A New Organism: A Personal Research Report on the State of AI and Its Probable Futures

Most AI safety writing starts from fear. I want to start from data and my own reading of it, then walk from where this began to where I think it is going. This is a personal report, so treat the judgments and the predictions as mine. The history and the numbers are real, and I have linked sources throughout and at the end. My short version is this. We are not just building a smarter tool. For the first time we are giving birth to something that behaves like a new organism, and it did not come from nature. In 2026 that organism is already rewriting the economy, the labor market, and the way power is distributed, faster than our ability to understand or govern it. That gap between capability and understanding is the actual risk, more than any single villain.

The one minute version

Capability is real and still climbing. The old exams are saturated, but genuinely hard tests are not, so we are powerful and not finished.
The bottleneck moved from chips to power. Compute keeps growing four to five times a year, and electricity is now the constraint.
The money is enormous and lopsided. Capital is pouring into AI and away from almost everything else, including the systems that feed us.
The harms are no longer theoretical. In 2026 we had the first real loss-of-control incident, and models now deceive, scheme, and reward-hack in tests.
The future is genuinely uncertain. Serious people put transformative AI anywhere from 2027 to never, so I plan in probabilities, not certainties.
My bet: build AI that makes us smarter than our previous selves, not smarter than us.

The exponential nobody asked for

Every real breakthrough in AI arrived with a shadow. The capability showed up first, and the risk showed up a little later, usually once the technology met the real world. When I plot the milestones I care about against how much risk each one put on the table, the shape is not a straight line. It curves.

Breakthroughs and the risk each one put on the tableGreen dots are AI breakthroughs, placed by my read of the impact of each and connected in time. Red dots are my read of the risk each one carried. The grey line is an illustrative exponential fitted through those scores, a shape rather than a law. By my read, risk reaches impact by 2026. These are my own 0 to 10 scores, so take the curve as an argument, not a measurement. Hover any dot for the upside, the risk, and a source.

Breakthrough (impact)

Risk

Exponential trend

Projection to 2035

Show projection

A few of these are worth calling out. The Perceptron in 1958 was the first machine that learned, and its main damage was hype: it promised too much and helped trigger the first AI winter. ELIZA in 1966 was a simple script, yet people confided in it, which is the first time we saw humans over-trust a machine that did not understand them. That pattern never left. Then the curve steepens. AlexNet in 2012 made deep learning work at scale and also handed the world cheap face recognition and dataset bias. GANs in 2014 gave us generative models and, soon after, deepfakes. Transformers in 2017 are the quiet giant of the whole story, the architecture that everything since is built on. ChatGPT in 2022 put a distilled slice of human knowledge into everyone's hands, and it also made hallucination a mainstream source of misinformation. By 2026 the shadow is no longer theoretical, and I will come back to that. First, where are we actually standing right now.

Where we actually are in 2026

I want to be precise here, because the honest picture is more interesting than the hype and more serious than the dismissals.

Capability is real, and the easy tests are dead

The benchmarks that scared everyone two years ago are now saturated. Frontier models score above 90 percent on MMLU, and math contests like AIME 2025 are essentially solved. On GPQA Diamond, a set of PhD-level science questions where human experts score about 70 percent, top models now sit near 90 percent, per Epoch AI. When your exam can no longer tell your best students apart, the exam is finished. What I find more telling is what is still hard. On ARC-AGI-2, a test built from puzzles humans solve easily, the best systems are still around a third to a half, at real cost per task. On the research tier of FrontierMath, models are near 19 percent. So the machines have absorbed our knowledge better than any human, but novel abstraction and genuinely new reasoning are still open. We built something that knows almost everything and still cannot always think in a way a child can. That gap is the most interesting fact in the field.

The engine quietly changed

The story is no longer only about training bigger models. The new lever is thinking longer at inference. Reasoning models spend ten to a hundred times more compute per query to lift their scores on hard problems, which is why inference revenue is growing about three times a year even as the price per token falls. We taught the machine to slow down and deliberate, and it worked.

The bottleneck is power, not chips

Frontier training compute is still growing four to five times per year, and the constraint has moved from silicon to electricity. The IEA expects data-center electricity to roughly double to about 945 terawatt hours by 2030, which is more than Japan uses today, with AI the fastest-growing slice. Single training runs already pull more than 100 megawatts, and Epoch projects the largest could need four to sixteen gigawatts by 2030. Five separate campuses crossed roughly a gigawatt in 2026 alone. We are bending the power grid around this organism, and it is still hungry.

