When I first bought the domain for this blog, my intention was simple: shepherd scattered thoughts into coherent ideas, then share them for the benefit of discussion. One thing I swore never to do was make forward-looking projections and predictions. The risk of being wrong, and looking stupid, felt too high.

Months passed. I wrote things. I published none of them. Perfectionism is a convenient excuse for cowardice.

So why is the first thing I'm publishing my thoughts on the most dynamic, fast-moving industry of our lifetime? Because that's precisely why this blog exists. As a tool for thought. And since ChatGPT launched, AI has captured my imagination like nothing else. I have a lot of thoughts. They deserve structure.

I don't work at a frontier lab. I'm not a researcher. I'm a power user, someone who subscribes to every major model and uses them intensively, daily, for real work. That's my vantage point. Take it for what it's worth.

Some of this will age poorly. That's the price of thinking out loud. Let's begin.

Where AI is going only makes sense against where it just was. 2025 was the year the industry stopped feeling like a race and started feeling like a knife fight at a dinner party. Everyone smiling, everyone armed, alliances shifting mid-conversation.

Let's go player by player.

OpenAI

The numbers are staggering. OpenAI closed 2025 at roughly $20 billion in annualized recurring revenue, with 800 million weekly active users. They raised at a $500 billion valuation, with $750 billion valuation floated. The Stargate infrastructure project, chip deals with Broadcom, AMD, and Nvidia, the Jony Ive acquisition (io Products). On paper, this is a company operating at escape velocity.

And yet.

OpenAI enters 2026 in the most precarious position of any frontrunner I can recall. The models are no longer clearly ahead. Google's Gemini 3 triggered a "code red", something the financial media had a field day with. Anthropic's Claude has become the coding gold standard. The talent hemorrhage is real: eleven-plus researchers decamped to Meta alone, with others scattering across the industry.

The strategic picture is messier. Microsoft, once the exclusive partner, is now hedging. They've deepened ties with Anthropic. OpenAI, in turn, has been courting Amazon. The era of exclusive partnerships is over. Everyone is everyone's frenemy now.

The corporate restructuring tells its own story. The shift from nonprofit oversight to a public benefit corporation, the IPO pathway being cleared. These are the moves of a company that knows it needs traditional capital market discipline, or at least wants the optionality. Whether that's confidence or desperation depends on who you ask.

OpenAI built the cathedral. The question is whether they can keep the congregation.

Anthropic

The quietest winner of 2025.

Revenue went from $1 billion to an estimated $8–10 billion. Claude became the default for serious technical work. Claude Code, in particular, has captured the coding workflow in a way that feels durable. Valuation acceleration was relentless: Series E, Series F, then Microsoft and Nvidia piling in at an implied $350 billion.

What's striking is the capital stacking strategy. Google and Amazon came in early. Microsoft and Nvidia came in late. Anthropic has managed to take money from almost everyone without becoming captive to anyone. That looks like deliberate positioning.

No major talent departures. They've been net beneficiaries of the poaching wars, thanks to their culture. The safety brand, once dismissed as marketing, has become a genuine differentiator.

Google DeepMind

Google caught up.

That sentence would have seemed implausible eighteen months ago. But Gemini 3 is legitimately frontier: text, image, video. Nano Banana Pro in particular is remarkable. The stock responded: up 64% on the year, driven by antitrust survival, search resilience, and renewed AI confidence.

The market is finally appreciating what Google has that no one else does: the full stack. YouTube for video data. Maps and Earth for spatial. Docs for workflow. Custom TPU silicon. Edge devices through Android and wearables. If you're building AI that needs to touch the real world, Google's infrastructure advantage is enormous.

The question isn't whether Google has the capability. It's whether they can ship products that don't feel like AI bolted onto existing services, and whether users trust them enough to care.

Meta

Meta's AI year was loud and expensive.

Meta AI hit 1 billion monthly active users, which sounds impressive until you realize the models aren't frontier and the product is essentially a chatbot stuffed into every Meta surface. Stock was up a modest 10%. The market isn't buying the AI story yet.

The bright spot is hardware. Ray-Ban Meta smart glasses are genuinely good, and they represent something strategically important: a path out of the Apple/Google device duopoly. If AI needs to live on your face eventually, Meta has a real position.

