations in papers or think that it’s ok to make up court cases in your legal arguments.

The “AI Bowtie” Framework

To explain to my 30-person team what I believe good AI usage actually looks like and when to use it, I regularly draw this picture for them on the whiteboard:

Reading it from left to right: The vertical thickness of the bowtie at any point is the amount of AI you should be using at that phase. Wide means lots of AI. Narrow means none. Five phases:

1. Research · Brainstorm · Discover

Lean on AI as hard as you can. Have it summarize twenty papers, generate dozens of competitor pitches, draft customer objections you haven’t thought of, role-play five different personas reacting to your idea. The bigger and weirder the funnel of inputs, the better the raw material later. What AI cannot do here: care which of the twenty options actually matters to you.

2. Synthesize · Prioritize · Distill

Dial AI down. You still use it to cluster and summarize. Then you take over. The model can stress-test your phrasing, flag contradictions, list counter-arguments. It cannot decide which idea is the one you are willing to bet six months on. The closer you get to a single concrete answer, the more it has to be yours.

3. THINK! No AI.

This is the brain-only zone, the soul of the work. The one paragraph you would write on a napkin if you had to explain what you are actually doing. The software architecture you can draw on a whiteboard from memory. The market positioning. The company strategy. The feature scope. The continue-or-pivot call.

If you outsource this to an AI model, you do not actually have a strategy. You have a plausible average of one, generated by something that has read everything and committed to nothing. You neither have a mental model in our head nor will you be able to defend it when challenged, extend it when the situation changes, or argue for it in a room of skeptics.

The soul of your work is what you discuss with co-workers, hone and commit to.

Subscribe now

4. Plan · Prepare · Architect

AI comes back, but on a leash. Write the implementation plan in your own words first. Hand it to the model for critique: where are the gaps, what does it not cover, what would a critical reviewer say. Tighten it. Break it into parallel workstreams. The model is now a fast colleague who reads carefully, not a vending machine that gives you an answer.

5. Execute · Implement · Disseminate

Open the throttle all the way back up. Generate the boilerplate. Draft the docs. Spin up agents in parallel. Here is the part most people miss: spend the saved hours on the things that would normally get cut for time. Make the landing page accessible to screen readers. Write the documentation properly instead of leaving it as a todo. Add the integration tests you would normally skip. Localize. Polish the empty states. The whole point of holding the soul yourself is that you earn the right to be lavish here, and your output ends up better than the pre-AI version you could have shipped.

tl;dr: Explore with AI. Define the Soul of your work with your Head. Ship with AI.

Personal plug: Currently my team at Altium is working on making Agentic Requirements Engineering a reality (news coming soon). To make sure that we keep ourselves sharp in our thinking while accelerating our work, I came up with the above concept.

Definitions: What is an AI co-pilot?

In my view:

An AI co-pilot is a specialized software that you can interact with while you are working on an array of tasks and that assists you in accomplishing them by:

• the provision of context-aware information that you need and/or

• the ability to execute tasks for you that are

  • simple but time consuming or

  • require computational abilities you don’t have

Some history on co-pilots

The idea of co-pilots that assist us in everyday life is not new. The vision for a digital assistant in the workplace had been around at least since the 80s; such as in this vision ad by Apple:

An also in pop culture, assisted working became an ideal to be desired, such as J.A.R.V.I.S., who helps Iron Man design his suit, defeat his enemies and much more.

However, the implementations of co-pilots in actual software has been terrible for the past decades:

While pretty much every person older than 30 years remembers clippy as one of its first major commercial implementations of an assistant, few of us shed a tear when it disappeared.

Later Voice Assistants, such as Apple’s Siri and Google Assistant and and Microsoft’s Cortana never lived up to the expectations; being mostly used to set a timer for its users.

Why did co-pilots fail in the past?

So why should anything be different this time?

There are 3 key reasons why co-pilots failed in the past and where their visions differed substantially from what users could experience:

1. Not understanding users’ intentions

   

2. Not saving significant time for a given task

   

3. Never offering anything the users could not have done themselves

   

Solving these flawed aspects by using Large Language Models (LLMs) are the first 3 of 5 aspects that have changed the game around AI assistants:

What makes a great co-pilot?

1. It understands the user’s intent

There is nothing more frustrating that trying talking to a cold and soulless machine when it does not even “get” what you are trying to do. With the release of ChatGPT this barrier has been removed. While the art of writing good questions for AI (“prompt engineering”) is still a skill worth developing, a good assistant does not rely on specific keywords or sentence structures to derive meaning from the users’ instructions.

For this it needs to…

• …understand multiple languages / colloquialisms / domain-specific language etc.

