With every significant event, aim to consider the first, second, and third order consequences of that event.

This is my rough attempt at considering the consequences of LLM tech so I can plan how I spend my time in the coming months.

Disclaimer: Obviously these are guesses and my own personal hot-takes. Take everything with a serious grain of salt. Only time will really tell the consequences of LLM technology.

First-order consequences of LLMs

We have seen these already.

• LLMs enable the quick automation of many basic tasks to a reasonable degree of effectiveness
• LLMs can help you automate tasks on huge amounts of data (at a cost)
• The ease of use of LLMs for basic tasks within their skillset has resulted in fast saturation of simple LLM wrappers for many use cases
• The primary barriers to entry exist around advanced use cases of LLMs. These take more time to reach market saturation.
• LLMs have received enough attention that almost everyone in tech is desperate to use them
• AI code generation is an immediate goal for many in the space
• People will be nervous relying on OpenAI forever
  • Open source models will be actively developed and used
• To find the business uses of LLMs that have not already been solved, you need to explore:
  • Industries underserved by tech in general
  • Limitations/weaknesses of LLMs
  • Areas requiring particularly high investment to build a useful solution
  • Novel ideas that are obscure enough to not yet have been considered

Second order consequences of LLMs

• Current LLMs are not effective on large amounts of data all at once
  • People code around the weaknesses of LLMs
  • It's hard to know which weaknesses will stay and which will evaporate
  • Fine-tuning of models specific to a business use case, where the business has a lot of data, may rise in popularity if it proves effective
• Giving the LLM a specific task is more effective than a general task
  • Curated graphs/chains with specific agents may be an effective solution
  • LLMs may force the user to narrow their prompts and break down the necessary tasks by asking questions back
• People find it hard to write sufficiently specific prompts
  • Frameworks that provide guardrails by curating prompt-flow will help mitigate this
  • Many engineers self-train in this skill
  • AI companies are trying to solve this themselves
• LLM evaluation is hard because it feels subjective
  • Evaluation frameworks will become more and more mature
• Open source models may lag behind flagship models in performance
  • Either: LLM models as a whole will plateau and open source models will catch up
  • OR: AI will hit a singularity
• Ability to "go wide" on massive amounts of data to accomplish new things - provided you can break the task down
• High quality data itself will increase drastically in priority
  • Data and data quality is vital to use LLMs to their best abilities
• Any tool that is "integrate-able" with an agent will be integrated
• The equipment to work with LLMs as a technology will be quickly developed
• Engineers will want to use tech they are familiar with, and will eventually move towards the best quality libraries
  • e.g. langchain/langgraph, while useful, is frequently broken and awkward to use. It will fix itself OR it will be replaced
• Niches that are popular/lucrative and well served by tech will adopt AI quickest
  • Likewise, niches least well served will adopt AI later
• Every procedure needed to hone LLMs for a specific niche will become developed and enhanced over time.
  • The integration of these procedures will also become developed and enhanced over time
  • Using AI to automate these procedures will be explored
  • AI companies will chip away at this market opportunistically
  • Open source options will make an effort to match AI companies in convenience
  • Solutions will emerge first disjointed (point-solutions that make part of it easy), then integrated (e2e solutions)
• Every procedure needed to build LLM apps will become developed over time
• Procedures that will be made easy:
  • Adding context
  • Specialised chains/flows/graphs
  • Data fetching & processing
  • Prompt engineering, clarifying follow-ups, etc
  • Interfacing (APIs, widgets, batch processing frameworks)
  • Evaluation
• The most common AI chains/flows will be abstracted and become their own endpoints (e.g. o1 self-evaluates)
  • This concept will apply to open-source models in some form as well - open source maintainers will replicate the secret sauce