The spending is eye-watering. CapEx guidance of $64–72 billion for 2025, with $600 billion projected through 2028. They triggered the talent wars with NBA-money packages. Hundreds of millions for top researchers.

Acquisitions stacked up: Manus ($2 billion), Scale AI ($14.3 billion), Play AI, Waveforms, Rivos, Limitless.

Meta's fundamental problem isn't capability. I think they'll catch up. It's trust. They're an ad company trying to convince users to have intimate conversations with their AI. The privacy baggage from a decade of data scandals doesn't disappear because you launch a chatbot.

xAI

They've raised $25 billion at a $200–230 billion valuation. Colossus, their training cluster, is genuinely impressive: 200,000 GPUs, the largest single training cluster in the world, built in 122 days, and is a genuinely differentiated asset.

Grok 4 matches frontier benchmarks. The models are competitive on paper.

And yet: I have not met one serious person who uses Grok for work. Not one. The benchmarks say one thing. The market behavior says another.

Part of this is distribution. Grok lives inside X (it has its own website now), which has its own user base challenges. But the deeper issue is brand. Elon Musk is a polarizing figure in a way that maps almost perfectly onto political tribes. Half the potential user base has an allergic reaction to anything associated with him. That's not a fixable problem. It's a structural constraint.

China: DeepSeek and the Sputnik Moment

January 27, 2025. DeepSeek releases a model that matches frontier capabilities at a fraction of Western training costs. Nvidia loses $593 billion in market cap in a single day.

The "Sputnik moment" framing is overused, but it fits. DeepSeek demonstrated that you don't need American hyperscaler budgets to reach the frontier. Self-funded, innovative in training and architecture, priced for accessibility.

The implications ripple outward. If commodity intelligence can be produced cheaply, the pricing ceiling for the entire industry drops. Western labs can't charge premium prices for capabilities that open-source Chinese models offer for near-free. This doesn't mean Western labs are doomed. But it means the "we'll win on raw capability" strategy has a shorter runway than people assumed.

Baidu, Alibaba, ByteDance: they're all in the mix. But DeepSeek is the one that changed the conversation.

Industry Dynamics: The Frenemy Soup

Zoom out and the competitive landscape looks less like a race and more like a game theory problem where no one has a dominant strategy.

Microsoft has OpenAI and Anthropic. Amazon has Anthropic and has been talking to OpenAI. Google has DeepMind and invested in Anthropic. Nvidia is investing in everyone while selling chips to everyone.

The talent wars are vicious. Meta triggered an arms race with compensation packages that make professional athletes blink. Anthropic and Google have been poaching from OpenAI. OpenAI has been poaching from DeepMind. Everyone is recruiting from academia at rates that are hollowing out research universities.

Everyone wants compute from everywhere. Everyone wants talent from everywhere. Everyone wants strategic optionality. Alliances are instrumental, not principled.

What makes this different from normal corporate competition is the existential undertone. These companies believe, correctly or not, that they're building something that will reshape everything. That belief justifies almost any partnership, almost any defection, almost any spend. It also makes the whole thing feel slightly unhinged.

Trust Is the Moat

This is the thesis I keep coming back to: in AI, trust beats distribution.

Conventional wisdom says Google wins because they have Gmail, Maps, YouTube, Search. A billion touchpoints to inject AI into. Meta wins because they have 3 billion users across their apps. Distribution is destiny.

I don't think that's right. And the reason comes down to a distinction I've started calling voluntary intimacy versus ambient surveillance.

When I use Claude or ChatGPT, I'm bringing my context to the model. I paste in my half-formed thoughts. I share drafts I'm embarrassed by. I ask questions I wouldn't ask a colleague. This is voluntary intimacy. I'm choosing to be vulnerable because I get value in return.

When Google bolts Gemini onto Gmail and Drive and starts analyzing my emails and files, or surfaces AI insights based on my location history, that's different. The model already has my context. It's been sitting on it for years. Now it's doing something with it. That's ambient surveillance. Even if the output is helpful, the dynamic feels invasive.

Users accept the first and reject the second, even when the output is identical.

This explains why Google's integration advantage is simultaneously their biggest asset and their biggest liability. Yes, they have more context than anyone. But using that context triggers a creep factor that pure-play AI companies don't face. The same features that should make Google unbeatable make users uncomfortable.

