90% : Strong Conviction

Personal notes for where we're headed over the next 10 years. While the future is never written in stone, I'm 90% sure of these outcomes. Past a decade, my confidence diminishes significantly.

LLMs will produce most of the functional code that's written in industry. The verification that this code does what users expect / intend is going to become critical. To create good software, it's not enough just to have a specification. You need:

• Contextual knowledge of the business. The problems, the objectives, the design constraints (financial, engineering, design, etc).
• The personalities at play. What do they expect it to do? How will they personally benefit from the development?
• Knowledge of the existing solutions; where data is located, interfaces to existing vendors, what toggles can be modified today?
• How will users use it? Will they follow the happy path? Does it need to be robust to adversarial inputs from outside?

In complexity theory, the definitions of P/NP distinguish problems that are solvable in polynomial time (P) versus problems that are verifiable in polynomial time (NP) but might be more difficult to solve in the first place. This relationship extends to how engineers write code today: the majority of time is spent writing functional logic instead of writing the tests¹. LLMs will flip this relationship. Engineers are going to spend most of their time defining the business objectives and expected handling of workflows (happy paths, edge cases, expected IO relationships) instead of writing the functional code.

Authenticity

Consumers are going to be awash in generated content. This will serve the purpose of private marketing (classic advertising or astroturfing campaigns) and more malicious state-actor attempts to sway public opinion. As a result, people are going to place a much higher emphasis on trust capital online. The value of real, provable, and authentic relationships are going to increase.

• Some trust might return to established organizations (the media, government, etc) but it just as easily might disseminate to community leaders. These individual leaders take different forms. Some might be known in-person though their local town, others may simply have a cult following of their blog, youtube shows, podcasts, etc.
• There may be a healthy medium to the size of these followings. Too large and they risk getting tarnished with organizational distrust. Too small, it's difficult to moderate and prove their legitimacy.

Technology can be part of the solution here but only a part. All digital signatures can be forged. Photos purporting to show some international event will have to be verified by people who are there, or journalists from abroad. For us to then trust what they say, we need some notion of identity online.

Governments are really in the best place to ensure this authenticity online, but there's little public appetite for them to do so. The big mainstream platforms are in the second best place - but they themselves show little appetite for doing so. It's likely that developing nations will figure this out before developed ones².

Market adoption

Most companies have become disillusioned with adopting new software. Managers have overseen one too many digital transformation projects that come in years late, millions of dollars over budget, and leaves the company in effectively the same state as it was before. Literature researches this as the productivity paradox, with some sources affirming and others challenging the conclusion³. Personally, I think this is more of a feeling than a quantifiable judgement.

Despite this trend, AI's biggest benefits can be accrued to companies and personas outside of Silicon Valley. Companies may not have seen much opportunity costs to conservatism in the past 10+ years because software hasn't changed that dramatically. But they will in the next 10. In periods of technical acceleration, the cost of losing out on technical gains becomes that much more pronounced. The opportunity here is in opex triage under the curve of ∫ workplace efficiency. That itself is going to be most highly impacted by technical acceleration.

SaaS software was mostly a focus on:

• Business model change (flat fee + support to monthly recurring revenue)
• Consumer-design brought to the enterprise
• The cloud and continuous integration allowing startups to launch a half-baked offering and quickly improve it

AI SaaS will be governed by:

• Quantifiable OPEX reduction in business costs
• Sense of opportunity cost for falling behind
• Business model change (utilization of models, or outcome driven revenue sharing)

AI has managed to captivate the imagination of people across backgrounds, because AGI feels so close at hand. Whether it is or not matters less than the worry of falling behind. This coupled with early case studies on business outcomes will likely buck historical hesitation to cross the chasm⁴.

Solved problems in historic ML research

Speech to text, text to speech, named entity recognition, and summarization will be fully solved problems. They already are in large part today. But we can assume that these primitives will be as stable as working with deterministic APIs will be today.

• When you increase the input space of software / businesses to anything that's text based, you also open yourself up to more unintended consequences. Realistic input fuzzing for these pipelines is going to become important to combat adversarial attacks. A whole new wave of SecOps is going to be focused on these attack vectors; especially once the first big hacks become public.

