How to tax AI, or rather the beneficiaries of AI, has become a topic of discussion. The assumption being that AI will lead to sweeping job losses, rather than job creation. I’ve compared the scale of this disruption to something like streaming video, at least with the current tech stack, instead of comparing it to a society-altering paradigm shift like the Industrial Revolution. Even if AI does bring on the Fourth Industrial Revolution, there is historical precedent that it would create more jobs than it would destroy. The Industrial Revolution was literal automation, just of the mechanical variety, and it unlocked massive amounts of productivity and growth until the United States decided to outsource this profit center in favor of profit-taking.

Even so, if automation does continue to advance and grow, it’s worth looking at how to offset the societal impacts of AI adoption and disruption, should it lead to permanent job losses.

Why Tax AI?

The reasons to tax AI beneficiaries are many. In the United States, our entire income tax system is based on humans earning and contributing at the state and federal level to provide basic civic services. Depending on your political leanings, you may agree or disagree with this arrangement, whether the government should be doing more or whether the government should be taxing income at all.

At present, taxes fund important services like Social Security, which is under threat after unchecked spending in Washington and a general disinterest in reconciling the budget. Other services include Medicare, Medicaid, the Department of Education, and the list goes on.

The central issue is that AI isn’t taxable for income. It’s not defined as a person with a Social Security number, it’s not necessarily a sentient being, and it’s really defined more as a service right now. How long things will remain that way is unknown. It’s a moral, ethical, and philosophical conundrum that our current political environment is not equipped to resolve, not in the least.

So while AI remains in this gray area, any job losses attributed to it are just gone with no recourse. There’s no ledger keeping track of which jobs were lost to AI and who was affected. There’s no compensation or ramifications for eliminating jobs in the name of AI, which is perhaps why we saw companies act when they had the chance.

AI Washing

2025 was a difficult year for Artificial Intelligence’s reputation, as company after company laid off droves of employees in the name of AI, at least that was what captured headlines. Earlier in 2026, reports were coming out about companies not seeing any discernible lift in productivity from AI. This echoed a 2025 MIT report about 95% of companies not being able to see any gains from AI integration into the workforce. So if AI wasn’t actually productive in 2025, leading into 2026, why was it being blamed for the layoffs at these companies?

Many of the companies that announced layoffs were already in precarious positions as businesses. Most of them were announcing these layoffs after a down quarter or multiple down quarters, so the reprieve of a layoff was likely a welcome one for CEOs looking for a way to cut heads for underperformance while framing it as a forward-thinking strategy.

Some of the companies, once Wall Street darlings, had become seriously embattled.

Amazon

• Over-hired during the pandemic.

• Margins under pressure due to tariffs.

• Expensive AI infrastructure buildout.

• Ongoing corporate restructuring.
  

Meta

• Spent tens of billions on Reality Labs.

• The company’s “Year of Efficiency” predates the AI narrative.

• Multiple restructuring cycles before AI became the explanation.
  

Block/Square

• Prior to announcing an AI restructuring, Block/Square had three consecutive earnings misses.

• Experienced general growth concerns across businesses.

• Investor pressure.

• The company had taken on efforts to improve and expand margins.

• Block had a unique amount of bitcoin exposure during a down cycle, bitcoin had crashed from 94K to 66.8K.
  

Salesforce

• Activist investor pressure.

• Public commitment to improving margins.

• Cost-cutting campaigns already underway.
  

In all of these cases, the companies were already under financial pressure. The pattern was one of contraction, not expansion, so the AI growth narrative isn’t a clean fit. That’s not to say growth and efficiency can’t coincide, but the timing of these announcements aligns more with cost-cutting rather than expansion.

Although Wall Street loves layoffs, they come with baggage for the CEO who presides over them. In most cases, they’re viewed as a sign of failure because they’re often a symptom of organizational failure to execute on some level. However, reorganizing for AI gave underperforming companies the perfect cover to have their cake and eat it too. Struggling companies were able to cut headcount while also looking like savvy business leaders.

Just as these jobs were lost to a black hole of uncertainty, so was the blame as it was thrown into the AI pit, something we’ve now seen the effects of as the vast majority of adults have turned on the technology. All this is to say that there isn’t really any evidence yet that AI has actually replaced any jobs, though its body count is climbing as its reputation plummets.

The Current Proposals

So now that we’ve established why we might need an AI tax and whether we have an acute AI job replacement problem, let’s look at the proposals being bandied about by tech and political leaders. These proposals range from improved measurement and tracking, workforce education initiatives, and universal basic income programs to more aggressive ideas such as transferring ownership stakes in AI companies, imposing wealth taxes, or levying excise taxes on massive data center build outs.

You will notice that conservatives are notably absent from this discussion. That is because, thus far, they have not put forth any major tax proposals specifically targeting AI. In their stead, I will reference British author Matt Ridley, whose book How Innovation Works provides a useful framework for understanding the tension between technological progress and government intervention.