• …have memory of your conversation to understand requests and additions in context

• …be able to ask concise questions back when user requests are unclear

Subscribe now

2. It saves significant time on mundane tasks

While the tasks themselves can be generic or completely domain specific, reducing implementation time by a factor of 10x makes all the difference:

Here it does not matter that the results might not be perfect or could also have been achieved by a junior person that took the notes; the fact that you have that result instantly at your disposal makes all the difference:

A good AI co-pilot feels like having a capable intern who assists you instantly.

It will make mistakes; its work won’t be necessarily on par with yours but it produces drafts for you instantly and with a bit of guidance lets you achieve your goals much faster.

3. It can do things you are not able to alone

While having an assistant that does things 100x faster than a human enables you to achieve super human results yourself; a great co-pilot goes beyond pure efficiency gains:

A great AI co-pilot makes you not only more efficient, but more effective.

While e.g. spell-checking, summarizing, pre-formulating drafts, etc. save you time and allow you to perform your tasks much faster, time savings are only the tip of the iceberg of what assisted work will look like in the future:

The real power of an AI assistant is when it helps you with complex tasks, such as:

• planning

• forecasting the implications of actions

• decision making based on data

• managing dependencies

While the underlying best technological approaches that will drive these abilities are still in their infancy (for the technically inclined, check-out e.g. Retrieval Augmented Generation (RAG) and Autonomous Agents), the best existing co-pilots today (08/2023) already start guiding users towards their results.

As humans we are very limited in our abilities to retain and recall information and to hold to many concepts in our head simultaneously. This is why assisted decision making that analyzes inputs, suggests alternatives and simulates outcomes for the user, allows them to take better decisions. This ultimately enables the creation (depending on the applied domain) better code, policies, medical treatments or products:

4. It lives where users work

While only the future will tell whether we will ultimately all have a single personal AI assistant that will follow us around all day or whether depending on the context we will interact with different assistants; it is clear that the co-pilots we will interact with will assist us inside the tools we use daily.

This allows the co-pilot to leverage previous work, get relevant context for each interaction, have domain-specific abilities and ultimately help with the task at hand.

This also means that many interactions with the AI assistant aren’t necessarily through voice commands or full requests like “improve this code”. Since the co-pilot is able to understand your intentions, its actions come in multiple forms, such as auto-complete, highlighting, smart buttons, popups and other forms.

Some of the best AI assistance features many of us use daily are not even visible in the form of an “AI bot”, but rather have become part of very subtle user-experiences. Many of us hit “tab” multiple times per day when writing emails with gmail forgetting that merely 5 years ago this functionality seemed so mindboggling, it drew laughs and cheers from a major crowd:

5. It makes you enjoy your work more, not less

Great co-pilots make you feel in charge and with access to super powers.

As new generations of workers and experts in each field will take for granted the ability to collaborate with an AI assistant, many companies will fail on the way to provide value to those users as they will turn out to be nothing more than a fancy gadget.

However, the assistants that ultimativley will shape our daily lives will be the ones that allow us to focus on the enjoyable parts of our work: being creative, making complex trade offs and design decisions,

In Conclusion

Large Language Models that are used in tools such as ChatGPT have brought us the technology that early assistants from the 2000s were missing to become great co-pilots by translating user intent into something the computer can act upon and giving us a tool to save time on mudane tasks. This technology however enables the true potential to be delivered by specialist solutions that make us not only more efficient but more effective, by providing co-pilots that live where we work, give us superpowers and make us enjoy work more; not less.

Personal plug: 

Given my enthusiasm for great co-pilots, you won’t be surprised to learn we are currently gathering feedback for our Assisted Engineering co-pilot for hardware engineers. Give it a try!

GPT-powered startups/initiatives have to overcome 3 major obstacles to become mid- and long-term successful. This article discovers how you can analyze their use case along the dimensions of “Productivity Enhancements”, “Non zero-sum-game Value” and “Moat = Value from Context” you can classify any GPT-application into 3 different levels of potential success.

The “Three-Hills” Model

While for people who follow the tech news there is no doubt that GPT1 technologies have started to disrupt all white collar industries at an insane pace and will have turned most jobs into assisted jobs by the end of 2024; it is currently much harder to predict who the winners of this race will be. This article aims at giving CEOs, investors and R&D budget responsibility a framework to assess the potential of GPT-based plays.

To help me as a CEO evaluate the long-term success likeness of GPT initiatives2, I’ve come up with the following simple but powerful mental model:

While many applications of GPT are amazing and therefore at first sight sound like great business ideas, in my opinion only GPT-plays that surpass “Moat Mountain” are actual business opportunities in the mid- and long run.

But let’s start analyzing step-by-step:

Productivity Hill

Most initial value propositions of GPT applications today are those that increase productivity significantly. 