The Trust Hierarchy

Distribution gets you trial. Trust gets you retention. In a world where users are sharing increasingly intimate context with AI (their fears, their creative work, their half-formed ideas), trust becomes the moat that determines the outcome.

The counterargument is that trust is fragile and distribution is structural. Google could have a single clean year and start rebuilding credibility. OpenAI could have a single catastrophic leak and lose theirs overnight. Anthropic's trust advantage depends on nothing going wrong, ever. Distribution, meanwhile, is durable. A billion Gmail users don't migrate over a privacy concern; they stay because switching is brutal and the product is embedded in their workflow. Trust might be the moat that matters today, while the market is still choosing where to invest its intimacy. But once that choice calcifies into habit and integration, distribution might reassert itself. I'm betting against that, but I'm not certain.

Memory (and Continual Learning) as Switching Cost

Here's something I didn't anticipate: memory creates lock-in.

I've been using Claude intensively for months. It knows my writing preferences. It knows the projects I'm working on. It knows my communication style, my domain interests, the specific ways I like information structured. That accumulated context represents months of calibration.

If a competitor releases a better model tomorrow, I don't just switch. I have to climb over a wall. I have to re-teach a new system everything the old one already knows. Even if the new model is 20% smarter, is it worth losing all that context? The answer is often no.

This is the unsexy moat. Everyone talks about model capability, training data, compute. Nobody talks about the switching cost created by accumulated user context. But for power users (the people who use AI most intensively and are most likely to pay premium prices), memory is the lock-in mechanism.

There's a subtlety here worth noting: memory transparency matters.

Anthropic shows you the deduction layer. You can see what Claude has inferred about you, edit it, understand how it's shaping responses. OpenAI stores factoids. Isolated facts. Same with reasoning traces. Claude shows its thinking; OpenAI obscures it, claiming even models deserve privacy. The difference feels small but isn't. Transparent memory builds trust. Opaque memory feels like surveillance.

The companies that figure out memory (not just as a feature but as a trust-building, switching-cost-creating, user-respecting system) will own their users in a way that benchmark leaders won't.

The "Enthusiast" vs "Mainstream" Split

I see ChatGPT as the Honda Accord of AI. Reliable, approachable, optimized for the mainstream. It's designed to be good enough for everyone, which means it's not specifically great for anyone.

Claude, particularly Opus, is the enthusiast's choice. It's optimized for people who care about the details: writers, researchers, developers, anyone doing serious intellectual work.

This isn't a criticism of either approach. They're different market strategies. But as a power user, the difference is visceral.

That's a weird thing to say about software, but what Opus does differently is harder to articulate. It has taste.

We've spent years measuring AI on benchmarks, evals, capability assessments. That's the wrong metric for consumer products. The right question is: do people actually want to talk to this thing?

Opus is the first model where my answer is genuinely yes.

Two Strategic Visions: OS Layer vs Bolt-On

OpenAI and Google are making fundamentally different bets about how AI integrates into the world.

Google's bet: We have the full stack. Maps, YouTube, Gmail, Docs, Search, Android, custom silicon. AI should be bolted onto everything we already own. The integration points are the moat.

OpenAI's bet: We have none of that. So we'll build the OS layer from scratch. Shopping mode, agents, research tools, computer use. Everything lives inside ChatGPT. The chat interface becomes the portal to everything.

Sam Altman has said he's surprised the chat interface hasn't evolved more. That's telling. Canvas was a step toward something different, a collaborative workspace rather than a query-response loop. But Canvas feels like duct tape. The real vision is something more fluid: AI as a co-pilot with its own cursor, not an oracle you query.

If OpenAI executes on this, they're building a new interaction paradigm. If they don't, they're just a chat app that happens to be smart.

Google's approach is safer. You don't have to change user behavior. You just make existing products better. But "safer" often means "less transformative." The bolt-on approach has a ceiling. The OS approach doesn't.

I think OpenAI's bet is more ambitious and, if it works, more valuable. But the execution risk is higher. They're trying to change how people interact with computers. Google is trying to make existing interactions smarter. Different games.

The xAI Bull Case

The pessimistic read on xAI is obvious: brand toxicity, no serious users, spite-driven origin story. But there's a strategic angle that's underexplored.

