“A person is smart. People are dumb.”, Tommy Lee Jones says in Men in Black. We’ve all experienced this. Your personal views about an issue might be nuanced. The message you liked about the issue, or the picture that got a bunch of hearts on Instagram is probably not as nuanced as your personal views.

It’s strange there’s not a name for this phenomenon beyond wisdom of the crowd. The act of having your intentions misrepresented or smushed into something bite sized for consumption by other people is something we face frequently. It’s the reason we post sappy LinkedIn messages, engage in boring conversations with people we’ve just met, and refuse to return bad vegetables to Trader Joe’s. We operate under the rules the crowd sets for us even if those rules are not something that align to our personal beliefs. This is of course different than law breaking, which we don’t do because there is a communal ethical standard or in the case of parking tickets, the inconvenience of fees or your car being towed. There is really only one disadvantage to not following the group: that you can possibly get excommunicated from the group.

You might alternatively find that the group agrees that a rule should be broken. Groups by their design are not rational. They attach themselves to charismatic narratives. The reason we post sappy LinkedIn posts is that we know those get the most engagement. The reason we engage in boring conversations is because it is polite. The reason we don’t return vegetables is because we picked them out. But turn each of these narratives over, and they clearly have holes in them. No one really enjoys the sappy LinkedIn posts. No one really enjoys the conversation. And those vegetables go in the bin and you lose your money. In each of these cases a self-defeating narrative has been created.

The wisdom of crowds is to lean towards charismatic narratives. This is not logical and not entirely representative of the members of that group. Groups are malleable from the inside if given more charismatic narratives. Often times, these risk excommunication but the upside is worthwhile.

I haven’t written in a while so this might be a bit choppy. If you liked it, let me know by liking or subscribing.

Subscribe now

Then it quietly crept back up. It recently crossed $100,000 and no one batted an eye. Because bitcoin, similar to AI, and similar to other phenomenons before it, went from outside interest to mainstream technology. Companies like Meta are paying millions for AI talent. The reasoning is simple: the road ahead is long, and there’s still a lot of upside. Also like crypto there is a lot of noise in the signal. This means sticking to fundamentals is important. It’s no surprise that about half of the engineers that Meta poached worked at OpenAI.

There are a lot of unsolved problems. Most enterprises are just using AI as a chatbot solution. Most patient intake is still done manually. Most meetings now have AI notetaking, but not AI summarization. There are still a lot of areas where AI can be leveraged to reduce friction.

We also do not have the right skillsets yet. Many organizations are downsizing other parts of their business to focus on AI-skilled talent. This misses the opportunity to leverage SME knowledge and upskill workers on the potential of AI. There are a number of functions that, augmented with AI, can work much more efficiently.

Models are also getting larger. While this is great for more complex tasks, focused and specialized tasks that need narrower scope are not well supported by these models. This means that to support the full range of tasks and subtasks, we will need smaller and faster models to not just execute but also to guardrail these executors.

Fortunately, there is also the appetite to get these things done. While the current AI wave started as a consumer product with ChatGPT, it is now largely being driven by enterprises. Enterprises stand to benefit greatly from the improved efficiencies from AI, and therefore are spending a lot of money on it. This means that AI is now being heavily driven by these interests and therefore we should see a number of powerful innovations coming to continue to fascilitate growth in AI.

Subscribe now

AI has a human problem. I don’t mean in the “I, Robot” sense, I mean in that humans mostly don’t know what to do with it. People will see that others have built chat bots and immediately go to do the same. This feels like a solution in search of a problem.

For some organizations there is no value add in owning the infrastructure. They end up overrunning on cost with a mess of infrastructure. Adding to this is the lack of foresight in how AI helps the problem they are solving. You should always, always, always start with the problem. What is it that is challenging for the user? Does AI solve it, or create a new problem?

If you’ve mapped the problem and AI provides some value add, first look to see if it has synergies with your organization. Are you handling lots of data anyway? Great, looks like a strong fit. Are you a business that doesn’t, but likes the opex savings? Plenty of companies will manage the AI in the equation for you, so you can focus on delivering your core competency.

Here, you need to make sure you are tracking what you are doing with AI. Without good visibility, there are going to be too many elements that can impact success. You should be able to see what inputs and outputs are at every step, and isolate if something breaks to where it broke. Ideally you have some metrics to do this that align with success at each step. Automated alerting would be great too.

You built an AI app, with visibility, that solves a real problem. Great! Now you need to make sure that it continues to perform as expected. Models will deprecate. Identify a path to validate new models that doesn’t impact your core solution, or that can easily be reverted. Remember those metrics from before? They should still be running in production. If they can’t, or it becomes too expensive, try another solution or metric.

And that’s it. If you liked this post, feel free to subscribe for more.