To be more precise: they help with tasks that fall into at least one of the following categories:

1. Speed Increase: Tasks that users could in principle perform themselves, but can now do it 10x faster in a GPT-assisted way. Examples are:

   1. Developers being able to write code fast by using the right commands and logic than looking up boiler-plate code in documentation and debugging themselves

   2. Marketing specialists, being able to write SEO optimized blog posts at 10x the speed they usually could 

   3. Students writing fast-writing essays about the french revolution instead of sitting through reading dry books and articles

   4. Looking up the best 8 things to do during your weekend trip in Lisbon, depending on your personal preferences

   5. Scientists being able to search, understand and dig into a corpus of millions of scientific papers to reach better conclusions on state-of-the art faster

2. Democratization of Abilities: Tasks that you yourself don’t have yourself the ability to perform but averagely capable people who are specialized at that not-too-niche job could. Performing them with GPT is extremely fast and free. Examples are:

   1. Writing a rhyming speech in the style of Shakespere for your best friend’s wedding that makes references to things that you have lived together.

   2. Review that long email to your boss for spelling, grammar, well-formulated arguments, tone, etc.

   3. A business owner being able to write a job ad in a descriptive and appealing way to get good applicants

   4. Consultants creating entire packages for small businesses including website coding and wording, optimized ad and marketing material texts.

   5. Formulation of grant applications with high chances of success for small research groups.

Now; before you go and start your SEO-optimiziation or Shakespearean wedding speeches startup, let’s evaluate those same ideas along the Three-Hill model.

All the above examples have cleared Productivity Hill. They all are “Level I” applications of GPT, meaning:

People will earlier or later adopt GPT technologies to solve problems on Productivity Hill and get significant value out of it. This will also significantly raise the bar for anyone offering these services professionally today.

Abonnieren jetzt

Tug-of-War Valley

The first challenge that any Level I application of GPT will face is the Tug-of-War Valley. If the application does only create value in a pre-GPT world but will be counter-acted in a post-GPT world through opposing GPT functionalities then they are not good business ideas.

E.g. While a single marketing specialist might be able to write 100s of SEO-optimized blog articles with GPT, thousands of marketing people doing that will within a few months completely break the ability of search engines to provide useful search results based on content. Search engines (or GPT-powered search capabilities) will need to find other ways of providing useful information to the user than by indexing mediocre SEO-optimized content. Therefore it will only take a few months until SEO optimization will play tug-of-war with AI enhanced search-engine improvements.

In Tug-of-War Valley most of the benefits of using GPT will be canceled out by an opposing force also using GPT.

Value Peak

The above immediately begs the question, which companies are able to escape the Tug-of-War valley and become a Level II company. The answer is: companies that provide value outside of an existing zero-sum game: Basically any mechanism that is not meant to act as a “proof-of-work” for a human (e.g. writing essays for job applications, writing online articles, personalized emails, etc.).

Many ideas won’t make it to Value Peak:

Looking at the above examples:

• Content marketing (example 1b) won’t work in the traditional sense anymore and cease to exist in the future. Therefore SEO article writing will cease to be a useful skill very soon and tools that help SEO marketeers either have to pivot or will also quickly become a fad of the past; whether GPT-enhanced or not. 

• Classical essay-homeworks (example 1c) will in a best case scenario either disappear from curricula and be replaced with more useful teaching/learning exercises or in a worst case scenario go into an arms-race with AI-detection algorithms for teachers.

• Wedding speeches (example 2a) will soon rely even more on very personal wit and humor than on the clever text form; at least after everyone heard the 10th GPT-generated best-man’s toast.

• Your boss might not care anymore about long emails (example 2b) that propose strategy changes, since a well written long mail won’t be anymore a signal for you having spent a significant amount of time thinking it through. In fact, she might stop reading emails altogether and just listen to a GPT-formulated summary of her inbox on her way to work.

• Job ads (example 2c) and applications won’t work the same anymore. Coverletters will be the first to become useless if they are written and also screened by GPT. But even descriptions and CVs will most likely not be spared.

• However the problem of distributing grants will get solved (example 2e) in the future, if they still rely on the written word, then the tug-of-war between writing and reading software will just worsen today’s state in which agencies already started to make the process meaningless.

However, any application that generates value beyond objects made for human consumption are able to reach Value Peak:

1. Real/Usable objects: whether it is a piece of software that has been coded using GPT that makes people’s lives better, a hardware product developed using GPT enhanced requirements, etc. Each of these objects has value by themselves and are not just means to an end, independent whether they have been generated with GPT capabilities or not

2. Experiences: like a piece of music being composed for your taste and pleasure, a personalized bed-time story for your kid or an RPG adventure you can play.

