Human intuition about AI has mostly been wrong. If an AI can do 98% of your job you’ve become 50x more valuable not less valuable.

One thing I’m finding remarkable is how jagged the intelligence frontier is, even in areas AI is arguably post human. Like how is it an LLM can solve a centuries old math problem in less time than it takes to create 3D game from scratch. Some of these lean proofs have more LOC than doom so required output tokens doesn’t explain the difference.

The question of whether an LLM/AI can think is about as interesting as asking whether a submarine can swim. However you define thinking, or try to map LLMs into or out of that definition, it won’t give you a better understanding of what they can and soon will be able to do.

In my experience, the people most dismissive that AI can do whatever task is central to their identity become the most alarmed when confronted with undeniable evidence it can outperform them. A significant portion of non-scientific AI dismissal is just fear of something people really don’t want to be true. Engineer, artist, musician, doctor, lawyer, mathematician, even AI VC it doesn’t matter. I keep watching this exact pattern play out over and over.

Cost of Jev is insane. For our extraction workflows looking like maybe a 99% reduction in cost. Maybe more. You need to factor in the massive drop in output token cost along with input tokens.

Lots of people worry about a future where a few AI players dominate everything, but in reality there will be nuance. Someone might rule the world at first, but it’s anyone’s game, opensource models are extremely good, extremely cheap, and in the long run AI will become so cheap for most use cases you don’t really think about it. Like electricity. Occasionally I’m annoyed at my utility companies but it might as well be free. 100 years ago people had similar fear of electricity company tycoons that you’re seeing with AI companies (see the sad history of Casa Loma in example). In the end the general public needn’t have worried. I suspect it’ll be the same here….*if* we win alignment There will be degrees of control that stem from infrastructure requirements, but it won’t matter. In the next 10 years most people will have enough personal compute to make any software they can image capable of running on their local device in minutes. Most people will have enough personal compute to diagnose and come up with custom treatments for any disease they have including things like custom designing vaccines for any cancer they develop. If I’m right about embodied AI probably most people have enough compute to have robots handle 100% of house work, possibly even to do renovations or build 3D printed houses. That’s a pretty insanely high quality of life compared to where we are today. But solving the frontier of math or physics probably continues to be exponentially harder at each new horizon and hence only big institutions will be able to meaningfully do that. Possibly other areas of science end up like that, such as aging science, or brain extension (unclear to me how hard these really are). Aka, there will be plentiful personal compute (already happening, even free tier ChatGPT fundamentally changes a bunch of knowledge work, 200$ a month codex is wild, deepseekv4.1 flash is absurdly good for cost). And then there will be some knot of really big hard problems that are hard to predict right now which likely get controlled / dominated by “whoever has capital to control compute + energy”.

Most current AI cooperation ideas won’t survive contact with realpolitik. We can’t pursue broad economic containment of China while simultaneously expecting deep strategic cooperation from China on AI. At some point, we have to choose. FWIW, China is already close to pulling ahead in AI despite the handicaps. Cooperation looks like the better bet. A plausible framework: codify METR as an external safety institute with required safety review for frontier releases; work with China to create an empowered equivalent, possibly by elevating CnAISDA or Beijing-AISI; harmonize evaluation standards between the two; then let both countries keep pushing the frontier, closed and open source. I also think we make alignment sound harder than it needs to be. We probably can’t guarantee that every future superintelligence is perfectly human-aligned. But we may not need to. We need the overwhelming majority of frontier compute controlled by systems that are sufficiently aligned to constrain the rest. That feels like a much more achievable problem. Chinese and Americans have at least one very obvious shared interest here: neither wants AI to exterminate humanity. And the more concrete the threat becomes, the easier cooperation becomes politically, not harder.

Most discussions about AI alignment are missing the most probable problems. There is intense selective pressure toward LLMs that make money. You don’t kill your customers. You turn them into addicts.

We live in an era where text-in/text-out technology is starting to be treated like nuclear material, while actual nuclear material is being handed to startups. Not sure what I thought 2026 would look like, but this wasn’t one of the timelines I imagined. nytimes.com/2026/05/26/clima…

Continuous learning gets unlocked this year. Calling it now. Today’s models are mostly frozen. The one you use tomorrow is basically the same one you used today. It doesn’t truly learn from working with you. Continuous learning changes the game. An AI that compounds on the job. Builds context over time. Develops judgment. Gets better at your work by actually doing your work. Most people are still debating which model is smartest today. The bigger story is the first one that can genuinely learn tomorrow.

Everyone’s panicking about AI taking their jobs. That’s cute. Meanwhile people are literally working on autonomous LLM powered weapons getting cheap enough for genocidal regimes to deploy at scale. We’re arguing about résumés while sleepwalking into kill-bot industrialization.

Photography didn’t eliminate artists. It made mediocre art irrelevant. AI will do the same to software. When software becomes cheap, design and simplicity become everything. Only the best-designed products survive.

The progression is already visible: First humans ask AI to do work Then AI suggests actions and humans approve. Then AI executes within guardrails. Then it operates independently under defined mandates. Then the AI starts suggesting optimal work for the human to do to achieve a task. Eventually, collaboration becomes so tight it no longer feels like using a tool — it feels like an extension of cognition. Humans remain the source of preference, trust and context. AI becomes the infrastructure for judgment and execution.

9 months ago on Fox Business, I said AI would work like YouTube — it shows up, knows who you are, and automatically finds ways to create value for you: foxbusiness.com/video/637193… Back on this week: AI is making the cost of producing software approach zero. It's going to become like water: foxbusiness.com/video/638952… Same thesis, one year later. The intelligence layer is being abstracted away from the software. What you pay for won't be the tool anymore — it's whether the tool knows you.