Subscribe now

I picked up a Switch 2 today. It is cool. I am not a big fan of Mario Kart so I just ended up playing my old Switch 1 games on it. The Switch 2 gives those old games some quality of life improvements. It doesn’t feel like I’m describing the best selling console in history. Why is the Switch 2 selling so well, and why was it initially panned?

When the Switch 2 was first announced, forums ignited with angry gamers. They were outraged. Nintendo had priced their console at $450, and games at $80. This was a 50% leap from the Switch 1, which launched at $300 in 2017, and a jump of 33% on games. In the same time, inflation has been about 31%. This did not bother the gamers though, as they demanded that Nintendo apologize for rising wage and development costs.

This explains the pricing of the hardware and software, but not the velocity of the systems flying off the shelves. The answer lies in a few factors.

First, Nintendo have proven that they are the IP champions. The Mario movie is the 17th highest grossing film of all time. Pokémon Scarlet and Violet sold over 10 million copies in their first 3 days, as did Legend of Zelda: Tears of the Kingdom. Compelling, differentiated experiences that are tied to Nintendo IP drive consumption of these platforms.

Second, the Switch 1 is a beloved console. It is by some estimates the second best selling console in history at 152 million units. Nintendo didn’t need to significantly change up their strategy given that it was already proven successful. Instead, they could rest on what made the Switch 1 successful and provide those small quality of life improvements. Additionally, and as is tradition for Nintendo, they have allowed some backwards compatibility extending the blossoming Switch 2 video game library with the mature Switch 1 library. Nintendo has also learned from failures such as the Wii U and forgone any naming subtlety by sticking a 2 at the end.

Finally, Nintendo has sent a clear signal in taking their time. Given there were 8 years between releases of Switch 1 and 2, Nintendo has made the statement that this isn’t fast tech. Many consumers, myself included, are overwhelmed with new iPhone versions each year, new LLMs each month, and new TikToks every day. Nintendo has shown their willingness to maintain their hardware and software with continuous updates and support. This de-risks the purchase for many consumers, who know if they buy a Switch 2 today it won’t be obsolete or unsupported in a year. This means the consumer has more patience and can build that in to their decision to own a console.

And that’s exactly what I’ve done.

Prompt engineering came out of this. It existed before but ironically being able to supplement your prompt meant it was a lot more important what your prompt was. Zero shot means you give no examples for the type of outcome you are looking for. Few shot means you give a few examples. By giving these examples we helped structure the RAG. This also meant we got even better answers.

Because LLMs responds best to well-structured, context-rich prompts, we continued to iterate on this premise. What if the LLM could go and fetch the right context at any given point? And thus tools were introduced. If the context retrieval requires taking action, we allow a tool to abstract that action.

Enter 2025. An LLM that can call a tool is called an agent. Agents help us by solving the last mile problem. The last mile problem in delivery is how do you get something from a distribution center to the front door of the purchaser. The last mile in LLMs is how do you take information returned by the LLM to solve a problem the user is facing. The last mile is typically what a person would do, eg. click through some buttons based on the information from the LLM. These agents can do that for me.

MCP came out of this. Model context protocol is a system that lets agents interact with tools in a more standard way. Traditional software is designed for human inputs. Humans input things slowly and expensively (cost of loading a design versus loading a schema) compared to agents. By letting agents have a standardized way to interact with tools and pull context, agents can do things for us faster.

Agents are able to use this standardized approach, or similar, to execute things on our behalf so we are managing outcomes, not processes. This frees us up to do things like worry about whether we are getting the right outcomes or be more strategic with our goals for these agents. Where aptitude for a human in a task can take several years, through reinforcement learning and virtualization, these agents can approach the same aptitude in days or hours.

This makes sense. LLMs are trained on the equivalent of 281,250 books for a 30B token model. On most standardized testing these LLMs are able to outperform top benchmarks. It’s no surprise that this influences the direction they will head in 2026.

So far, we have primarily focused on text. Context and tools in text. We are sort of aware that LLMs and agents can interact with other modalities, but it is under leveraged. Most production use cases are text based, both on input and on output. But the facts support additional modalities. Most premiere models support images and audio and video inputs. The reason these use cases have not been adopted is because they are more difficult to monitor and guardrail. In other words, the risks currently outweigh the benefits. Next year the scales tip towards benefits. The premiere models have been testing and updating their multimodal capabilities all year, and as we have previously pointed out, a year is a long time for an LLM. So it seems realistic that next year we will see image, audio, and video generation that is production ready where we can also guardrail and monitor effectively.

At Meta, she was responsible for building out their advertising business. Instacart chose her to be their CEO in 2021. Notably at Instacart, Fidji was responsible for acquisitions of Caper AI, Rosie, Eversight, and Wynshop. These acquisitions enabled growth in key areas for Instacart and provided additional advertising channels. This makes her a perfect choice for, as Altman says,

This rolls nicely into OpenAI’s purchase of Windsurf. Windsurf is an AI coding tool. Others, such as Cursor, have tracked at a reported $100 Million in new ARR in the last 5 months. Growth in the LLM wrapper space, applications which wrap LLM’s to optimize existing workflows, is rocketing. These applications are also expanding the TAM, or total addressable market, of AI.