3. Solving personalized problems: like providing recipes for the ingredients that you have in your fridge, helping you prepare personalized worksheets for your students, working out the logistics of your dream trip or any other problem that only you have in a certain constellation

Value Peak applications are not a fad; they will not go away anytime soon; because there is a real demand for them.

Displacement Canyon

Companies and Initiatives that assist human activities and provide clear value in a post-GPT world can reach Level II. The activities that they assist don’t just short-cut human work but create value by themselves and will therefore most likely keep existing and be valuable in the mid-run.

However, even if these applications have been thought through with specific use-cases and UX/UI, they will constantly have to compete with “generic” GPT solutions (such as ChatGPT + plugins). Why should a developer use a dedicated coding co-pilot platform instead of using a generic text assistant that can also write code?

In the past start-ups would usually get asked by investors: “what prevents the big FAANG companies from outcompeting you if they dedicated their resources to solving the same problem you are solving?”. The analogous question now will be “why should anyone use your solution and not just be assisted by a generic GPT application”?

Companies that don’t manage to climb Moat Mountain will stay in Displacement Canyon and will constantly have to compete for their users with OpenAI and similar players; most likely losing more often than winning, since the abilities of these generic assistants improves rapidly.

Moat Mountain

The only way to get to the “Basin of Success” (and therefore a Level III GPT application) is to build up a moat that generic GPT-applications (such as ChatGPT) can’t or won’t compete on directly. 

Although there might be more in the future, today it seems like there are 3 key moat-categories that startups and initiatives can implement to escape the risk of being displaced by generic GPT applications:

1. In-Context & Collaborative Features:
   Generic GPT functionalities are amazing at assisting single users with their task at hand. However, any use-case that requires collaboration of multiple people (think figma, google docs, etc.) for reviews, commenting, knowledge combination, history-tracking, dedicated access rights, etc. require a dedicated platform. When these platforms become GPT-enhanced, they become more valuable to the user, but a generic GPT assistant does not (yet?) offer context or features beyond a conversation.

2. Gated Knowledge/Data
   While any useful Large Language Model has been trained on an enormous corpus of publicly available data, gated knowledge can vastly improve the outputs of these models for specific applications. That data can be one of the following:

   1. single-client specific data: e.g. company-internal code that allows the application to suggest internally developed functions and coding styles; all legal memos written at a law firm or similar3

   2. domain specific data: e.g. databases of non-public data, such as all hardware design specifications of thousands of engineers using an engineering software

   3. complex to parse data: e.g. biomedical data or music that needs to be specifically indexed and accessed to be of value through a GPT-enhanced interface

3. Edge Computing / Offline Use Cases
   There are many reasons why some GPT use cases would need to run locally: e.g. many people might feel more comfortable that a privacy-first personal assistant that has access to all your medical data, appointments, emails, etc. did not access or store the data in the cloud. Also use-cases that enable you to interact e.g. with industrial equipment or similar, might not need to be as powerful as ChatGPT to understand verbal instructions, but are probably better kept off-grid.

Companies with a business model that relies heavily on these differentiators will very likely not only be able to attract many additional customers through the GPT enhancements, but also be able to successfully defend them.

In Conclusion

There are 3 levels of GPT-companies/initiatives. Level I applications are useful but will disappear shortly, Level II will create mid-term value but are probably not good business ideas and Level III can leverage the technology while still providing a compelling business model.

Let me know in the comments whether you agree and which (surprising?) conclusion you’ve reached when analyzing ideas with this framework.

Personal Plug:

Given my thoughts around long-term successful GPT-powered applications, it won’t come to surprise that at Valispace we just released Assisted Engineering functionalities that catapult hardware developers into this exciting near-term future today.

Since developing hardware products is a real value add and collaborative features + gated domain-specific data create moat, we are confident as a Level III company to change hardware engineering forever.
Expect JARVIS-like hardware engineering to become a reality very soon.

Given that ChatGPT likes to invent and imagine things, introducing AI Assisted Hardware Engineering sounds like a recipe for disaster. Why use an unpredictable AI in the design of planes, cars and rockets; where being exact matters so much?

By diving into actual hardware development tasks that engineers have to perform regularly and looking at concrete AI generated engineering artifact examples it becomes clear that Assisted Engineering will lead to faster and better design of safer products; always as long as we understand its limitations and how you have to pair current technologies with math-capabilities to make engineers 10x faster but their designs still reliable and safe.

Predicting the future of Engineering

Background

As an aerospace engineer who has worked in the past on one of Europe's biggest Satellite Programs (MTG) and as the CEO of an engineering software company that has helped hundreds of modern companies digitize their hardware engineering, I am often asked about what the future of engineering will look like. 