Dario Amodei, CEO of Anthropic

Reading through Amodei’s proposals, they’re not surprisingly, very well reasoned and well thought out. The guy thinks about AI all day and all night. It’s funny how policies are either hyper-reactive, as a result of something that just happened, or as a result of some extrapolation into the distant future based on any number of circumstances. Amodei’s policy proposals are the latter.

• Measuring and Tracking, We touched on the hazard of this earlier. All the jobs that were lost in 2025 are just lost in a wash with any other data; they aren’t attributed to AI in any specific way. Maybe this was intentional, maybe not. Amodei stresses the importance of tracking this data in order to accurately impose any sort of legislation, penalties, or benefits.
  

• Pro-Employment Benefits, He’s more vague about this one. It’s the carrot-versus-the-stick approach to get employers to retain talent and not give in to their instinct to cut costs and fully automate. The forms he suggests are grants and training programs, but this would likely look like tax credits or some other well-known tool to incentivize businesses to retain talent over machines.
  

• Long-Term Economic Support, This is essentially universal basic income, which, depending on your political leanings, should be implemented today, even without the threat of AI, or is a reason to be burned at the stake. He shares some different mechanisms for this, but the reason why this would be introduced would be for a Depression-style disaster caused by AI job loss.
  

Bernie Sanders, U.S. Senator

Senator Sanders, unsurprisingly, is recommending some version of federalizing AI companies via a stock transfer. Once a company reaches $200 million in annual sales, this triggers the tax, which automatically transfers 50% of the company’s stock to a sovereign wealth fund. From there, the fund is managed by a committee appointed by the President. He argues that every American could see an annual $1,000 dividend from the fund, which isn’t nothing, but it’s also not going to make up for mass unemployment.

There are a few issues here:

• Discouraging VC Investment, Startups have long been reliant on VC money to get started. Once this policy goes into effect, investing in an AI startup is somewhat pointless for VCs looking for another home run. If the company will just be taken away by the government at a certain point, or if there’s a hard cap on growth, there is no real point in investing in AI startups anymore.
  

• Creating Market Volatility, There’s the issue of when a company reaches $199,999,999 in annual revenue and the stock just craters as literally everyone divests from it, not wanting to own what will ultimately be a different company with a different board now that the government owns so many shares. There’s also the possibility that the extra shares come from massive dilution.
  

• A Powerful but Unqualified Board, The people managing the fund will likely be sitting on the board of a number of AI companies now. I would like to believe that the President would be able to select qualified people to sit on these boards and make smart decisions about these companies, but Washington’s track record with tech, cybersecurity, and now AI doesn’t inspire confidence.
  

Elizabeth Warren, U.S. Senator

I like Elizabeth Warren. I always think her policies are very thoughtful. I know people give her a hard time for some stuff that happened 40 years ago, but she’s proven to be a thoughtful legislator. Senator Warren’s proposal goes more after the picks and shovels, not the AI companies themselves, which is actually pretty clever.

• The Excise Tax, This is the heart of her proposal, and it really shows an understanding of the issues and what would impact the industry, raise capital, but not discourage innovation. This is really a tax on the massive, and massively unpopular, data center buildout. The beauty of this is that by taxing the data centers, she points out that you’re also taxing the companies investing heavily in AI. She adds that the tax would scale with the size of the data center, which is a nice touch as well.
  

• Taxing Investments in Technology, This is a nuance that has gone unnoticed for a long time, but this loophole got blown open recently, which is why data centers are popping up everywhere. Senator Warren notes, “Right now, companies pay payroll taxes for their workers but get tax breaks for investing in technology,” so the more expensive the buildout, the bigger the tax break.
  

• Wealth Tax, Past versions of this tax from Senator Warren were something like 2% on people making $50,000,000 annually, and it capped at 3% for billionaires.
  

Matt Ridley, Author of How Innovation Works

If you sneered during the last two proposals, you may like what Matt Ridley would say on the issues of taxing companies and the effects it has on innovation. I’ve only read one of his books, How Innovation Works, but from that I can apply some of his notions to this problem. Nuclear power is still a controversial topic even in 2026. In Ridley’s book, he notes how it was slowed and essentially killed off following the disasters at Chernobyl and Three Mile Island.

In How Innovation Works, Ridley argues that bricolage (tinkering and combining existing ideas) is paramount to the success and proliferation of innovation. Once the government begins regulating and interfering, the technology reaches a point of stasis, no longer able to evolve and grow freely as it once did. He uses nuclear power as an example, citing how the technology hadn’t been able to advance for 40 years. It was only when fission began to round the corner that nuclear technology started to seem safe again, but fission is a fundamentally different approach to nuclear power than what was used at Chernobyl or Three Mile Island.

“Because it is a monopoly, government brings inefficiency and stagnation to most things it runs; government agencies pursue the inflation of their budgets rather than the service of their customers; pressure groups form an unholy alliance with agencies to extract more money from taxpayers for their members.”

