6 AI Realist Predictions for the Future
This article has been 100% written by a human. The target audience is myself in 2–5 years.
Foreward: An homage to specificity
Some things are easy to predict. I can reasonably guess whether it will rain this week, whether a restaurant will be good, or whether the federal budget will grow next year. But these would be bad predictions- not because they’re incorrect- but because they’re uninteresting (far more damning). Many poor predictions follow this pattern:
- Tomorrow will look like today.
- Startup X will fail.
- Event Y will happen eventually.
These are bad predictions because they reveal nothing. In the short term, tomorrow is mostly like today. Most startups do fail. And “eventually” never gets any closer. They’re not falsifiable, and anyone who makes these predictions will be technically correct- the absolute worst type of correctness.
For these reasons, the predictions below will be (a) specific (b) definite-term, and (c) predict how the future will be different from the present.
A response to Leopold Aschenbrenner
A few months ago in June 2024, a white paper called “Situational Awareness” appeared. The author, Leopold Aschenbrenner, is a former super-alignment researcher from OpenAI. Situation Awareness extrapolated current trends in compute scale-up, algorithmic efficiency, and other AI “unhobbling” to arrive at the following conclusions:
- AI scaling will hold. From 2019 to 2024 we saw 6–8 orders-of-magnitude (OOMs) in model progress from GPT-2 to GPT-4. Improvement of ~1.5 OOM per year will continue.
- Superintelligence (ASI) will emerge by 2027. This is presented as inevitable so long as effective compute continues to grow at its current pace and the above scaling laws hold.
- Big Tech (Meta, Google, et. al) will generate $100B in revenues from AI by mid-2026. AI will shortly become the largest revenue category for US corporations, and these companies will spend $1T on AI infrastructure by 2027.
Situational Awareness filled me with skepticism. It was the same skepticism that I felt after reading Human Compatible or Superintelligence or countless Yudkowsky articles. I have been attempting to understand AI risk for years, and despite my best efforts to “FOOM-pill” myself, I’m doubtful humans will ever be subordinate to ASI.
Two groups of people extol the existential risk posed by ASI. The first are genuine effective altruists who internalize the expected cost of any extinction risk. The other is those who benefit from the massive marketing hype: top labs and the surrounding AI safety complex. With billions in funding and newfound political influence, the incentives to inflate capabilities are strong. I present an alternative view: realistic, medium-term predictions.
My alternate, “AI realist” predictions
Prediction 1: Human-level agents will not emerge and US GDP growth will not exceed 5% by 2027.
The history of AI can be told as a series of benchmarks. A test is established that would herald in “AGI” if beaten. The benchmark is then surpassed and quickly explained away as a poor benchmark, and the goalposts are reset. The list is long (Chess, Go, image recognition, self-driving, the Turing Test, etc). These tests have typically been narrow. Modern definitions of AGI normally mandate some domain breadth, plus the ability to learn new tasks with few examples. We no longer find League of Legends AIs impressive- they need to quickly master multiple games, just like skilled humans.
There are of course those who claim AGI is already here. Let’s assume that an AI can ace any intellectual human test (GPT-4 can already pass the bar exam in the top 10% of all test-takers). I argue that this does not constitute AGI.
How then will we know when AGI has arrived? The rapid self-improvement from a true AGI will likely break through every pre-existing test. All known benchmarks would quickly saturate, and it would become extremely challenging to design new ones. AI systems would also exceed humans on these tests for any time budget (whereas humans outperform current systems when given unlimited time). Beyond benchmarks, inter-agent communication and economic activity would quickly grow, perhaps surpassing human activity. Agents would collaborate, trade, compete, negotiate, hire humans- perhaps even create art. Scientific progress would accelerate, and humanity would be ushered into a new era of either abundance or subordination, depending on your proclivity.
The economic impact preceding either of those outcomes would be measurable in GDP. United States GDP growth has been hovering around 3% for several decades. Whether or not AGI will suppress white-collar wages, the net increase would be substantial. In writing about AI risk and reward, Stanford economist Chad Jones relies on population as the primary variable affecting GDP. At scale, AGI would be equivalent to adding millions or billions of silicon-based, graduate-level workers into the economy. By that calculation, US GDP could climb into double-digits. Goldman Sachs more conservatively estimates 0.4%. I find Tyler Cowen’s prediction most likely: the increase will be approximately 1%. This will be unnoticeable year-to-year, but significant when compounded over several decades.