While having discussed broad ideas in the past, December 2022 has been a pivotal point to shaping my opinion: I am now convinced that the future of engineering is assisted.

What changed?

At the latest with the release of ChatGPT, it became clear that Large Language Models (LLMs) are able to suggest, complete, improve and summarize text extraordinarily well. While thousands of articles have been written about the pros and cons and possible implications, I want to focus solely on how hardware engineering will be affected by this technological shift. 

In this article I will refer to engineering features that are enabled through AI and more specifically through so-called Large Language Models (LLMs) such as ChatGPT as “Assisted Engineering”.

Today: How we develop cars, ships, planes, rockets and most other physical goods 

If you are a daily practicing engineer in an industrial setting, feel free to jump to the future section.

V-Model and Documents

Hardware engineering in general follows the so-called “V-Model”. Without going too much into detail, the key take-away is that in principle it can be simplified into a 5 step process:

1. Specify: You start by formally specifying your design (“How fast do I want the car to be able to drive?”) 

2. Break-Down: then you break these needs down into requirements for subsystems and parts (“How much torque does the motor need to generate for that?”).

3. Design: Next you work on a design that can fulfill those requirements and is optimized for cost / quality / time (“Which battery do I use for the car and what is the distance the car can travel with it?”)

4. Verify: You now verify subsystems and components through analyses and tests (“How many RPMs do I get when I put the motor on a teststand?”)

5. Validate: In the end you validate the fully integrated product; ideally with a customer (“Let’s drive this car and see how fast and far it goes.”)

There can be dozens or thousands of engineers involved in the above steps; depending on size and complexity of the product involved. 

Today the development of products takes place in specialized software (CAD for physical design, Simulation Software, etc.). There have been approaches in the past decades to manually build up models (see SysML), but in the day-to-day life of most engineers in industry the communication between engineers and teams is mostly document-based (Requirement Specifications, Interface Documents, etc.).

How is product safety ensured today?

Safety from a product-design perspective is mainly achieved using 2 key mechanisms:

1. Products are developed according to standards. These are lists of requirements that describe properties of the product itself, the process of developing as well as the accepted methods for testing/verifying it. Sometimes these standards come from external agencies and sometimes these are company-internal standards.

2. In certification and qualification processes, the company has to prove that the product and processes used to develop it indeed have followed the standards.

When it comes to the certification process a rule of thumb is that the more dangerous failures of products are (think cars, planes, etc. vs. smart watches), the heavier the certification process is to prove that the design was done correctly.  

This means that while engineers are always responsible to develop good products, engineering as a discipline does not rely on single engineers not making mistakes. 

Today this means that engineers might make e.g. a mistake in their spreadsheet that calculates critical properties of a product, but through a process of design reviews and qualification tests this single mistake will be caught before it goes into production.

This is very important to be understood as we assess later whether besides spreadsheets and documents an AI assistant can also be used to derive critical engineering elements.

The Future™ 

Every engineering action will be assisted. What this means concretely is that the following processes won’t be done by a single person anymore, but by a person or team collaborating with an AI.

Let’s dive into what this looks like concretely with technology that is available already since the beginning of 2023:

Going from 0 to 1

Lots of engineering work consists in coming up with solutions to novel problems. However, staring at a blank page is as much a problem for creative engineering work as it is for writers of novels. It is much easier to start from a draft that you edit to your needs than to try thinking of everything yourself.

Assisted Engineering allows you to brainstorm within seconds a starting point to many potential aspects. Consider that you are e.g. tasked to design the landing gear of a drone. Instead of trying to think of all the aspects that might be relevant to this piece of equipment you can start discovering which aspects your solution needs to cover:

And suddenly we are ready to think about lots of aspects. Not only about the landing gear itself, but also to think about drag forces, failure cases, material-choices, shock, monitoring and risk management.  

These requirements are far from being good or final, but within seconds we are in the middle of thinking about the detailed design. The Results have accelerated us and we can now start diving deeper into multiple aspects:

And here we are. Within minutes we have reached the point in which we are able to consider and formulate requirements that are testable. 

Do these requirements need work and refinement, before sending them out to a supplier? Absolutely. Do the values need to be thought through and checked with other stakeholders? Definitively. Did we arrive here 10x as fast as we would have without Assisted Engineering? You bet. 

Retrieving existing information

Engineering work also contains a non-trivial amount of trying to find the correct information that is needed at a given point for decision making. These may be values from datasheets, requirements, verification methods, etc.

Structured engineering data, such as in PLM and EIM tools is already helpful, since data exists in these tools in machine readable format instead of PDFs and other documents. However, Assisted Engineering capabilities will allow engineers to ask questions to retrieve data easily, both when interacting alone in front of a PC or when enhancing a conversation in a meeting room.