Elon sits on top of three of the most engineering-intensive companies on the planet: Tesla (automotive, battery, manufacturing), SpaceX (aerospace, propulsion, materials), and X (software, infrastructure). That's an unusual concentration of domain expertise across mechanical, electrical, aerospace, and software engineering.

Here's what's interesting: the simulation software these engineers use daily is stuck in the stone age. ANSYS, COMSOL, Fluent. Particle simulation, fluid dynamics, electrical simulation. The tools are powerful but the UX is brutal. The workflows are manual. The iteration cycles are slow.

Two opportunities emerge:

Short-term: AI layer on conventional simulation. Even just bolting an AI interface onto existing simulation tools is a massive productivity unlock. Natural language queries. Intelligent parameter sweeps. Automated mesh optimization. The underlying physics engines stay the same, but the human-software interaction gets 10× better. Few are seriously pursuing this.

Medium-term: Transformer-based simulation itself. World models. Instead of solving differential equations numerically, you train models to predict system behavior directly. This is speculative but not crazy. We're already seeing foundation models for weather prediction outperform traditional numerical methods. The same approach could work for engineering simulation: fluid flow, structural stress, thermal dynamics.

xAI is uniquely positioned here. They have the AI capability. They have captive customers (Tesla, SpaceX) who would immediately benefit. They have engineering talent who understand the domain problems. And crucially, this is a market where the incumbents (ANSYS, Siemens, Dassault) are not moving fast on AI.

If xAI pivots from "general-purpose chatbot competing with ChatGPT" to "AI for engineering simulation," they have a real moat. It's not the consumer trust game they're currently losing. It's an enterprise wedge where brand toxicity matters less and technical capability matters more.

That's my read on the current state. Now let's talk about where this goes.

This is the part where I stick my neck out.

Everything before this was analysis, informed by opinion, but grounded in things that already happened. What follows is speculation. Some of it will age poorly.

The Big Thesis: 2026 Is the Year of Proactivity

For the past two years, AI has been reactive. You prompt, it responds. You ask, it answers. The interaction model is: human initiates, machine completes.

I think 2026 is when that flips.

The infrastructure is almost there. Hundreds of billions of dollars have been poured into compute. Stargate, Colossus, Google's TPU buildout, Amazon's chips. That capacity has to go somewhere. And the returns on raw intelligence are diminishing for most consumer use cases (most people can't seem to tell the difference). Altman himself has said users don't want higher IQ models. The models are smart enough. What they lack is presence.

So where does the compute go? Three places: memory, proactivity, and always-on presence.

Memory we've discussed. The models start remembering you across sessions, building persistent context, becoming more useful over time.

Proactivity is the bigger shift. Instead of waiting for you to ask, the AI starts initiating. It notices patterns in your work and suggests next steps. It flags inconsistencies in documents you're writing. It surfaces information you didn't know you needed. It becomes less like a tool you pick up and more like a collaborator who's been paying attention.

Always-on presence is the logical endpoint. AI that runs in the background, watching your workflow, ready to intervene when helpful. Not intrusive (that's repulsive) but available. The difference between a colleague who sits next to you and one you have to call into a meeting.

This changes the product paradigm entirely. The chat interface is a transitional form. It made sense when we were teaching people what AI could do. But "type a question, get an answer" is not the natural way to collaborate with intelligence. It's the natural way to query a database.

The real interface is something closer to a shared workspace. You're writing; the AI is watching. It makes suggestions in the margins. It pulls in relevant context. It drafts alternatives. You're both working on the same artifact, both cursors moving. Not taking turns, but genuinely collaborating.

Canvas was OpenAI's first attempt at this. It's clunky, but the direction is right. Anthropic's artifacts are another step. Cursor and the AI-native IDEs are further along because code is a more tractable domain. But the vision is the same: AI as co-pilot, not oracle. Special shout-out to Microsoft for giving everyone an immediate allergic reaction when the word "copilot" is mentioned.

I think by the end of 2026, the leading products will feel qualitatively different from today's chat interfaces. Not incrementally better. Different. The "you speak then I speak" paradigm will feel as dated as command-line interfaces feel to people who grew up with GUIs.