Prediction 2: AI will not recursively self-improve by 2030.
While Aschenbrenner doesn’t explicitly mention “foom” or “singularity”, he echoes many of his safety contemporaries in arguing that exact point. The thinking is that once AI reaches researcher-level intelligence, it will explosively accelerate. Ever-smarter intelligence will be able to generate cleverer algorithms and better exploit available resources in a positive feedback loop.
Sufficient intelligence is the remaining limit for recursive self-improvement. We already possess AI + human self-improving algorithms (HITL) and the scaffolding for purely AI self-improvement is being built.
However, current AIs have not demonstrated a meaningful ability to reason outside existing training distributions. AI alone will not conduct experiments, generate breakthroughs, or advance the frontier of science by the end of the decade. AI will accelerate human scientists by eliminating grunt work (hugely valuable in itself), however, the highest value science will be accomplished by humans for the foreseeable future. This is important because high-value science — breakthroughs in materials, chips, algorithms, etc. — is needed to create AGI. Unbounded recursive self-improvement will only occur when AI can regularly generate, leverage, and compound scientific discoveries.
Prediction 3: AI will be cited in Fields Medal discoveries by 2030 and will help solve a millennium problem by 2035.
In 1792, a teenage Friedrich Gauss carried a notebook where he would jot down mathematical ideas. While looking at prime tables, he noticed that the frequency of primes less than a number n traced the curve n/log(n). This is a remarkable discovery for many reasons: curve-fitting hadn’t been invented yet, the prime tables he referenced contained numerous errors, and this discovery — that of the prime number theorem— would later build into the Riemann Hypothesis. By every definition, it was genius: a better abstraction from higher dimensional, out-of-distribution thinking.
Current AI agents are unlikely to solve a Millenium Problem or win a Fields Medal. However, deep learning will be used by scientists in all fields. Of those fields, pure math and mathematical physics are considered the most difficult, abstract, and IQ-demanding. It therefore stands to reason that as AI climbs the intelligence ladder it will impact these fields last.
Indeed, the accomplishments of deep learning in physics so far are modest: using CNNs to identify Calabi-Yau manifolds; uncovering new statistics in elliptic curves; and generating new conjectures in Knot Theory, to name a few.
However, I find this progress promising. Unlike the agentic approach, AI which is grounded and guided by human intuition gives researchers an entirely new set of tools to tackle big problems. This is especially true for top-down mathematics or problems with significant data size. In the previously mentioned Calabi-Yau example, there are ~500 million manifolds discovered so far, and deep learning is vastly more efficient at identifying candidates than existing computational geometry approaches. In mathematics, breakthroughs often come from applying a tool or technique from one domain to another. AI is one such new vector of attack.
Although I am less confident predicting which millennium problem will yield next, my guess would be the Birch Swinnerton-Dyer conjecture.
Prediction 4: AI Safety will converge around disinformation.
Like nuclear weapons, AI provides enormous leverage to bad actors to wreak havoc. The conversation around AI safety has focused on ASI- super-intelligent machines capable of escaping any network. The more likely (even inevitable) scenario is not a few super-intelligent machines that hack critical infrastructure, but millions of 100 IQ bots.
This has already begun. LLM slop has wrecked online search. Product marketing via thoughtful, human content is dead. As the ratio of AI to human-generated content grows, the state of truth-seeking worsens. In 2025, Google will readily hallucinate details or regurgitate ChatGPT outputs as fact, and these falsities are only reinforced by further training on generated outputs. Bad actors can further interfere with poisoning attacks.
Millions will be duped by fake digital personas. This will swing elections, create and destroy reputations, and advance false narratives. Hidden influence networks will emerge with none of the safeguards of existing ad networks and propaganda will be distributed to billions of people.
Beyond this, real people will be impersonated: presidential candidates engaged in salacious acts; a podcaster advertising fake products; and a crypto scheme promoted by dozens of top technologists. The resulting deluge will saturate our feeds and mar our ability to distinguish authentic media. Without solutions, the internet will become a swamp. Manipulation at scale will become a defining problem of our time. Proposed solutions are in their infancy.