Imagine that an engineering team is considering using a Raspberry Pi for an early prototype for a sensor in a foundry that they are developing. When someone brings up the fact that the ambient temperature might be too high for the electronics to survive, the pace of the meeting can continue high, if instead of it being interrupted by people looking through documentation you can just ask during the meeting:

Triggering Simulations from text

But of course not all of engineering is based on text alone: engineering requires parameters, formulas, physical units and their interactions to derive meaningful designs. But even in this case, assisted engineering capabilities will allow the interactions to be sped up significantly.

Once you combine natural text input with access to data-driven engineering and simulation tools you can on the fly generate accurate answers for hypothetical scenarios. For me, Valispace’s video (from 2018!) visualizes best what such an interaction can look like and why the difference in getting an answer changes how fast your engineering can advance.

Large language models are (with good prompt engineering) already great today at translating human intent into parameterized functions. 

As long as the engineering data is now stored in a collaborative engineering environment with calculation engines, Assisted Engineering can provide accurate answers to loosely formulated questions.

Reviewing

Reviewing the work of other engineers under a specific view-point is an essential every-day task. Many of those aspects rely on knowledge, experience and clear system understanding of the engineer. However there are also many reviewing activities that solely focus on finding duplicates, bad descriptions, unintentional inconsistencies and much more. All of those can be automated to speed up the engineering process significantly and to reduce costly mistakes before even starting the manufacturing process.

For this example I have on purpose introduced a few mistakes into the design description of the Lilium Jet from their website, which an Assisted Engineering system can spot right away:

Similar functionalities will be able to review documentation, specifications, test procedures vs. design, etc. etc.

System Modeling

Designing a product is obviously not only something that happens within the requirements space, but at some point an actual system needs to be described with numbers and models.

Assisted Engineering will be able to extract the structure and initial model-information from text alone.

Being able to convert human language into the initial version of system models removes the entry barriers to model based design that classical approaches have been struggling with.

An Assisted Engineering model that then is specifically trained to suggest relevant formulas in combination with mathematical optimization algorithms is then able to generate meaningful modeling results within a very short time.

Documenting

Documenting a hardware product’s properties and behaviors is an essential part of an engineer’s work, even if not one that engineers usually enjoy a lot. Documentation might be customer-facing (such as user manuals), internal documentation for fellow team members or documents that summarize the state of verification to a certifying body.

In all cases documents have to be carefully reviewed before being released but their preparation and keeping them up-to-date won’t be time consuming anymore at all in an Assisted Engineering environment:

While not being an extraordinary summary, it is a human readable paragraph that contains the key information and has been generated within seconds.

As engineering iterations are becoming shorter and products contain more variants, Assisted Engineering is the best way of allowing engineers to focus on the actual engineering work and be assisted by a never tiring AI.

The risks of using AI (specifically LLMs) for hardware engineering

What LLMs should be used for in Engineering

Hardware engineering is one of life’s areas in which you don’t want companies to make mistakes. You want to be sure that the car is safe, that the bridge has been calculated using the correct forces and that whoever gave the go-ahead for passengers to board a new plane is sure that it will land safely again. 

This makes AI generated content in the engineering process the “Best place to start & worst place to end”. Just like a researcher should not treat a wikipedia article as a reliable source to be cited in her final paper; it would be a waste of time not to read up on new concepts that are new to her on wikipedia first.

Concretely that means that generating initial design-outlines, brainstorming solutions as well as summarizing and expanding existing work all are tasks that the engineer can be perfectly assisted with. The AI does not release the engineer from the responsibility of her work but is another tool under her belt, just like spreadsheets, simulation software or other tools have been in the past. 

How to complement AI with math and engineering structure?

Language is just one piece of the engineering puzzle. Designing products well relies heavily on values being correctly derived and calculated; simulations to use the right in- and output to calculate performance parameters and boundary conditions and requirements to constantly be checked.

These kinds of problems are “math problems” and should be treated like that: AI (at least the kinds available in early 2023) are not built for and not suited for this kind of task. So the most important step to be taken by any Assisted Engineering software is to allow users to move from the “text” to the “numbers” domain and back easily:

Numbers and physical units that appear in AI generated text need to be converted into parameters that can be linked with formulas amongst each other. Optimization should be left to mathematical optimization algorithms that can produce amazing results (both to calculate optimal parameters as well as to generating shapes) without the need for neural nets.

How to ensure that AI assisted products are safe?

The answer to this question turns out to be quite simple. Even today we don’t trust a single engineer to do their work “correctly”. We have built systems of verification and certification to ensure that the final product is safe.