Third order consequences

In General

• LLMs will automate a considerable quantity of things which previously required manual work
  • Service businesses that fail to adopt LLMs while their competitors do will be outcompeted in price
  • Billing models in some industries will need to change (e.g. hourly to flat-fee)
• Many professionals will be made substantially more productive by LLMs as routine tasks are reduced from hours to minutes. Lawyers, software engineers, etc... Any given professional will personally accomplish more tasks than before.
  • Junior roles in various professions may be affected, since they typically add value by taking on the most routine tasks as they learn the business
  • The supply/demand economics of hiring in numerous professions will be affected as existing professionals are able to get more done with the same hours
• People will use LLMs as a crutch; they will lose the skills they delegate to LLMs, or not learn them in the first place
• Effective sales & marketing will become a higher priority to businesses, relative to the services they offer
  • Why? Many services deeply rely on their ability to hire people who can accomplish the service being offered at all. But if their workforce can accomplish more, and the pool of people who can accomplish the same tasks grows, competition will grow as well. It will be vitally important for a business to market their work, as they will face more competition than before.
• Creating trust in your solution above others will become even more of a priority as you face competition

To Software

• AI code generation will be solved. It already has been solved to an extent. It will be solved more and become increasingly cheaper, easier to do, and more flexible.
  • There will be multiple ways to do it and tools to use as numerous people focus-fire on the problem. With a bit of luck, some of them will be entirely open-source.
• Software engineering as a profession will exist in some form for as long as LLMs struggle to manage the essential complexity behind building solutions to problems
  • But the barriers-to-entry may be significantly reduced. Anyone who can clearly articulate how a problem should be solved may be able to solve it with code generation.
  • Common problems with standard solutions will take less time to deliver - LLMs will have seen these more often
  • Big-picture software architecture may become a higher priority
• Software engineering has been increasing its own productivity for decades. All existing trends resulting from increased productivity will be accelerated substantially by LLMs.
• As LLMs make software engineering easier, more features will be built - there will be less delay between idea inception and delivery. The same engineering team will be quicker to deliver.
  • Feedback, analytics, A/B testing, and experimentation will become a significantly higher priority to determine which features add value and which detract
    • Existing analytics solutions will do well (e.g. Mixpanel) as analytics is prioritised more
    • Demand for a good open-source analytics solution will grow
• Understanding the problem to solve and the best way to solve it will become a substantially higher business priority
  • We have already seen this phenomenon as the pool of software engineers has grown and the technologies available have made them more productive. An ugly solution that barely works may have passed historically by having solved the problem at all. In the modern day, many people are able to solve the problem. The most successful business solves the problem exceptionally well.
  • Businesses that solve a problem poorly, but only exist by merit of having solved it first and successfully marketing their solution will be replaced by new businesses that solve the problem exceptionally well.
  • Market research, psychology, and UX design will become even more important
• Competition will increase dramatically for any software business idea.
  • If you have an idea, you probably aren't the only one.
  • Implementation is currently the most substantial blocker to someone with an idea.
  • The barriers to implementation will begin to fall more and more, so more people will be able to implement ideas similar to yours.
• Numerous novel ideas that have not yet been tried will be tried
• Consumers will become even more choosy. Understanding them will be even more vital for a business's success.
• Finding problems that could be solved better will be an even more major problem for entrepreneurs
  • A framework for finding and understanding problems could be useful
• Engineering solutions "that come with everything you need" will dominate
  • Vercel already does this to an extent - but the maturity of the solution (and others) will grow immensely

Other thoughts

• If you make a simple wrapper around an LLM, building a moat will be hard. You need to be in a sufficiently niche and obscure space to do this.
• Be close to the research. This industry is evolving quickly.

Current state

• The software industry is focused on RAG, graphs, tools - code tools that use existing models
  • As it becomes easier to use open source models, adoption will become more common
  • As fine-tuning becomes easier, adoption will be more common if it is proven to be effective
• Code can already be generated for web apps in integrated environments (bolt.new)
  • Fully-fledged e2e code generation and testing environments will be developed for all kinds of code
    • VSCode will be popular for this & with plugins
    • AI that can generate AI tooling for a domain niche will be explored (and in some cases already is)

Other areas of exploration

• Generative Multimedia
• Software startups
  • Every operation needed to build a startup will be made quicker
    • More startups will be created
    • Content will be generated
    • LLM startup ideas will be wildly common
• Software tooling
• Industry specific niches

Listen to this post

Prefer listening to reading? Feel free to listen to an AI generated podcast discussing this post