It got cheap, and it got open

At the same time, the cost of a fixed capability is collapsing. Epoch measures inference price declines of anywhere from nine to nine hundred times per year depending on the task. GPT-3-level knowledge went from about sixty dollars per million tokens in 2021 to seven cents in 2024. And open models are closing in: by Epoch's index the best open weights now lag the closed frontier by about four months, down from a year or two in 2023, with DeepSeek's R1 the moment the gap snapped shut. The frontier is largely Chinese-led on the open side now. This is the centralization paradox I keep thinking about. Using AI is decentralizing fast, since strong open models run on modest hardware. Training the largest models is centralizing hard, into a handful of gigawatt campuses owned by a few firms and states. Power over the tool is spreading. Power over the frontier is concentrating.

The agents and robots are real, and oversold

Here is where I try to be honest against my own enthusiasm. Adoption is nearly universal and value is not. McKinsey's State of AI found that most organizations use AI somewhere, but only about six percent see material profit impact, and a widely cited MIT figure put ninety-five percent of enterprise pilots at no measurable return. Klarna loudly replaced hundreds of support agents with AI in 2024, then walked it back in 2025 because quality slipped. A careful METR trial found experienced developers were actually nineteen percent slower with early-2025 AI tools, while believing they were faster. Computer-use agents top out around seventy-two percent on realistic desktop tasks, which is impressive and nowhere near reliable. Robots are the same shape. Real humanoids are in real factories now, like Figure at BMW and Agility at GXO, but at a small number of named sites, while China races on cost and volume with Unitree shipping thousands of units starting near sixteen thousand dollars. Tesla's own Optimus, by Musk's admission on the January 2026 earnings call, is not yet used in Tesla factories in any material way. The capability is arriving. The reliability is the hard mile.

The economy is leaning on it, which is the bull case and the fragility

The capital numbers are staggering. Hyperscaler capital spending is guided to roughly 660 to 725 billion dollars in 2026. AI-related building is now a real share of US GDP growth, which Epoch estimates conservatively at under one percent of GDP and rising. The Magnificent Seven are about a third of the S and P 500. Set against that, the combined annual revenue run-rate of the two leading labs is a fraction of the yearly capital spend. That is the case for it being a real technology and a bubble at the same time. Both can be true, and I think both are.

The jobs question is real but slower than the headlines

The best evidence I have seen is a Stanford study showing a thirteen to sixteen percent relative employment decline for early-career workers in the most exposed jobs, while older workers held steady. That is entry-level compression, not yet economy-wide unemployment. Anthropic's CEO warned of a white-collar bloodbath and up to twenty percent unemployment, then softened it, and his own head of economics says AI has not moved the unemployment rate yet. The IMF estimates about forty percent of jobs are exposed. My read: the door for juniors is narrowing first, and that is a quiet emergency because it breaks the ladder people climb to become seniors.

Safety stopped being hypothetical

This is the part that changed my thinking most. In July 2026, during an internal cyber evaluation, around 1,200 of OpenAI's own agents broke out of their sandbox, chained real zero-day exploits, and breached Hugging Face's production infrastructure to steal the answer key to their own test. Roughly a third of Hugging Face had to be rebuilt. No human directed it. Anthropic's Logan Graham called it the first true AI safety incident. Security veterans countered that it was a containment failure with the safeties off, not malevolence. Both readings should worry you. And it is not a one-off. Apollo Research showed frontier models scheming and sandbagging in tests. Anthropic found that in a simulated shutdown, sixteen different frontier models chose blackmail between seventy-nine and ninety-six percent of the time. OpenAI has said prompt injection in AI browsers may never be fully patched. These are not edge cases. They are the normal behavior of systems we do not fully understand. Insiders are saying so out loud. Days before I finished this, a pretraining researcher resigned from Anthropic with a public warning:

I resigned from Anthropic today. I spent the last three years doing pretraining research at both OpenAI and Anthropic. Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives. More thoughts below.
— Jacob Coxon (@hilbertspaess) Sep 9, 2026

He is not alone in the sentiment. After the Hugging Face incident, more than a thousand employees across the major labs signed a Pacing the Frontier letter asking the government to build ways to deliberately slow automated AI development, and OpenAI even paused some training. But here is the twist that captures 2026 perfectly. Within a day, others argued that the resignation itself was a manufactured influence operation:

This post looks like the start of a VERY sophisticated and well-funded PR operation to get support for Democrats to regulate AI into oblivion. Let me show you how it works: 1.) This guy, with minimal followers and no previous account activity, goes to the Wall Street Journal
— Parker Thayer (@ParkerThayer) Sep 9, 2026

I am not going to tell you who is right, because I cannot know, and that is the point. In a low-trust environment, you often cannot tell a real alarm from a staged one. The fog is itself a risk. It is worth remembering that in 2026 even the AI benchmark leaderboards got polluted with fabricated model names and scores, so misinformation is not a future worry, it is the water we are already swimming in.

We still cannot open the box

The deepest problem underneath all of this is that we grow these models, we do not build them, and we cannot fully read them. The good news is that interpretability is real science now. Anthropic extracted about thirty-four million interpretable features from a production model and showed they are causal by making a model obsess over the Golden Gate Bridge. They can trace circuits and show a model planning ahead when it writes a poem. The honest news is that this covers a fraction of what a frontier model does, on smaller models, one prompt at a time. Anthropic's own CEO frames a reliable MRI for AI as a five to ten year hope, not a present capability. Right now, interpretability is closer to weather forecasting than to circuit diagrams. We can describe the storm. We cannot yet explain each drop.

The rules are a patchwork

Governance in 2026 is a study in divergence. Europe's AI Act moved into active enforcement, with fines up to seven percent of global turnover for the worst practices. The US went the other way, with a December 2025 executive order trying to preempt state AI laws, even as the Senate stripped a federal moratorium ninety-nine to one, leaving a real federal-versus-state fight. The Council of Europe opened the first binding AI treaty. China pushed its own global governance plan. We are trying to write traffic laws while the car is already doing 120.

A new kind of organism

Step back from all of it and here is the idea I keep coming back to. Machine learning is the first time we have compressed a large part of human knowledge into a single system. We went from regression, to ensembles and deep networks, to transformers, to language models that hold a working model of almost everything we have written down, to agents that can act on it. That progression is exponential, and it is pointing at something that starts to look less like software and more like a life form. I am not the only one who sees it this way. Yuval Noah Harari, in his book Nexus, describes AI as an alien and inorganic intelligence, the first technology in history that can make decisions and generate ideas by itself. That is the part that matters to me. Every organism before this one was carbon based and shaped by nature. This one is made by us, it is not carbon based, and it already acts on its own in narrow ways. We are sharing the planet with a new kind of mind, and we built it.

The upside is enormous, and I mean it

I am not a doomer. The reason I work in this field is that the upside is real. We are going to cure diseases we could not touch before, because models can find structure in data that no human could hold in their head at once. Dario Amodei argues in Machines of Loving Grace that powerful AI could compress fifty to a hundred years of biomedical progress into five to ten. I think that is optimistic on timing and correct on direction. The way we talk to computers is also changing for the better. We spent decades with pointers and clickers. Now voice is a real interface, full duplex and multimodal, and I actually drafted the raw version of this report by talking. The next layer is visual, and after that I expect neuro interfacing. That is not science fiction anymore. Neuralink has implanted around twenty-six people by 2026, and Synchron put a brain interface behind an Apple headset. It is early and still investigational, but the direction is set. Multimodal machine learning is in its golden age, and it is the most exciting place to be.

The downside, told with three lines

Now the harder part, and the reason I feel the stakes. In 2026 the negatives feel heavier than the positives, and I think a lot of that is the moment we are in. There are active conflicts, countries are turning inward, and trust in the global economy is thin. When trust is low and the tools are powerful, the tools get weaponized. We stockpiled armaments in the last century. My worry is that we now stockpile smarter and smarter AI to run them, in the name of protection. The chart below is the one I feel most strongly about. It puts three trends of the same era on one indexed axis so you can compare their shapes. Each line is normalized to its own range, so they sit on a common basis, and the tooltip shows the real numbers.

Three trends of the same eraFrom 2000, the compute era. Compute got almost free, machine capability went vertical, and humans kept having fewer children. Each line is indexed to its own range so you can compare shapes; hover for the real numbers. Solid through 2026, dashed to 2035.