AI can be used to help engineers derive their designs faster. The same way we spot an excel-formula mistake or a documentation copy-paste error we are as likely as to find problems that have been introduced by the AI.

Making sure that our calculations and numbers are derived from well understood formulas and mathematical algorithms over black-box neural networks allows us to make use of both the creative power of a smart digital engineering colleague who assists us in building better products.

Conclusions

In Assisted Engineering we should treat AI Assistants as our favorite colleagues in the room: They have a lot of creativity and help us derive design solutions faster. They do not get tired of performing tedious and repetitive tasks for us. They have a great memory and their insights bring design tasks forward 10x faster than if they were not in the room. Since they joined our company we have been able to iterate faster; make less mistakes and come up with better and cheaper designs, leading us to faster time to market.

But just like with any other engineer, we would not trust them and get rid of our quality control or certification process; because they might very well make mistakes as well.

Personal plug: 

Given my enthusiasm for how the future of work will be Assisted, it won’t come to surprise that at Valispace we just released Assisted Engineering functionalities that catapult hardware developers into this exciting near-term future today.

All above queries have been generated with ChatGPT. All images have been generated with Midjourney.

Abonnieren jetzt

While Bing might currently be mostly interested in the “assisted search” use case, I doubt that this will be the most society and job-transforming use case of ChatGPT in the near future.

What is ChatGPT (and a LLM in general) good at?

In principle, Large Language Models (LLMs) such as ChatGPT use a half-finished text and always predict the next word which has a very high probability of fitting in (great technical write up here). This means that by definition LLMs are great at coming up with creative solutions to how the text should continue.

What this means in essence is that while creative writing is a natural process for this type of AI aspects such as logic, math, truthfulness or politeness don’t come naturally and need to be added as functionalities.

Since ChatGPT is trained on existing data and uses its existing corpus of learning together with some inserted randomness, its results will likely always be fitting roughly into the mainstream idea of what a mediocre text would look like.

What are daily applications for creative writing?

When trying to predict how in the near future the world will likely be disrupted, it makes sense to focus on applications of what the AI is great at already and to assume that this is where the biggest immediate potential lies. 

So where do we perform “creative writing” in our day to day activities? 

Some cases are very obvious, such as school essays, amateur poetry, wedding speech writing etc. Others are less obvious, such as writing technical product specifications, finding inconsistencies in legal documents, composing music and many more. Let’s dive into some of them and see how the near future is going  to quickly be transformed in these areas.

What does Assisted Everything imply?

School and Homework → towards Assisted Teaching and Assisted Learning

Given that is an obligatory task and kids are smart and tech savvy, it is no wonder that ChatGPT technology is used immediately to finish homework assignments in no time. While educators' first impulse was to ban students from using ChatGPT and later some ideas emerged of assignments that cannot be faked, smart teachers and education experts have quickly found that  embracing it instead is the right strategy. Teaching students how to use this new technology is essential, since it is here to stay and probably a better idea to teach kids how to use it responsibly than to pretend it does not exist. So teachers are encouraging students to use ChatGPT, but to highlight how they used it in their work.

But the biggest potential for disruption does not lie in how students will game thesystem; it lies in Assisted Teaching, in which teachers will have an assistant at their side helps them prepare better lessons, personalize learning content, make content more contemporary and much more. 

• Imagine being able to update your high school seniors chemistry class on polymers to include content, references and jokes from the TV show Breaking Bad within minutes.

• Imagine that when you are being asked “why do we need to learn about vectors in math class?”, you are able to mass-customize exercises with examples from all your students favorite hobbies.

• Imagine being able to teach the differences of Old English dialects using Elon’s tweets instead of 7th century texts.

While inspiration for concrete suggestions for teachers will spread fast, students will also be able to learn more about any topic through Assisted Learning themselves:

• Imagine being able to take paragraphs from complex subject matter books and get a simpler explanation in plain language within seconds when you are struggling with a new concept.

• Imagine generating for yourself a geography quiz by just writing 5 lines of text instead of going endlessly through flashcards

• And if you don't trust the AI to output correct content, teach your students critical thinking by using Assisted Teaching to write history essays and let your students find mistakes in it.

Law → Towards Assisted Litigation

Instead of arguing whether AI is suited to handle legal cases by itself or letting it pass bar exams, the more interesting use case of the current technology is Assisted Litigation. While there are definitively many hard things that lawyers have to do, a lot of work includes summarizing big batches of documents, finding inconsistencies, formulating extensive memos, etc. These all are tasks that require a fair amount of creativity, but allowing a lawyer to edit a “quite good” draft is going to make lawyers far more efficient. So maybe the first use case is actually rather going to be the Assisted Paralegal who relieves the young associates from some of their all-nighter work.