How Ridley’s notions apply to AI and taxation/regulation, Using these statements as a guide, I would hazard a guess that Mr. Ridley would be against taxing and regulating AI at this stage, just a hunch. His main reasons would be that a catastrophe actually hasn’t happened yet, and even if it had, it’s not worth losing the ability to keep advancing the innovation in favor of whatever benefits the taxation or regulation might bring.

My Humble Proposal

Taxes aren’t just a way to raise money. While they do serve to fund many of the services we have, we have somewhat lost the plot on why we have taxes in the first place. Taxes are intended to encourage or discourage behavior in a society. A prime example is how we’ve essentially legislated away smoking tobacco, only to have younger generations revitalize it via vaping.

I can’t ignore how impactful indoor smoking bans were in reducing smoking, but the knockout punch was the taxes on cigarettes. A pack of cigarettes in my hometown of Minneapolis is $15.00 a pack. Sure, inflation is partly to blame, but these prices have been climbing for years as we try to tax away tobacco use.

I say all this because when we impose taxes on something nascent like AI, we have to be careful that we don’t stifle innovation and that we are encouraging or discouraging the desired behavior from businesses.

Tax the Replacement, Not the Adoption

Much of the angst around AI is about job replacement, yet many of the policies are focused on company ownership, data centers, or the complete destruction of the labor market. These are all treating the symptoms, not the underlying issue we want to address. I’d measure the two things that matter here: headcount and revenue.

For established companies, we would use the 2024 tax numbers to set the baseline for future years. New companies would start fresh, since they don’t have prior records to measure against, and generally speaking, the size of new companies is typically smaller, making the measurement less important. The reason for using 2024 is that it provides a reference point before widespread AI adoption.

The goal is not to stymie AI adoption or penalize a company for being good at AI. If a company can run 100 agents to great effect, good for them. If you want to spend $500,000,000 on tokens, be my guest. The thing we’re watching here is headcount relative to revenue.

Headcount goes up and revenue goes up? Great.

Headcount goes down and revenue goes down? Unfortunate, but acceptable.

Headcount goes down while revenue goes up? Now you’ve got a problem.

Why the Ledge, Not a Smooth Curve?

A gradual curve has no forcing function. There has to be a decision point between adopting more AI and hiring more people. With a tiered system that includes a steep ledge, there are critical points where a company would have to think hard about paying for more compute or bringing on more employees. Bringing on more employees has the added benefit of increasing the threshold for AI usage.

Here’s a rough breakout:

Under 20 Employees

• Exempt.

• We want experimentation, and this is where new ideas and exciting things happen.
  

50+ Employees

• As long as workforce reduction remains under 10%, exempt.

• A 10% year-over-year reduction can reasonably be viewed as normal attrition.

• This could potentially be gamed through repeated annual reductions, though the effect diminishes as the workforce shrinks.
  

10%+ Headcount Reduction from 2024 Baseline or Year-over-Year

• This is where RPE (Revenue Per Employee) comes into play.
  

How RPE Is Factored In

• RPE flat or declining: Exempt
  Explanation: A company is losing money and/or their layoffs are commensurate with their losses. There’s no need to kick them while they’re down. This would also allow for AI adoption to help struggling companies get off the mat.
  

• RPE up modestly (1%–9%): 10% tax
  Explanation: Fairly straightforward. There is some kind of disconnect between the headcount and the revenue; either the company has laid off workers and benefitted from it, or the company is outperforming relative to their headcount.
  

• RPE up significantly (10%–19%): 30% tax
  Explanation: Cranking up the pressure on companies. The RPE is signifying a more dramatic disconnect between revenue and headcount, so we tax accordingly.
  

• RPE up sharply (20%+): 50% tax
  Explanation: The company has completely abandoned their RPE ratio in favor of unmitigated growth without hiring workers in kind. This is the ultimate goal of the policy: to stop this behavior or fully replacing workforces in favor of outsized growth.
  

It’s a Simple Solution to a Complex Problem

I was racking my brain trying to figure out how to track compute, tokens, and whether we would need agent IDs or some other form of employee identification. All of those things could be gamed and would require new tools and legislation. What I like about this solution is that it uses tools already at our disposal to provide answers today.

One additional benefit of this mechanism is that it elegantly encourages the behavior we actually want. If an employer wants to add more AI, there are points where they must choose between AI and a human worker. At the same time, there are benefits to hiring humans because doing so also expands their AI allowance. The RPE ratio also allows policymakers to adjust thresholds over time to fit the needs of the economy and labor market.

Take a company with 1,000 employees in 2024 that later lays off 600 workers. It doesn’t necessarily need to hire back all 600 in the exact same positions, though I would require that new hires be full-time employees, ideally American citizens, although Title VII bars discrimination based on national origin. I say this because the RPE wouldn’t count outsourcing or contractors against the ratio.

Additionally, generally speaking, offshoring jobs has been a very unpopular economic policy. RPE also has a halo effect of reining that in, whether it’s through shell companies, satellite offices, or vendors. I’d leave it to the legislators whether to include H-1B visas in the RPE ratio, though I do want to note that tech companies are outright exploiting H-1Bs, which is working out badly for both the exploited workers and American workers.