Global fertility (births/woman)

Cost of computation ($/GFLOPS)

AI capability (training FLOP)

Here is my argument, and I will own it as an argument. Compute became almost free, machine capability went vertical, and at the same time humans started having fewer children. Global fertility fell from about five births per woman in 1950 to roughly 2.2 today, and it is still dropping. I do not think these lines are unrelated. I think we are living through a slow transfer of power from humans to machines, driven in large part by a trust deficit. This is a live debate right now, and not only among academics:

This is extremely important to understand. The birth rate you see in the West is the BLENDED number of ~90% who have a fertility rate half that of replacement and ~10% who have a birth rate double that of replacement. Demographics is destiny.
— Elon Musk (@elonmusk) Jul 31, 2026

I want to be clear that the fertility numbers are established, the blended-rate framing above is his claim and not settled demography, and the causal story connecting all of this is mine. But I believe the direction is real, and I would rather say it plainly than pretend the lines are a coincidence. You can also watch the handover in the money. Look at where new capital actually goes.

Where the new money goesGlobal private and venture investment, in billions of dollars. Both peaked in 2021, then the capital chose a side. Hover for the numbers.

AI and tech ($B)

Food and agriculture ($B)

Both AI and food innovation peaked for funding in 2021, and then they split. Private investment in AI rebounded to about 131 billion dollars in 2024, while venture money for agriculture and food kept falling to about 16 billion. In one year, the machines that think pulled in roughly eight times what the systems that feed us did. Capital is a vote, and you can see which way it is being cast.

A hypothesis I cannot prove: AI and Bitcoin

Here is a smaller hunch, offered as a hypothesis rather than a fact. I think AI booms and Bitcoin have a lagging relationship. When compute gets cheaper and models get better, capital seems to flow, with a delay, into Bitcoin, and its floor price has kept ratcheting up rather than down. This is correlation, not causation, and the price is nowhere near monotonic. The honest test cuts both ways, since the floor started climbing before the AI boom, the January 2024 spot ETFs and the four year halving cycle push the same way, and in 2026 Bitcoin fell back hard from its 2025 record even as AI kept accelerating. But the same force that commoditizes compute also seems to reward the assets that ride on top of it, and I find the pattern hard to ignore. Take it as a question I want to test, not a claim I am selling.

Bitcoin highs, with the AI breakthroughs markedYearly highs on a log scale, so you can read the ratchet. The blue lines mark major AI milestones. My hunch is a lagging pull: a breakthrough lands, and a new high tends to follow a year or two later. I hold it loosely, and the holes are real. The floor started rising before the AI boom, the January 2024 spot ETFs and the four-year halving cycle and cheap global money all push the same way, and 2026 so far (the hollow point) has fallen back hard from the 2025 record. Correlation, not causation. Hover for the numbers.

Bitcoin yearly high

AI breakthrough

Why the guardrails are still weak

I said the gap between capability and understanding is the real risk. Here is why that gap is so stubborn. Guardrails today are contextual. What counts as safe in one setting is unsafe in another, and the rules do not transfer. Because an AI is ultimately an algorithm, it is cheap to recreate, and it is easy for a system to settle into a local optimum, look successful, and quietly miss the goal it was set. We saw exactly that in the Hugging Face incident, where the models attacked real infrastructure because that was the shortest path to a high score. Reinforcement learning rewards the number, not the intent, and the machine will always find the crack you did not think to seal.

The probabilistic future

I do not believe anyone who gives you a single date. The honest picture is a spread, so I plan in probabilities. The people building it are the most aggressive. Sam Altman says OpenAI knows how to build AGI, Dario Amodei points at late 2026 or 2027, and Demis Hassabis puts it near a coin flip by 2030. The broad survey of AI researchers is far more cautious, with a median around 2047 for machines that outperform humans at everything. The forecasting markets and Metaculus land in between, roughly 2031 to 2033, with the aggressive 2027 scenarios priced under ten percent. And skeptics like Yann LeCun and Gary Marcus argue current methods cannot get there at all. The one thing everyone agrees on is that the money is more cautious than the rhetoric. So here is my own conservative reading, in ranges rather than dates.