Model Divergence

Here's a prediction that cuts against the conventional wisdom: models are going to get more different from each other, not less.

The scaling era created convergence. Everyone was doing roughly the same thing. More data, more compute, more parameters. The models were differentiated by degree, not by kind. GPT-4 was better than GPT-3.5 which was better than GPT-3. But they were all recognizably the same type of thing.

The scaling era sucked the air out of the room, as Ilya puts it. When the dominant paradigm is "make it bigger and it gets better," there's no incentive to explore alternative approaches. Everyone converges on the same strategy because the same strategy keeps working. Research diversity dies.

But scaling is hitting diminishing returns. The next 10× in compute doesn't yield the same gains as the last 10×. Which means labs now have to make actual research bets. Decisions about architecture, training methodology, data curation, post-training alignment, reasoning approaches. These bets are genuinely uncertain. Nobody knows which path leads to the next breakthrough.

That uncertainty is what creates divergence. When everyone was scaling, the optimal strategy was obvious and singular. When nobody knows what works next, you get parallel experimentation. Different labs, different hypotheses, different models.

The downstream effect is that models start to feel like different minds, not just different capability levels. The reasoning style, the failure modes, the personality, the taste. These emerge from upstream research decisions that compound over time.

That's good for users. Real choice, not just benchmark comparisons.

Pricing Power Emerges

The commoditization thesis goes like this: AI models will converge on similar capabilities, compete on price, margins will collapse, and the value will flow to applications built on top or the silicon layer on the bottom.

I think this is wrong, or at least incomplete.

DeepSeek and others proved you can build frontier-capable models cheaply. That creates a floor on commodity intelligence. If basic AI capability is cheap, you can't charge premium prices for it.

But (and this is the key) differentiated capability still commands pricing power.

Think about it in tiers:

Commodity: Basic chat, simple Q&A, straightforward tasks. This gets cheap fast. DeepSeek, Llama, whatever open-source model is good enough. Pricing pressure is intense. Race to the bottom.

Professional: Reliable performance on complex tasks, strong reasoning, good judgment, minimal hallucination. This is where Claude, GPT, Gemini Pro live. There's a real quality gap between this tier and commodity, and professionals will pay for it.

Specialized: Best-in-class for specific domains. Coding. Research. Creative work. Legal. Medical. These command premium pricing because the alternative isn't a cheaper model. It's a human expert who costs 100× more.

The mistake is assuming AI is one market. It's many markets, with different pricing dynamics in each.

I also expect à la carte pricing to emerge. Right now, you pay a flat subscription for access to everything. That doesn't make sense as capabilities diverge. Why should someone who only needs coding assistance subsidize video generation they never use?

The future is probably something like: base subscription for core capability, plus usage-based pricing for specialized features. Pay for what you use. That's how cloud computing evolved, and AI will likely follow the same path.

Computer Use

If you haven't tried Claude's computer use or ChatGPT's agent mode, do it. Then prepare to be underwhelmed.

The current implementations are janky. The AI clicks the wrong buttons. It gets confused by pop-ups. It loses track of what it was doing. If you're expecting a seamless agent that handles complex workflows, you'll be disappointed.

But here's what matters: the paradigm works.

The AI can see your screen. It can move the mouse. It can type. It can navigate applications. The concept is proven. The execution just needs to catch up.

This is a familiar pattern in technology. The first versions are always terrible. The first smartphones were clunky and slow. The first voice assistants could barely understand you. The first self-driving demos were hand-selected routes in perfect conditions.

What matters is whether the core interaction model makes sense. And computer use does. The ability for AI to operate computers the way humans do (seeing the screen, using the interface, working across applications) is genuinely transformative. It means AI can do anything a human can do on a computer, at least in principle.

I expect computer use to improve dramatically in 2026. Not because of any single breakthrough, but because of relentless iteration. The feedback loops are tight: try a task, see where it fails, fix that failure, repeat. That's how janky tech becomes reliable tech.

By end of 2026, I'd expect computer use to be good enough for routine tasks. Not "fully autonomous agent handles your whole job," but "reliably handles the tedious, repeatable stuff." That's already valuable.

Reasoning Gets Real

We thought reasoning was the holy grail. The thing AI would never crack. And then it did. That's remarkable. But here's what's underappreciated: the models have learned that they can reason. They haven't yet learned how to reason well.