Ultimately, the company controls whether its RPE ratio is in balance or not. If the company has innovated a novel manufacturing method that doesn’t require as many factory line workers, resulting in 600 layoffs, it could offset some of that ratio by rehiring 300 employees in sales or other areas of the company.

The intent is not to create more overhead and bureaucracy for AI providers or the companies using AI. Nor is it to discourage AI adoption, especially if it’s helping companies grow. What we want to penalize is one thing: removing human workers and replacing them with AI workers solely to increase profits.

Is it a perfect solution? No. I’m not a legislator, and there is likely an attorney somewhere who has already turned this policy concept into Swiss cheese. What I’m aiming for is a framework that helps us adopt AI without knee-capping ourselves. Looking at Senator Warren’s policies, I could see them working well in concert with something like this.

Nothing is Certain

Not to be dramatic, but Benjamin Franklin’s famous quote, “Nothing is certain except death and taxes,” could prove ironically accurate for the AI sector if the wrong approach is chosen. The longer we wait, the more likely this policy area becomes defined by knee-jerk reactions rather than long-term thinking.

Already, many of the current solutions being proposed seem punitive in nature instead of seeking to avoid the unintended consequences of an AI-driven economy. Part of that is due to the self-inflicted public relations disaster the AI sector has created for itself. The proprietors of frontier models are in full pitch mode as they attempt to justify their valuations, which means projecting far into the future where their models achieve uninhibited success.

Forecasting mass job losses only stimulates a defensive response from the public rather than a willingness to accept and adapt to the technology. My suspicion is that, after the IPOs, some of the hype will die down as people gain a clearer understanding of what is actually behind the frontier models, how these businesses operate, and what can reasonably be projected.

Until then, it bears repeating: nothing is certain except death and taxes.

Yann LeCun made headlines last year when he left Meta, but he didn’t just leave the company; he threw cold water on the entire AI sector on the way out and at a time when excitement and enthusiasm were reaching a fever pitch. He’s back in the news for criticizing Elon Musk and xAI as a “failure,” in part because most of the founding team has left the company. The man is unafraid of controversy, but he saves his most savage barbs for LLMs (Large Language Models), declaring the technology a dead end.

To be specific, when LeCun was criticizing the technology, he was criticizing its ability to achieve AGI (Artificial General Intelligence); think Jarvis from Iron Man. When people talk about “the singularity” and are fearful of AI’s ability to become sentient, they’re generally talking about AGI, even when they don’t mention it specifically.

LeCun left Meta, made this bombastic declaration, and then… nothing; but what if he was right? Yann LeCun has been right about AI before, and he’s generally considered to be one of the “Godfathers of AI,” so when he declares LLMs a “dead end,” you would think people would listen, unless they didn’t like what he was saying.

If LeCun is right, then that means that billions of dollars have been invested in the wrong technology, that is, if you’re trying to create Artificial General Intelligence. That would pose a threat to OpenAI, Anthropic, Google, Meta, and any number of companies who have pushed their chips in on this specific branch of Artificial Intelligence.

It’s true, there’s a lot of money riding on Large Language Models and Gen AI leading to an AI revolution similar to the Industrial Revolution. LeCun’s assertion comes at an inconvenient time for those heavily invested in the technology, but the question remains, is AGI a requirement for success, or can these technologies succeed on their own merits to deliver the first phase of this revolution?

A Whole New World

LeCun’s World Model concept is actually quite brilliant, which probably isn’t surprising. To understand why he feels we need to take this next step into the World Model, let’s first take a look at LLMs and why he feels they’re insufficient to tackle the tasks we’re asking of them, or at least what we expect of AI.

He’s not wrong. LLMs have been derogatorily referred to as autocomplete on steroids, or some variation of that notion. The characterization is apt, and it gets to the point of what LLMs try to do, which is to say, they use machine learning to find themes and return outputs/answers.

The reason why all the foundational LLMs indiscriminately scraped nearly everything from the internet to train on is so when you input “What are some themes in Steve Jobs’ biography?” the LLM first has the terms “biography,” “Steve Jobs,” and any number of other phrases. It can then make its connections with the text it’s sucked in to give some kind of answer, though that answer may be a total hallucination.

The issue with language is that it’s already abstracted from the original thing. It’s the Magritte’s Ceci n’est pas une pipe (“This is not a pipe”) painting with every input. This has a knock-on effect that means the LLM doesn’t actually know anything; the tokens (words and phrases LLMs use to process requests) are abstracted from the things they represent. So Steve Jobs isn’t Steve Jobs to an LLM; it’s just the words “Steve Jobs,” and those words appear frequently together. They also appear in the context of certain phrases and descriptions. If you brought an LLM to life somehow and asked it to point out a picture of Steve Jobs, it wouldn’t be able to do that, not without some kind of external tool, in which case the tool is doing the work and the LLM is just expressing it.