There are of course other aspects to AI safety. Self-driving cars, for example, present questions about what safety threshold is acceptable or whether we require observability. These niche safety debates, once at the forefront of the field, will diminish as the technology is adopted and the risk is internalized. Just as we no longer consider autocomplete, recommendation systems, or machine translation to be advanced AI, the current fight around self-driving cars ensures that the rollout of subsequent autonomous systems will be uncontroversial.
Prediction 5: LLM progress will dramatically plateau, but vertical AI will still emerge as the next great software market.
First, there was the model. Then there was the chatbot (model + RLHF + context). Now we have workflows (static DAGs), agents, and reasoning (inference-time scaling). Chain-of-thought or process models are progressing by using LLMs for higher-order planning and relying on deterministic models, generated code, or external sources like Wolframe Alpha for number crunching and lookup.
It’s increasingly clear that the current plateau in scaling is the result of exhausting the ~1T of text tokens on the internet. Synthetic data holds promise in expanding this volume of data by several OOMs, however, it is only applicable in domains where you can formally verify outputs. We can therefore expect code-gen to improve and math benchmarks to continue to saturate, but synthetic data seems unlikely to unlock true AGI. After ~10 years of pre-training, humans are quite sample-efficient and don’t require trillions of tokens to self-improve.
Even with slowing model progress, vertical AI (Cursor for X) will nonetheless emerge as a massive new SaaS category. Just as ERP, CRM, and PM software are all well-defined categories, “Copilot for X” will become a must-have feature for knowledge workers this decade. Industries will design and adopt this software at their own pace with a vertical-specific emphasis on data security, workflows, and integrations. Just as horizontal, general-purpose software emerged at the end of the SaaS wave, the long tail of human capability and intuition will be hard-fought over many years.
Even if the IQ of models doesn’t improve rapidly, there is still a decade of applied AI work as this technology disseminates. Search, prompt engineering, retrieval/augmentation, interpretability, personalization, user experience, and offline/local models are all areas with large room for improvement.
Prediction 6: The cost to create a photorealistic video will approach zero by 2030.
There are approximately 100 exabytes of video on the internet, about 1000x greater than the volume of text data. We don’t possess the methods or compute to leverage even a fraction of this, yet we already have cheap, open-source, photorealistic models. Future models will have coherent internal representations of the world, be highly interpretable, and give far more precise control over what pixels should change in what manner. By 2030, anyone will be able to generate or edit an image or video of anyone or anything imaginable for approximately zero dollars.
This is already possible, but there are constraints. Image models are biased and overfit, preferring attractive celebrity faces and dramatic lighting (a consequence of how data is selected and voted on for RLHF). Video models are limited in their number of frames, and do not accurately simulate physics.
However, unlike with AGI, the path to overcoming these challenges is clear. SoTA image models require tens of millions of dollars to train and there are dozens of smaller players rightfully competing. SoTA LLMs require billions of dollars to train and there are less than ten companies worldwide who have the resources to train a leading model. If Google were to turn its attention toward image space and spend a few billion dollars training on YouTube data, it would dominate with relative ease. The release of Veo-2 suggests that Google may be starting to overcome its hesitation around liability (IP, abuse, impersonation, etc). ByteDance, through TikTok, already has several exabytes of video and Alibaba has demonstrated significant video prowess. Pressure from startups like Runway, Luma, Pika, OpenAI, and others will not relent.
Closing thoughts
I want to give credit to Leopold. Unlike his former boss and least-specific man of all time, the predictions in Situational Awareness will either happen or be disproved by the end of the decade.
Readers may notice that I didn’t dispute his 3rd prediction- that Big Tech will generate $100B in revenues from AI by mid-2026. On this particular forecast, I agree. In fact, depending on how you measure “AI revenues”, Aschenbrenner may be underestimating. Microsoft is currently claiming a $10B annual run rate from AI, and I suspect Google has both incentives and justification for attributing search revenue to AI initiatives.
Broadly, that is my prediction for the medium-term impacts of AI. This massive newfound leverage currently belongs to the handful of companies with the balance sheets to support billions in R&D and the technical ability across chips, data centers, model development, product, and distribution. That is a fairly short list of incumbents with extraordinarily deep moats. It follows that AI is less likely to upend the existing technocratic world order than it is to enforce it. Humanity is certainly at an inflection point, but the future doesn’t belong to the machines.