OpenAI itself does not have a growth problem. Per many reports, it is still growing at a staggering rate. OpenAI does however have a problem capturing some of the value it is creating. LLM wrappers are an excellent way for OpenAI, and presumably its competitors, to go after higher margin opportunities. Additionally these businesses are already reliant on foundational companies such as OpenAI to wrap their applications around. OpenAI can simultaneously lower COGS, LLMs in this case because they have lower prices than what they offer publicly, while increasing margins on the service of those LLMs. Once they have these additional channels, they can also focus on advertising and upselling other OpenAI services and applications in these LLM wrappers.

Fidji has exactly the right experience to be successful here. Naturally her experience in acquisitions and advertising is foundational. Beyond that, she also led Instacart through an IPO. OpenAI is getting a skilled executive who has experience with the roadmap they have ahead.

11x, Cody, and kay are all examples of companies that straddle the line between tech and teammate. 11x breaks out the two hireable personas into Alice and Julian. Cody and kay are the names respectively of those two agents. Across those three platforms, there are similarities in approach.

First, you’ll notice that the names given to these agents are more personable. This makes it more natural to hire and interact with these agents. Humans have a natural tendency to be skeptical of paradigm shifts in technology, so this allows for a more familiar way to utilize these teammates. Second, these agents have a focus on training. Across all three web pages the terms train and learn are prevalent. Instead of more technical terms like fine-tune, these agents are “onboarded”. This re-frames the problem as a more familiar one where resources are used on enablement instead of implementation. Finally, these new employees are integrated as always on team members. They are not seen as compute but rather as hard working, 24/7 co-workers. With interfaces such as Slack or email these agents become indistinguishable from other employees.

If we treat these agents as employees, we also want to interview them appropriately. Just because generative AI ROI is estimated at $3.7 dollars per each $1 spent, does not mean that will be the realized amount. The new framing of hiring these agents doesn’t change the fact that onboarding them will be the same rigorous process as any other software in an organization. There is even a higher chance it will be more rigorous as these agents require broad access and may send requests to a third party that contain highly sensitive information. Before continuing, ensure that the agent is the right way to solve the problem you are facing.

Assuming it is, the first thing you should do is a background check and get references from other teammates of these agents. These references should include not just the value of these agents, but also focus on how the user is interfacing with it. This should include what they found helpful when they first started working with the AI agent.

Identify weaknesses as well. Find the edge cases or limitations where the agent doesn’t do as well. Either confirm you can solve for those issues without the agent, or make a request for a feature. Think about what happens if the agent is unavailable or goes down.

In the same way you hire for the future aptitude you may need, the agent will need to evolve as the problem space does. Ask where they see themselves in the next few years. Confirm they can handle all of the shifting context and responsibilities given the role they have to play.

Assuming you do your due diligence, the agent can play an integral role in your organization. Ensure that they get the support they need when they need it. The problems they solve for are efficiency focused, and therefore without this support, they become bloat. Efficiency gains are also better when they are shared. Make sure the agent is easy for many users to make part of their existing workflows. All that is left to say is good luck and may this new agent be a successful addition to the team.

Subscribe now

Thanks to Kwabena Ohemeng for review.

Generative AI developers rely on transformer models, the standard for LLMs and the T in GPT. Leading models are trained using Reinforcement Learning from Human Feedback or RLHF, where humans rank outputs so the model learns to predict and optimize for responses people prefer. Innovations such as Deepseek’s synthetic-data review have slashed costs without sacrificing performance, letting models improve through continual experimentation. Through this experimentation, these models get better aligned to goals we set for them.

Currently we are finding efficiencies in the transformer model approach. This is a period of what Thomas Kuhn would call “normal science”. We accept the reality of the transformer approach and are trying to optimize it. This process is incremental, and with each jump we are seeing some improvement over the previous model in exams such as the uniform bar exam (sometimes not). Risk and compliance are still top issues for adoption, but adoption is growing as improvements de-risk these models.

Think back to the Attention Is All You Need paper that introduced transformers. This was a paradigm shift. Transformers didn’t just improve on the previous method, they provided a new approach that made generative AI cheaper to use and more effective. When introducing this kind of paradigm, you also unlock opportunity that wasn’t previously available such as non-deterministic chat bots.

Introducing paradigms is a radical act. You are in effect saying the world works differently than how we all think it does.

Hinton mentions that humans aren’t rational, they are resonant. Paradigm shifts need to resonate with a lot of people to be effective and adopt the paradigm. Depending on the amplification and application, ideas can resonate quickly or slowly. Some paradigm shifts also need to find the right audience to be effective.