Which brings me to my next point: today we compensate for LLMs’ shortcomings by using scaffolding, a secondary layer which can be comprised of machine-learning functions with specific actions or traditional functions executed in standard programming languages. The end product is something greater than the sum of its parts; it can give a really smooth experience, if a deceptive one.

That’s not good enough for LeCun because, let’s be honest, it’s a hack. What he’s looking for is a single approach that can take inputs and contextualize them without scaffolding, or perhaps the scaffolding is more specific to giving the model context about the world around it.

In LeCun’s world model, the model compensates for these shortcomings. Using my previous example, the model would presumably know that Steve Jobs was the founder of Apple, that he was adopted by an Armenian family, that he founded Next and co-founded Pixar, and that he sadly passed away from pancreatic cancer in 2011. Additionally, the model would presumably know that Tim Cook is the current Apple CEO and that the new CEO John Ternus starts on September 1st, 2026.

This is all stuff LLMs don’t have context for yet, not without tooling and scaffolding, so you can see why LeCun is dubious of them as a viable path to AGI. So while LeCun’s model sounds magnificent and like the correct approach, it also seems like a purist approach to AI versus a pragmatic approach to AI. That distinction is important because LeCun is a scientist and researcher first.

My question is, does that cloud his perspective on LLMs and their practical applications as he pursues the perfect solution for Artificial General Intelligence? More importantly, just how far off is this solution? We have LLMs today; the world model hasn’t even been demoed yet, so while the world model sounds intriguing, there is no “world” model until it can meet the rigors of the real world itself.

Is it AGI or Bust?

LeCun’s whole premise for leaving Meta and taking his position on LLMs is that they aren’t the path to Artificial General Intelligence (AGI), but is that really a requirement for AI to succeed in the short term, or are his goals different from the sector at large?

LeCun isn’t the only one talking about AGI. Sam Altman has used AGI as his North Star for OpenAI and frequently uses it to intimidate the world at large for reasons unknown.

Just a quick recap of Altman’s history discussing AGI:

2023: Altman discusses the possibilities of AGI in a blog.

2024: Altman predicted AGI would do most marketing work by 2029.

2025: Altman said AGI is generally pretty well understood, which is different from existing.

Obviously, this didn’t happen in real terms. We had companies adopt AI on some level and deploy agents via MCPs, but they were far from self-sufficient, singular beings.

Early 2026: Altman shifts the goalposts on what AGI is (10:00), about what the definition of AGI is.

Today: Altman declares that college students will graduate into a world with AGI in existence.

So clearly, these two tech luminaries have different opinions on the timeline of when AGI will descend upon the unsuspecting public and whether it will happen at all with the current tech stack.

We’re still only scratching the surface of agentic AI, and really, agentic AI can more be defined as AI working on behalf of a user, not necessarily in a sentient way. That’s a really important distinction because AI working in a domino effect of functions and LLM processes is very much doable today; sentience, not so much.

So is the former enough, or is the Valley going to be hard-lined about what qualifies as agentic AI? Moreover, if it works, do users care? The bar today is to make AI functions that are easier, faster, and better than the previous generation.

As someone building a product, I can tell you that the bar is nosebleed high. The Overton window of what’s acceptable has shifted from rudimentary prototypes to a refined, high-functioning product. There is no more passing along crudely taped-together MVPs to users and clients to get feedback. These products are competing in the real world; you’re not trying to beat the idea of something or something offline that is slow and clunky, you’re trying to beat something slick, fast, simple, and polished most of the time.

I’m a pragmatist, so if you ask me, no, I don’t think that AGI is a requirement for this wave of AI to succeed and to bring about meaningful change. This wave of AI doesn’t represent sentience; it’s more like streaming video. Mind you, streaming video unlocked an enormous marketplace, one that upended all of broadcast TV and continues to grow in unexpected ways. I make this comparison because LLMs change how data is processed, similar to how streaming video changed how footage was transmitted. Both go through a unique process that allows a transformation and ultimately increases the speed, but also the nature, of the original thing. Streaming video turned footage into bits, and LLMs turn words into tokens.

I feel that analogy can give us a sense of what we’re in for in terms of disruption until we land upon a new method of contextualizing our world for machines. That’s really what LLMs are trying to do: use this universal tool, language, to provide context.

This is ultimately why I believe Yann LeCun to be correct about LLMs. I’m very much eager to see what he or others come up with in their pursuit of transforming our world into something comprehensible for machines, but for those of us building products today, there is an opportunity right now to build on something that, while imperfect, has the potential to unlock amazing things, maybe not sentient things, but amazing nonetheless.

Finally, we explore the growing AI public relations challenge: mounting skepticism around AI, the massive data center buildout underway, and whether advances in edge computing could leave the industry with more infrastructure than it ultimately needs.

When LLMs launched to the general public, part of that introduction was introducing the concept of tokens as a means of transacting with the models. At the time, the tokens themselves seemed innocuous enough; it was almost like the next generation of binary code to some, a method of machines understanding things that humans just needed to come to terms with.