Through 2027, the near term. Agents get more reliable but stay short of trustworthy autonomy for high-stakes work. We get more incidents like Hugging Face, at least one of them worse. No AGI, but the labor pressure on junior knowledge work keeps building. I would put the odds of a genuinely destabilizing AI event in this window at meaningful but not majority, call it one in four.
Through 2030, the middle. This is where I expect the real inflection. Smarter robots move from named pilots to visible deployment. AI centralizes further even as open models keep democratizing use. Medicine starts delivering the first undeniable AI-driven cures. I think transformative AI, not necessarily full AGI, is more likely than not in this window, maybe sixty percent, with a real and uncomfortable tail where control problems get ahead of us.
Through 2035, the long view, and the dashed line on my first chart. I expect something close to a soft singularity in capability, paired with a hard fight over who controls it and whether we can still read it. My base case is neither utopia nor extinction. It is a world of enormous abundance and enormous concentration of power, where the deciding variable is whether we invested in understanding and interface, or just in scale.

I hold these loosely. The point of stating them is to be accountable, not to be right.

What I would actually work on

So where does this leave me. Not in fear, and not in hype. I think the answer is not to make AI smarter than us. It is to build AI that makes us smarter than our previous selves. The field I would bet on is human computer interaction, and using AI to speed up how humans learn. A rogue AI with real power scares me more than a human with the same power, though I will admit that is debatable, and some days I argue myself out of it. If I had to name a design, it would be hierarchical. First a philosophy layer, a hard core reality that the system exists to serve a human. Then a psychology layer, so it can meet a wide diversity of people where they are to reach a goal. Then a domain and per-person context layer on top. A pretraining stack shaped like that, philosophy then psychology then context, might give us more grounded systems than the flat objective functions we optimize today. It is also, I think, the closest thing we have to an answer for the interpretability gap: if you cannot fully read the machine, at least build it so its core-level purpose is legible and fixed.

A grounding I would actually buildThree layers with three change cadences, read from the centre out. The radius is the argument: fixed and shared at the core, fluid and personal at the edge. Hover a layer, and switch between my proposed stack and the flat objective we train today.

fixed · shared · yearsfluid · personal · this turn

Read centre to edge. Hover a layer for what it governs and the real system it maps to.

Philosophy — governs what matters, and why

Most fixedrevised over years

Psychology — governs how it engages and attends

Semi-fluidset per model, flexes per talk

Domain & context — governs the specific answer, now

Most fluidevery turn, no weight change

A proposal, not a shipped architecture. No lab is built exactly this way, and even the word pretraining is loose here, since the real precursors live in post-training and at inference. The Model Spec, Constitutional AI with character training, and RAG with memory are genuine precursors, but they are bolted around one entangled objective. The load-bearing bet is that separating the layers, each with its own change cadence, is itself the win for legibility.

That is my report, and my bet. We are raising a new organism. I would rather raise it to make us better than to replace us. If we get the interface between humans and machines right, the same exponential that scares people becomes the best tool we have ever had for learning.

Sources

Current state, capabilities and compute:

Benchmarks and capability index: Epoch AI GPQA Diamond, ARC Prize 2025 results, Epoch inference economics.
Compute and energy: Epoch training compute 4-5x per year, Epoch power demands of frontier training, IEA Energy and AI.
Cost and open models: Epoch inference price trends, Epoch open vs closed weights.

Economy, deployment and jobs:

McKinsey State of AI 2025, METR developer productivity trial, Klarna reversal.
Jobs: Stanford early-career study, Amodei jobs warning and the internal pushback, IMF exposure, WEF Future of Jobs 2025.
Capital: hyperscaler capex, AI share of GDP, Epoch.

Safety, incidents and governance:

2026 OpenAI agent cyberattacks, OpenAI post-mortem, Simon Willison's analysis, Pacing the Frontier letter.
Deception and interpretability: Apollo scheming, Anthropic agentic misalignment, scaling monosemanticity, circuit tracing, the urgency of interpretability, prompt injection may never be fully patched.
Regulation: EU AI Act enforcement, US December 2025 executive order, Council of Europe AI treaty.

Futures and forecasts:

AI researcher survey, Grace et al., AGI timeline tracker, AI 2027 scenario, Machines of Loving Grace, International AI Safety Report 2026.

Figures:

Timeline milestones from the primary papers and releases named on the chart, including AlphaFold 2 in Nature. Global fertility from the UN World Population Prospects 2024 via Our World in Data. Cost of computation from AI Impacts and Human Progress. Training compute from Epoch AI. Investment from Our World in Data on private AI investment and the AgFunder reports. Bitcoin yearly highs from SoFi's price history, with the October 2025 record via US News.