Software Development → Towards Assisted Coding

Software development is probably the most advanced area when it comes to implementation of the Assisted Everything principles:

While sitting together with another human developer to look at the same code and work together (“Pair Programming”) is known for increasing code quality and job satisfaction, it is not widely spread, since it increases the amount of person hours that go into a project.

This is probably the key behind the success of the Assisted Coding tool “Github Co-Pilot” that only one month after its launch already was used by 400.000 programmers. In Assisted Coding, the AI suggests code that fits the context that you are currently writing. It is able to execute tedious repetitive tasks, but also to creatively complete code in languages that the user might not even be very familiar with and propose automatic tests for the existing code.

According Githubs own research, developers that use their Assistant are 55% faster in writing code and feel more productive and fulfilled with their jobs.

Creating Art → Towards Assisted Writing, Composing and Design

2022 already saw a major shift in the visual art world. With the revolution of Midjourney and similar tools first two things happened:

1. A major debate regarding copyright, ownership, artistic merit and other topics emerged, as users started using existing artists’ names in their prompts and styles were copied

2. Artists who used these Assisted Painting services regularly actually became very good at combining and remixing, together with the AI and manage to create amazing and unique outputs.

With ChatGPT, a similar shift is now happening in the world of written art. The sheer amount of new services popping up to assist users in writing anything from children's books to condolence cards,  from slide presentations to love letters, from witty flirt lines to sermons, is only starting to scratch the surface of what is possible. 

Where it gets even more interesting though, is use-cases in which the creative inputs can be text, but the outputs are of another medium: 

• Drama and Moviescripts are a great example where the written word is later translated into other art forms. Assisted Writing of screenplays today already produces results that make professional writers claim “with a bit of editing I could take that to Netflix: just need to finesse a little bit”. 

• But even in less obviously connected areas language-generating AI can change everything: Assisted Composing allows audi artists to create music from text.

• Even complex visual art forms such as apps or websites might soon be easier to design with an Assisted Designer than by yourself, even if you are a professional. 

Hardware Development → Towards Assisted Engineering

Hardware development at first seems like a very math and physics intense discipline that does not immediately come to mind when thinking about LLMs in AI. However, actually a whole lot of steps along the product development lifecycle are text based: Everything from writing specifications and product requirements, to analyzing and reviewing datasheets, preparing test procedures and technical documentation is text-based and can therefore be improved through Assisted Engineering. Since this is a topic that is very close to my heart, I’ll dedicate a few posts to this topic, starting with a general overview of why products that are developed using Assisted Engineering will counterintuitively be safer than todays’ products; even if current LLMs are terrible at math.

Similar to the Assisted writing case, the first applications of LLMs in engineering modify text alone (such as improving requirements, preparing progress reports, summarizing technical datasheets, etc.), but there is even more potential to be unveiled when text describes logical models of our hardware and by altering text we are able to alter the design itself, which I will explore in future posts which you won’t miss if you subscribe to this substack.

Leading Teams → Towards Assisted Management

While it is unclear whether being assisted in writing performance reviews or reviewing  applications in Assisted Hiring will make the processes better or meaningless, only time will tell. 

However it is very obvious that summarizing calls and drafting minutes of meeting, reviewing KPIs for consistency and clarity,  brainstorming pros and cons for decisions, answering questions of newly onboarded staff and similar tasks will speed up daily tasks in management immediately and probably be hard to work without within a short period of time.

Conclusions

Taken AI generated content as-is with today’s technologies are seldom extremely creative but consistently provide above mediocre results if prompted the right way. 

So while the outputs by themselves won’t change the world (yet?), the interaction with these mildly creative texts have the potential to speed up human work extremely. They are a great starting point for any process that relies on a loop of rapid creative creation and continuous improvement: 

Why should teachers, lawyers, programmers, writers or engineers sit down in front of a blank sheet of paper, if they can work with an assistant who provides them with drafts, improves it according to their wishes and takes over the tedious parts of it?

Ultimately, efficiency in any white collar job depends on the increased rate of iterations. And if we are able to reduce the time of iterations with ourselves (coming up with the first draft of anything) and with others (relying on the Assistants functionalities of reviewing and proposing), we all get to create more in less time.

There is a literal race amongst new startups to provide solutions and tools in every imaginable vertical is what makes me certain that within the coming 2 years most white collar jobs will be assisted.

Personal plug: 

Given my enthusiasm for how the future of work will be Assisted, it won’t come to surprise that at Valispace we just released Assisted Engineering functionalities that catapult hardware developers into this exciting near-term future today.

All shown queries have been generated with ChatGPT. All images have been generated with Midjourney.

Subscribe now