As the business of AI matured, the concept of tokens, and more importantly the economy of tokens, remained fluid. For the record, tokens are phrases or bits of text that can be parsed by LLMs (Large Language Models) as part of their natural language processing; this, in turn, allows the LLM to function and make all the interesting connections it goes on to make.

That pricing has fluctuated as the cost of processing and building out the infrastructure of AI has ballooned up. Moreover, the way the tokens are used has changed quite a bit as well; what was once a simple back-and-forth transaction of prompt and response has been replaced by multi-agent, parallel-path, fully agentic workflows, churning through tokens the entire time.

Gizmodo just reported on a lawsuit where Anthropic reportedly misled users about tokens and token usage. Anthropic has been fiddling with their pricing model for a year or so now, and it’s drawn the ire of many a user, including myself. I’m a bit surprised to see them named in a suit like this, but token burn is real, and someone was going to feel it eventually.

Super Powered Approximation Machines

OpenClaw/Moltbot kicked off a fully agentic frenzy that calmed a bit but has not gone away entirely, likely because some form of the agentic workflows in these proto-builds gives us a glimpse of the future, which is why people got so excited by them. For those unfamiliar with OpenClaw, OpenClaw is an open-source framework for integrating a foundational LLM and sending off sub-agents to achieve tasks. This functionality has been essentially replicated by MCPs, though the automations still need some Python scripting and scaffolding to be executed.

What came from this was an obsession with agents and sub-agents: how to fully automate workflows into these machines that don’t need human intervention at all. People only asked if they could build something and stopped asking if they should. What happened was a lot of cognitive surrender to the agents, giving up logical decision-making to the machines and letting them spin until they came to a conclusion. Eventually, the agents would come back with an answer or something close, but only after spending a king’s ransom in tokens.

Frequently, I see Opus put forth in these fully agentic flows, which are totally hands-off. When I’m working with Claude Code, I’m working hands-on and document nearly every session to give Claude the best possible chance at success, and it still makes critical errors. One error just from today: it flat out told me I hadn’t deployed a function that has been deployed for over a month, but its memory had gone stale, so it was about to go on a completely destructive path. I had to stop it, remind it we actually had this function deployed, and send it the error code from the backend so we could get back on track, at which point we were able to resume work.

Using Anthropic’s own diagram’s I want to point out the issues with these workflows, the premise is logical but the conclusion is not. What these workflows assume is that AI has reasoning ability, it does not. AI has more of a funnel or probabilistic method of making decions, so instead of taking a linear course of action, it sort of winds it’s way to a solution.

In this evaluator arrangement, the agents are assigned different roles. The example would be: Agent 1 writes the code, Agent 2 checks it against the PRD, and when the criteria are met, reports back to the human. This isn’t necessarily bad. My only issue is that what is being outsourced to Agent 2 could easily be avoided through clear documentation up front, spending time writing specs, and answering the agent’s questions as it goes. The user is missing out on opportunities to clarify issues like the one I flagged about the stale memory, as well as finding opportunities to improve the flows and learn from the AI as the process progresses.

One final note on this diagram: this loop can go on and on for several cycles before an output is reached. There’s no way to avoid token churn entirely, but in setups like this, there is always a risk of waste.

This next diagram is where things get spicy. This is where token churn is happening. This is where the AI, a machine with no reasoning ability, is being asked to make assumptions, draw conclusions, and extrapolate on whatever the user has input. As it uses its machine learning to connect the dots and plug the holes, it will eliminate possibilities and narrow in on what will eventually be the correct answer, or something like it. What’s interesting about this flow is that they aren’t showing the quality control of the evaluator phase, but I suppose it’s meant to be like assembling different components of a machine.

I saw a version of this diagram where the planning and the research were happening in parallel to each other, which is completely counterintuitive. How is it possible to plan without the research being complete? There seems to be this notion that AI can completely upend all logical constructs, even time-tested sequences, that somehow it’s become stateless and able to execute everything and anything all at once.

Suffice it to say, when I see workflows like this one below, I cringe. There is no code owner here; there is no domain knowledge. Beyond that, the orchestrator stage is burning a fortune trying to compensate for the cognitive offload the human deferred to it.

Artificial Intelligence is a misleading term. It gives the impression that we’ve created a little person inside a machine that’s making decisions; we have not. The best we’ve done is we have created an extremely good and expensive approximation machine. It will get you to the answer, but it will need to approximate along the way, and it will charge you the entire time, where a little planning and human intervention could save you a lot of time, effort, and tokens.

Tokenmaxxing is dead! Long live Tokenmaxxing!

Lest I get tokenmogged for this take, I’m glad tokenmaxxing is dead. Tokenmaxxing was dreamed up by Silicon Valley leaders as a way to track AI adoption and usage, but it was clearly thought up by someone who thought that tokens were free or something. At Meta, they had leaderboards to track token usage, which are now humorously out of commission.

If anything, though, it highlights how little leadership in these companies seems to understand AI, tokens, and the space at large. Like most things, though, it didn’t hit home until the bill arrived. That’s when we started seeing articles about companies budgetmaxxing on their agentic AI spend.

