”We’re moving into the AGI era” announced Greg Brockman on September 6th, “AGI has arrived” pronounced Jensen Huang on September 7th. These both followed the release of the new flagship model from OpenAI, Astra.
The definition to AGI is that the AI matches human level intelligence across cognitive domains. ASI is the version of AI that surpasses human intelligence.
Today it is getting harder to find knowledge work tasks where the publicly available frontier models cannot complete the task if they are given access to the right data and tools (just like a human would need for the same task). We can still argue that there are tasks where the smartest humans are capable of making better decisions with the available information, especially when emotional intelligence is a part of the process. System design is a key part to get the enterprise workflows functioning. However, I would argue that 90% at least can be done as well or better by AI today. A majority of those tasks do not even need the smartest models with the highest reasoning effort which means many are becoming token efficient.
This leads to the price of most knowledge work will start approaching the cost of tokens. Ranking sales prospects, coding a new product feature, answering support tickets, reconciling books, creating a presentation etc., are cheaper and more efficient to do with AI. This doesn’t mean AI will do all knowledge work. It doesn’t always know what to build. Choosing what to do in an infinite universe of options is not AI’s forte. Attention to detail and taste still matter a lot. Human service will continue to provide huge value.
Every company needs to understand what kind of place they have in this post-AGI world. How do they deliver value on top of the commoditized work tokens can provide? Integrations, connectors and data vendor-locks might seem like a defensible position today, but that will not be the status quo for long.
There is a lot of opportunity in the near-term future by just doing the basics well. Understanding system design and processes of specific industries is key. The job of a founder today is to recognise that the capabilities of LLMs have aggressively outpaced the diffusion of AI to enterprises, SMBs and consumers. Every legal services company knows they need an AI solution and have and will be buying one soon. There are many industries where companies will be aggressive buyers of solutions in the next 1-3 years. Founders can identify those markets and provide great solutions with appropriate human service level to sell AI-driven outcomes.
Venture Capital investors need to look typically 5-10 years into the future with the current standard investment vehicle lifetimes. The most resilient markets will be the ones that see aggressive expansion with AI lowering the cost of delivering results. Coding was the first obvious one where we’re now building orders of magnitude more software. The fastest growth companies will be built in the current cycle in markets that grow in a similar fashion. If AI does not accelerate and grow the underlying market, and the task can be done by commodity level intelligence we have come to expect from AI, that market can be a tough space to create a large company in.
An auditor will have more of their work done by AI and the high paying human task's value will a certificate of trust. As the industry builds robust evals and trusted workflows the requirement of human input in the audit process will shrink in the future. Agents will have more access to our systems and can autonomously find the context needed for the audit agent to make its analysis. If we expect the audit process to be completed at a commodity price, the only advantage will be the provided certificate of trust, potential proprietary data within an organisation (like the Big 4 have from all of their previous audits) to post-train better quality AI to do the work with better token efficiency and human service. It is unlikely to see corporations audit their companies more often so it would be logical to expect the market for auditing to shrink. This would be an example of a tough market to enter unless the incumbents struggle to retain their level of trust.
Adversarial markets like cybersecurity, fraud prevention and litigation are clear market opportunities. The demand to protect ourselves scales with the opponent's adversarial capacity and they get access to the same cheap intelligence. Customer demand growth is close to guaranteed, and the bar keeps moving forwards with foundational AI improvements so it never commodities.
Not every token is valued the same. Current frontier models can achieve results with less tokens and are capable of some tasks that the lesser ones cannot complete. In the future I expect many large companies to utilize their internal expertise of the processes and system design with data to create optimized models with reinforcement learning to provide the best price with a high trust level. I suspect this will lead to a barbellification of most industries, where we will see a very small number of large winners and the rest are much smaller.
However, we cannot underestimate the difficulty of simply creating better specialized models. Curating the right data and the training alchemy is a rare skill only a small number of organizations are capable of doing. This skillset is not scaling fast in our workforce. It’s an extremely important position for most companies and very well compensated one. I expect we will see in the next model generations improvements in this field. One of the next generation of frontier models will be able to work on post-training models, guide teams and more expert data companies will help serve these quests. This means access to compute and great data companies will stay extremely relevant in the coming years.
This is the most exciting time in the history of technology and startups. You can just build things and take on harder ideas as your mission is not a crazy decision. We now live in a post-AGI world where every founder needs to understand how will their company stay relevant and navigate these changes in the availability of abundant intelligence at constantly lower costs. Very few incumbents are safe in their position and they know it.