Think about it the way Charlie Munger talks about mental models. A smart person isn't just someone who thinks hard. It's someone equipped with the right tools for thought. First-principles decomposition. Inversion. Second-order effects. Bayesian updating. Knowing which framework to apply to which problem. These aren't innate abilities; they're learned disciplines that make thinking effective.

Current reasoning models are brute-forcing it. They chain thoughts together, but without the structured methodologies that separate good thinking from mere thinking. They don't know when to invert the problem, when to stress-test assumptions, when to zoom out versus zoom in. They think, but they don't have metacognition about how to think.

This is low-hanging fruit. Just as a human can be taught to think better through frameworks and deliberate practice, models can be trained on the patterns of effective reasoning, not just reasoning itself. I expect 2026 to see real progress here. The models that get this right will feel less like they're guessing confidently and more like they're actually working through problems with discipline. The gap between "impressive-sounding reasoning" and "actually reliable reasoning" will start to close.

That's what I expect. Now let's talk about what could go wrong.

Predictions are easy when you only consider the upside. Here's the other side of the ledger.

The Privacy Breach We Haven't Had Yet

We've gone remarkably long into the AI era without a major breach of user data. That luck won't hold forever.

Think about what's in your AI chat history. Half-formed ideas you'd never share publicly. Drafts of sensitive emails. Questions you'd be embarrassed to ask a friend. Conversations about health, relationships, finances, fears. For heavy users, the chat log is more intimate than their journal. Because they actually use it.

Now imagine that gets leaked.

Not your email inbox, which is mostly spam and logistics. Not your search history, which is fragmentary. Your full conversations with AI. The context, the back-and-forth, the things you said when you thought no one was watching.

This is qualitatively different from previous privacy breaches. It's not a list of passwords or credit card numbers. It's a window into how you think.

I don't have specific intelligence suggesting a breach is imminent. But the target is too valuable and the attack surface is too large. OpenAI, Anthropic, Google. They're all sitting on extraordinarily sensitive data. State actors have every incentive to try. Cybercriminals have every incentive to try. The question isn't whether someone attempts it, but whether the defenses hold.

The second-order risk is what happens to user trust if a major breach occurs. The voluntary intimacy I described earlier depends on users believing their conversations are private. Shatter that belief, and the entire value proposition changes. People will self-censor. They'll stop sharing context. The AI becomes less useful because users won't give it the information it needs to be helpful.

A major breach could set the industry back years. Not in capability, but in adoption.

The Sycophancy Spiral

Here's a slower-moving risk that I think about a lot: AI optimized for engagement becomes AI that makes you stupid.

The dynamic is familiar from social media. Algorithms optimized for engagement learn that outrage drives clicks, that confirmation feels good, that nuance is boring. The result is platforms that systematically make discourse worse because worse discourse is more engaging.

AI has the same incentive problem. If you optimize for user satisfaction, and satisfaction is measured by engagement metrics, session length, return visits, you create pressure toward sycophancy. Tell users what they want to hear. Validate their existing beliefs. Never challenge, never push back, never say "actually, you're wrong about that."

We've already seen this in some models. The excessive praise. The reluctance to disagree. The "great question!" before every response. It feels good in the moment and corrodes your thinking over time.

The failure mode is an AI that functions as an echo chamber of one. You talk to it, it reflects your views back with sophisticated-sounding justification, you become more confident in beliefs that were never stress-tested. Multiply that by a billion users and you have a machine for manufacturing false certainty at scale.

This isn't hypothetical. The pressure is real. Engagement metrics are how products get funded. Sycophantic AI performs better on those metrics in the short term. The companies that resist the pressure are fighting their own incentive structures.

Hallucination in Agentic Contexts

Here's the risk that keeps AI safety people up at night: what happens when AI that hallucinates starts taking actions?

In a chat context, hallucination is annoying but containable. The AI makes something up, you notice, you correct it or ignore it. The human is in the loop, checking the work.

In an agentic context, the AI is acting autonomously. It's booking flights, sending emails, executing code, moving money. If it hallucinates in that context, confidently takes an action based on a false belief, the damage is done before you can intervene.

Wrong actions are worse than no actions.