Fortunately, by reading this blog you are likely the right audience for this idea. While the new paradigm isn’t here yet, it is starting to take shape. You can already use generative AI to take back your time for a number of different use cases. As an example I start coding projects faster, generate basic patterns, and debug my code with the help of AI agents. Delegating these actions to an AI agent helps me focus on the more valuable parts of coding. This has made projects which would otherwise take me weeks take me days to complete.

Models getting better is normal science. The bar I would expect from a paradigm shift is much higher. It needs to transform the reality of the problem altogether. The way I foresee AI doing this is by anticipating my intent. If my vision is correct, the next paradigm shift will come from not just how we interact with the model but how it interacts with us. These models should be able to predict what prompt we may want to serve them and proactively resolve the request. Think of this as a recommendation and enablement engine for action. Minority Report but with better guardrails and more boring application.

Thus we return to our agency. We shape the world around us by interacting with it. With a AI agent, we can interact more rapidly and therefore learn more rapidly. Daily friction removal will help us be more effective with our time. Eventually, we will be able to predict where friction may lie and proactively remove it, taking prompting out of the equation for humans. The question for after that reality is whether the next paradigm will be found or whether it will find us.

Subscribe now

Thanks to Pratik Bhavsar and Kwabena Ohemeng for review and feedback.

“Who would I be if I weren’t seeking anything?” - Haruki Murakami

It is easy to want something. There is no effort or cost required. You can want to exercise, to check out a restaurant, or world peace. You can have as many wants as you can imagine. Without urgency or means, wants do not get addressed. So they stay exactly that, just things you want.

If you need something, it has a cost. Other things are put on hold for needs. Needs are known even if you don’t necessarily know how to resolve them. No one advertises needs. Someone might make you want a specific thing that solves a need, but the need is innate to the problem.

It can be challenging to know the difference between a want and a need, but typically a want is dressed up as an abstraction of a need. They claim to deliver the same thing that you need, but when you look closely they are only mimicking it. Social media likes? That’s a want. Real friendship and companionship? That’s a need. Likewise, you shouldn’t de-prioritize needs for wants. There is a stigma attached to seeking help if you are a man, such that you may not want to look weak. But as you can tell, that’s a want. If you need help, you have to do what is right for you.

The problem is that we are a society of wants. It is beneficial to want. If you do not want anything, you are like a monk and of no use to anyone but yourself. But we are, for the most part, not monks. We do not exist in isolation, but amongst a sea of others. Those others can also want the same things. Wants help drive us to goals that we couldn’t achieve individually, like going to space.

Needs are essential for an individual and wants are essential to community. A strong need is something that impacts you, and a strong want is something that impacts your community. Both wants and needs serve a purpose. Your perspective on wants should shift to questioning how they fit in communally, and your perspective on needs should be to prioritize them for yourself.

Some people are tremendously talented. They don’t just capture one discipline, they capture multiple. Kubrick is another good example, having made critically aclaimed movies stretching from space to horror to comedy. Marie Curie has two nobel prizes, one in physics and one in chemistry.

The connecting thread is that their talent was sculpted by adversity. Certain difficulty is unavoidable, such as geopolitical or circumstancial issue, and this is by no means advocating for seeking out adversity. Instead, if you are working on something, and you face adversity, and that things comes out the other end, it is a better thing for it. Competition makes you a better competitor.

That means systems should be self enforcing. Being in a market makes you better for that market. This is why time is a proxy for experience. Just by existing as an artist or a scientist, you enforce that you are that thing. It follows that you also face more adversity the longer you are that thing. This feels wrong though. At a certain point, people accept you are that thing and the adversity drops off.

Once you reach the plateau you can do one of two things. You either challenge yourself by iterating on your own ideas, or accept that you have reached your level. If you continue to iterate you will find adversity even in correct ideas. This is because you have a perspective that few others (except those with the same time in role) will understand fully. Eventually they will catch on that your idea is right and it gets cannonized. They will act like it was obvious in hindsight and reject the notion that they ever doubted you or the idea.

And yet talented people live for that. That tiny spike immortalizes them forever. They spend some portion of their lives getting antagonized to eventually have an impact that is so outsized it makes it worth it.

The first conclusion we take from this is that experience plus drive is ultimately the formula for finding competence, not just experience. If you have not pushed yourself in the last several years you are no longer in the same competition as people who have. The second thing is that challenging ideas is good. Where possible, you should challenge ideas. This doesn’t have to come across as crass. You can challenge ideas by first acknowledging what you like about them. If you are working with someone who is driven they will appreciate the challenge. This leads into the final point. You should want your ideas to be challenged. Seek out adversity on your ideas wherever possible. If you are not open to it then you will be doing yourself a disservice.