The core of tokenmaxxing, though, is leadership: well, 1. wanting to track their employees in a big-brothery way to see who’s adopting AI; 2. not knowing how to measure adoption of this technology any other way.

As disconcerting as the first point is, the second point is even more so. I think we’ve all been able to see some measurable improvement to some things thanks to AI, but with it we’ve also seen a lot of slop, slippage, and general malaise. AI blends in with regular work so well, and when done well, it fits seamlessly into existing workflows, so how do you even prove you’re using it?

The token economy as the new “cloud”

The branding of “The Cloud” was always a funny way of saying off-site servers, albeit an incredibly effective one. In a way, tokens are the same thing; they’re something very functional that has been abstracted and monetized. Time will tell if this truly is the correct way to value and price AI usage. As we’ve seen with tokenmaxxing, it hasn’t stood up to the test of HR and the workplace yet.

We do know that businesses and users are being charged in this way, though, and that could prove to actually limit the growth of the AI economy. If integrating with foundational LLMs is so expensive and onerous that it becomes a sizable part of a company’s burn, then it’s going to slow the growth of both companies: the foundational LLM provider and the startup.

In building my startup, I’ve chosen to remain LLM-agnostic and have chosen to host open-source options to control burn. As prices fluctuate and features are lobbed off, it’s impossible to build a business on the few leading foundational LLMs. I also think that more foundational models will be around the corner; there’s more disruption yet to come. I could be wrong on all of this, but it’s hard for me to believe this is the final, most perfect version of Large Language Models. At least, those are my two tokens.

Until recently, it’s been a rough time for startups or any company looking to go public. After a deluge of IPOs in 2021 via the traditional practice of pricing an IPO with a banking partner and issuing shares, or the oft-scrutinized and under-regulated process of SPACs (Special Acquisition Companies), the IPO market dried up due to soaring inflation and volatility in the market.

Defying all logic and historical precedent, the market continues to rise, avoiding meaningful corrections, much less the dreaded “bear market,” as America’s K-shaped economy masks many of the historical indicators of a recession. In other words, “markets only go up,” to use a Reddit aphorism, in this new world, but it may not be strictly due to the health and performance of the economy. The market and the investor class have somehow detached themselves from the actual economy as the investments continue to pour in, or at least not leak out.

The economy, the market, and the nation continue to march on despite rising inflation, war, isolation, and disruption to trade. We just received a 6% PPI print that, paired with a 4.2% CPI print, again stimulated hopes of a rate cut. I mention all this to say, not much has changed since 2022. Inflation is still hot, and there’s no functional way to bring inflation down without crashing the economy. The thing is that Wall Street doesn’t care, but moreover, these IPOs may be more about how these companies can’t afford to wait any longer.

Do the Mythos of Anthropic Need Debunking?

I’ve been an advocate of Anthropic and Claude, using it for years to build Group Table. It’s hands down the best foundational LLM and has been even when the narrative changed the “winner” of the AI race about a half-dozen times over the past two years.

Setting my bias aside, Anthropic has been focused. They have decided to own AI coding and branch out from there into other areas of specialty and expertise. If you look at the benchmarks they measure themselves against, it’s mostly specialty work: coding, legal, biology, cybersecurity, health, things that require domain expertise and recall.

Anthropic has been using this expertise and focus to go big-game hunting for corporate money and, without revealing their books, it seems to have been effective, maybe too effective. A company reported burning through $500M in one month due to no Claude usage limits. On a related note, Uber burned through its entire annual agentic-coding budget by the end of Q1.

The benefits of signing up large corporate accounts early on are clear. Deep-pocketed clients can commit to longer and larger contracts than individual users. If Anthropic is successful, perhaps they become the Microsoft of AI with corporate leaders.

These deals come with some undesirable side effects, though. First, several corporate leaders have come out loudly and proudly declaring that they are cutting staff due to AI. So even if Anthropic is immediately benefiting from lavish contracts, the entire space is suffering from what I would call a bit of a PR crisis as the revolt against AI grows ever louder.

The second, somewhat related issue stems from the sales pitch Anthropic’s sales team needed to deliver to land these contracts. Presumably, to codify any sort of commitment, the sales team had to promise some sort of productivity return. Again, Uber COO Andrew Macdonald was cited as saying, “If you’re not actually able to draw a direct line to how much useful features and functionality you’re shipping to your users, that trade becomes harder to justify.”

For me personally, aside from my work at Group Table, AI has sped along several tasks, but it hasn’t outright upended the existing pipelines, at least not yet. The deltas between the numbers laid off, the gains in productivity, and costs associated with those gains are significant. In my contracting work as a product designer, I’m attempting to preach the importance of managing token burn and building efficient workflows and pipelines, which seems to have caught on as token-maxxing appears to be dying off.