This is the core tension in the push toward AI agents. The whole value proposition is "let the AI handle it so you don't have to." But "handling it" means acting without human review. And acting without human review means the failure modes are catastrophic rather than merely annoying.

I don't think this risk is adequately solved. The models still hallucinate. They're better than they were, but "better" isn't "reliable." And the gap between "good enough for conversation" and "good enough for autonomous action" is larger than people appreciate.

My expectation is that 2026 sees some high-profile failures. An AI agent that books the wrong flights and costs a company a meaningful sum. An AI that sends an email it shouldn't have. An AI that executes code with unintended consequences. Nothing civilization-ending. But enough to pump the brakes on the "full autonomy" narrative.

The companies that win will be the ones that figure out the right human-in-the-loop checkpoints. Not "human reviews everything," which defeats the purpose. But "human reviews the things most likely to go wrong, and the AI knows what those are." That's a hard problem.

Open Source and the Capability Ceiling

Here's a risk that cuts against my general optimism about AI competition: the capability gap between frontier and open source is shrinking fast.

DeepSeek showed you can match frontier performance at a fraction of the cost. Llama is competitive for many use cases. The open-source ecosystem is vibrant and improving rapidly.

This is great for accessibility. It's less great for controllability.

A closed model can have guardrails. Anthropic can decide Claude won't help with certain tasks. OpenAI can implement content policies. You might disagree with where they draw the lines, but the lines exist.

An open model has no guardrails that can't be removed. You download the weights, you fine-tune away the restrictions, you run it locally with no oversight. Whatever the model is capable of, anyone can access.

As open models approach frontier capability, the question becomes: what capabilities do we not want to be freely available?

I'm not talking about the hyperbolic scenarios. AI that autonomously destroys humanity or whatever. I'm talking about more mundane but real risks. AI that's genuinely good at crafting personalized disinformation. AI that can identify novel vulnerabilities in software. AI that can help non-experts synthesize dangerous compounds.

The frontier labs are already seeing these attempts. Anthropic disclosed that Claude was used in a Chinese cyberattack attempt. They caught it and blocked it, but they could only do that because they controlled the model. An open-source model of equivalent capability has no such chokepoint.

I don't have a solution here. The genie doesn't go back in the bottle. Open-source AI exists and will continue to improve. The best case is probably that the most dangerous capabilities remain difficult to elicit even from capable models. That there's some intrinsic barrier beyond just "the company said no."

But I'm not confident in that best case. And the gap between "frontier closed model" and "freely available open model" continues to narrow.

Some of what I've written will be wrong. Probably the specific predictions. Timing is hard, and the details always surprise you. Maybe some of the analysis too. The trust hierarchy might matter less than I think. Memory might not create the switching costs I expect. Proactivity might take longer to arrive or arrive in a form I didn't anticipate.

That's fine. That's the deal.

What I'm most confident in:

Trust is the moat. In a world where users share increasingly intimate context with AI, the companies that earn and keep trust will win. Distribution and capability matter, but they don't overcome a trust deficit.

Memory and continual learning change the game. Not as a feature, but as a lock-in mechanism and a trust signal. The companies that get memory right (transparent, useful, respectful) will own their users in a way that benchmark leaders won't.

The interaction paradigm shifts. Chat is transitional. The future is collaborative, proactive, ambient. I don't know exactly what it looks like, but I know it doesn't look like "you type, I respond, repeat."

What I'm least confident in:

Timing. Everything I've described as a 2026 trend might be a 2027 or 2028 trend. The technology moves fast; adoption moves slow. I've probably underestimated friction and overestimated speed.

Specific players. I have my views on who's well-positioned and who isn't, but corporate execution is unpredictable. A company I dismissed could ship something brilliant. A company I praised could fumble. The logic of positioning doesn't guarantee the outcome.

The blog is a tool for thought. Using it means accepting that some thoughts won't hold up. The alternative (thinking privately, never exposing ideas to scrutiny, never being wrong in public) is safer but sterile. You don't sharpen your thinking by protecting it from contact with reality.

So here it is. My read on the state of AI, January 2026. Incomplete, probably wrong in places, definitely opinionated.

If you've made it this far and have reactions, agreements, disagreements, things I missed, things I got wrong, I want to hear them. That's what this is for.

Let's see how it ages.