50-90% : Medium Conviction

Cheap capital will not return

My basic argument: US debt load increasing will eventually need to print more dollars to cover payments to creditors. This results in more US denominated spending where other countries will spend dollars outside of debt instruments. They'll use returns for construction or their own economic stimulus projects. The Fed will have to keep interest rates high enough to keep larger inflation under control.

Some challenges to this:

The US benefits from issuing debt in its own currency. This position affords it greater leeway in managing debt without resorting to the printing press, especially if debt growth is matched by GDP growth.

• I assume the US Dollar will remain the world's default reserve currency well into the next 10 years, but just because it has a preferable debt agreement does not free it from servicing its obligations. ⁵

The assumption that creditors will spend US dollars, thus potentially fueling inflation, does not account for the possibility that they might instead reinvest in more US assets, including debt securities. This reinvestment can actually fund the deficit without necessarily increasing the money supply or spending in the US economy.

• This is a fair challenge. I imagine this will come down to the specific state of each economy. Current indications in 2023 looking at international markets that are facing debt squeezes don't paint a particularly rosy picture for re-investment.

Labor movements

Growing worry among some jobs that they might be replaced or already are. I've heard this specifically from designers, copywriters, and executive coaches. In some cases the results are quantifiably lower quality but they're still "good enough" to get the job done. This might encourage a resurrection of unions and labor movements that push for limits on AI in industry, or some financial profit sharing for companies where they are affected. See dockworkers during containerized shipping or car manufacturing for similar historical examples⁶.

There might also be a sense of inevitability with these changes, so people are proactively encouraged to learn new skills, go into training the models explicitly, or something else.

Interdisciplinary understanding

A big missing skillset in the market will be looking at business problems and actually determining how they can be refactored into an ML suitable problem. The tighter the IO contract on tasks, the easier we can benchmark and evaluate different models and approaches. We need bounds on performance of these systems before deploying in the wild, so we need to featurize it somehow to a binary outcome.

This knowledge might come:

• Internally, via a new generation of workers that intuitively understands what LLMs/ML are good at and where they fall flat.
• Via new job placement, where companies want in-house expertise to develop these skills.
• External contractors, who define some of these problems and point to internal/external vendors that can help solve them.
• External businesses, who can sell a solution to a problem that they know likely exists.
  • This change-management hurtle to enterprise sales will likely remain similar to today. They'll need to convince companies to modify processes, whereas companies might be motivated to just automate their existing workflows.

Building software remains easier than building things

Our software world is going to look dramatically different than it does today: more personalization, more delegation of tasks. This is going to be true in a world where things still look mostly the same. Buildings are going to be built in the same way as they are today, with similar designs and similar materials. Self driving cars might occupy some of the lanes, but they'll be a minority. AI isn't going to dramatically morph the layout of the physical world on this time horizon.

Compiled languages

Over the last 10 years, we've seen a surge in the industrial adoption of Python and Node. I suspect this is due to their simpler syntax, robust library support, and being pretty quick for POCs⁷. Compiled languages are going to see a resurgence in the next 10. We want to have systems that will throw errors early - so LLMs can quickly get feedback, modify their approach. Even better will be compiled languages that can yield problems at different levels of abstraction; syntax on the line, thematic flows in the large application, etc.

This trend is coupled with decreasing cross-language barriers. All languages are Turing complete; their main difference is in the syntax and in the libraries. When libraries can be automatically rewritten into another language, and the syntax becomes less important as people write less code, the best language will be the one with highest performance and the most robust compiler error handling.

<50% : Low Conviction

Low conviction things here could go either way. I imagine them being trends but I'm not sure which way things will tilt.

A lull in AI adoption

Achieving the last ~20% of ML performance is always the most difficult. And meaningful AI accelerants take time, especially when core research must be scaled up iteratively to training tasks that cost hundreds of millions of dollars. Discussion of AI has reached such a fever pitch - alongside with an overestimation of what it's currently capable of - we could risk seeing a corporate backlash and more caution around adoption.