As for the company specifically, Anthropic did a bit of a road show, wooing/scaring the geriatrics in Washington with its new Mythos/Fable model. It could be debated how wise it is to brand your own foundational model as “dangerous.” Certainly, this is a PR play; it all is. These are LLMs supported by machine learning at the end of the day, and the whole Wizard of Oz effect can seem outsized in a country that doesn’t take cybersecurity seriously and doesn’t seem like it could if it tried, at least not with the people who currently occupy office.

As of writing this, Anthropic has been forced to roll back the model again, perhaps part of the PR, perhaps the PR backfiring. It’s really so hard to say with the IPO theater that goes on and in this political environment. Truly, anything is possible. That said, needing to roll back a just-released model isn’t a great look, no matter how “dangerous” it is.

All this is to say, with an alienated public due to the PR disaster surrounding “restructuring around AI” and these contracts expiring due to lackluster returns on productivity, I wouldn’t be at all surprised to see some level of pullback at some point for Anthropic. That’s not to say that AI can’t and won’t become a permanent part of workplace culture. In fact, I think it already has. It’s just that the hype and the narrative have gotten ahead of the tech by a large margin.

Is the Window Closing for OpenAI?

Two years ago, the world was declaring OpenAI and ChatGPT the undisputed winners of the AI race. Now it’s looking like the company is in serious trouble, and it has nothing to do with Elon Musk’s lawsuit, though that probably didn’t help anything.

No, a series of unforced errors and strategic missteps have put the company in a difficult position as it helplessly sets money on fire. The company plans to burn through $115 billion through 2029, which was reported last year under rosier conditions.

Since then, the company has released minor improvements to their models to mixed results. Bringing on Jony Ive hasn’t seemed to have helped the company change its fortunes, at least not yet. There was a minor revolt this year when an estimated 2.5 million users unsubscribed from the platform because the COO donated funds to conservative efforts and, by proxy, ICE.

My personal experience as a user is that the product has lost its edge. ChatGPT 4.0 was pointed and would work toward an answer, a solution. Then the subsequent models started aiming toward the middle, regressing to the mean. It seemed like the team was attempting to appeal to everyone and ended up making a product for no one. Now today, the product does this weird, annoying both-sides-ism on literally everything. I’ve formed a habit of turning to ChatGPT, and I keep forgetting how annoying it’s become because it refuses to give me a concrete answer when I’m looking for one. It just gives me both sides and questions me at every turn. It’s not working toward an answer; it’s working toward engagement.

I posted on LinkedIn a couple of months ago in response to a video that was circulating that showed ChatGPT hallucinating a timer result. The original video went viral as “proof” that LLMs “suck” and aren’t ready for showtime. It was evidence of the latter more than the former. My post focused on something I continue to try to stress when people are getting angry about AI: these things aren’t magic; they’re interfaces. In this case, the interface wasn’t hooked up to anything. Unlike deterministic programming that errors out when there should be a dead end, LLMs hallucinate, and that’s what happened here.

Back to the business, though. So awkwardly, in November of 2025, OpenAI’s CFO said an IPO is “not on the cards right now.” Then, in April, it was reported that OpenAI’s CFO and Altman don’t see eye to eye on when to IPO. Then, on June 8th, OpenAI filed to go public. I think it goes without saying that none of this sounds promising at all. As an outside observer, the behaviors don’t suggest a company trying to maximize value, growing every quarter. It feels like a company that’s trying to find an exit. The recent activity with private equity doesn’t instill confidence either. PE has become the death knell of companies, not the harbinger of good things.

Update June 16, 2026: Today it was reported by Tech Crunch that Chat GPT’s market share had fallen below 50% for the first time ever, somewhat confirming the assumptions in this article.

Perplexity and 2028

Perpetually marching to the beat of its own drummer and really in a different league from these foundational models is Perplexity. Being that Perplexity is in the application layer, it makes sense that it would come later.

This isn’t really a statement about Perplexity per se. I don’t know their business model, token consumption, efficiencies, and whatnot, but this is such an important aspect of the AI economy: the applications that are spawned from the foundational models of GenAI.

So, as you may have guessed from my piece, I think these IPOs are premature, but I also don’t think that will matter. There will be pullback and maybe some consolidation, but in the long run, I think this space will thrive. The companies like Perplexity, the companies that develop on these models and find ways to bring the technology to consumers, will create this AI economy, not Anthropic or OpenAI.

However, that hasn’t happened yet, so we have this tension, this angst, and anger. I often cite a fact about the Industrial Revolution, that the Industrial Revolution, specifically textile mills led to job loss initially, but ultimately it led to a fifty-fold increase in productivity.

Now I have to be very honest about that statement. Job loss is real, and it hurts. I’ve experienced it while trying to support a family, and I still haven’t moved on from it emotionally. Also, productivity is a slippery term. Productivity, especially in an AI economy, might not mean jobs, though I ultimately think it will as we introduce more robotics, channels, and ways of doing things.

I don’t know if I’ll be right about AI and the future. I hope I am, but it’s incredibly hard to say. All I know is some very rich and powerful people will be even richer and more powerful after this while we all look on and wonder if the bet they made for all of us will be worth it.