The Artificial Investor - Issue 41: 2025 predictions
My name is Aris Xenofontos and I am an investor at Seaya Ventures. This is the first post of the year, where we will go through our AI predictions for 2025.
Beforehand, let’s evaluate our 2024 AI predictions with 👎/ ⚖️/ 🎯 (far from reality/ more or less/ spot on), so you can tell if this post is worth reading🙂.
If you prefer to listen to this issue, click here for the audio version.
Powered by ElevenLabs
📓 2024 AI prediction assessment
1. The First AI Giant's Fall: 🎯Spot on!
Not one, but three AI scale-ups fell (Inflection, Character AI and Adept).
2. Mobile Intelligence 2.0 with integration of hardware, OS and LLMs: ⚖️More or less
Apple launched Apple Intelligence, but only launched some basic new AI features with iOS 18.2.
3. Supply/demand balance & price stabilisation of AI hardware: 🎯Spot on!
All AI startups we know have gotten cheaper and easier access to GPUs in 2024 than 2023.
4. AI deepfakes cause a dispute in a national election: 👎Far from reality
AI deepfakes rose but did not reach the levels of proliferation required to clearly influence an election.
5. AI as a catalyst for a military victory: ⚖️More or less
Not exactly a military victory, since the Ukraine/Russia war is still ‘live’, however, autonomous drones, face recognition, etc. have made everyone talk about the age of battle asymmetry (Davids’ ability to defend against Goliaths).
6. The birth of truly AI-dedicated chips driving innovation and performance gains: 🎯Spot on!
We saw true AI-dedicated chips launched by startups, such as Groq and Cerebras, leading to lightning-fast model performance.
7. Investment in AI will be lower than the record levels of 2023: 👎Far from reality
Investment in AI in 2024 was higher than in 2023. We haven’t changed the phase of the hype cycle yet.
8. Every Tech giant will launch its own AI chip: 🎯Spot on!
Literally every hyperscaler announced its own AI chip: Meta MTIA, Amazon Trainium, Microsoft Maia, Apple Baltra and Google Axion.
9. Fortune500 hire Chief AI Officers, but only a few enterprises scale AI: ⚖️More or less
Chief AI Officers didn’t really become a thing. But Enterprise AI adoption at scale indeed hasn’t met expectations yet.
10. The first non-Transformer model launches & challenges the status quo: 👎Far from reality
Transformer-based models have certainly demonstrated their weaknesses, which have been addressed by alternative architectures (e.g. state-space models like Mamba), but we are far from changing the status quo.
11. AI dominates Oxford's English language lexicon expansion: 🎯Spot on!
Generative AI, AGI, prompt engineering, training data and hallucinations are all words that have entered the official English dictionary in 2024.
12. The Dark Knight, Apple, is once again a strong second-mover: 🎯Spot on!
The BigTech company launched Apple Intelligence and demonstrated its focus on user experience as opposed to technological edge (for the sake of it). A deal is made with OpenAI to access its models for free (in return for access to its client base).
13. The revolution of senses: the first truly multi-modal AI model is unveiled: 🎯Spot on!
Not one, not two, but three true multimodal AI models were launched in 2024 and wowed everyone. OpenAI 4 Omni, Gemini 2.0 Flash and LlaVa (a vision version of Meta’s LLama).
Wrapping up, we had 54% big hits, 23% partial hits and 23% misses. Not bad. Although, it could also mean that we could have been bolder with our predictions last year.
Now, let’s go into our 2025 AI predictions and try to make them bolder.
🔮 2025 AI predictions
We will structure our 2025 predictions around the biggest AI debates. Here it goes:
Most people will acknowledge we are an AI funding bubble…but, it won’t burst yet.
Value will finally transition from the Infrastructure to the Application layer (within the consumer, prosumer and small-business market).
This will be the year of robots and robotaxis…compensating for the disappointment of agents.
The AI Cold War will get worse before it gets better, but we see a path to a peace treaty.
The homes of most AI innovators, California and the UK, will introduce strict AI laws.
AI models beat all humans in logical reasoning, but 2025 reminds us how far we are from superintelligence.
AI hardware prices increase and availability becomes limited, as the world gets excited about inference-driven models.
(Bonus prediction) 2025’s emerging AI topics: explainability, data curation, solving the “hallucination problem”, building systems around models, and, of course, AI agents.
Let’s get to it.
🫧 Are we in an AI funding bubble? If so, when will it burst?
We think that 2025 will be the year that most people acknowledge there is an AI funding bubble, but it won’t burst this year.
We expect AI funding in 2025 to surpass the 2024 levels, driven by emerging funding sources, such as (the return of the) crossover funds, infrastructure funds, and sovereign wealth funds (as they get increasingly convinced of the strategic importance of AI sovereignty). Also, the interest in AI assets is likely to push the reopening of the IPO window and we will have more AI IPOs in 2025 than in 2024. This will drive AI investor returns, which will generate further interest in AI assets. Nevertheless, most people will accept the fact that large private AI companies, such as OpenAI, are overvalued.
🔮 I would be a fortune teller if: There were two additional AI 'decacorns' in 2025, five AI IPOs, and some OpenAI shareholders sell their shares below the valuation of the last round
📉 Is AI a technological hype? If so, are we going to get over the hype cycle in 2025?
We believe that the GenAI technology has strong value and will scale, but 2025 will be a mixed year of adoption.
Value will finally transition from the Infrastructure to the Application layer within the consumer, prosumer and small-business market. This will be driven by multimodality (voice and vision) and small model quality. We will see multiple new unicorns and a couple of killer apps that gain great traction and make a big buzz.
We don’t expect the same penetration in the Enterprise segment. Output quality/ reliability challenges will limit the adoption within low-risk use cases.
🔮 I would be a fortune teller if: The first GenAI (application-layer-only) application reached 1m users, the first SME GenAI (application-layer-only) software reached $100m ARR, and the percentage of enterprises with GenAI solutions at scale and in production increased by 33%.
🌍 Can AI have a true impact on the world or is it just about summarising text?
We anticipate 2025 to be the year that AI has a true impact in the world thanks to advancements in robotics and autonomous vehicles. With China leading the way, and the US following suit, we believe that we will see robots dominate factories, enter our homes in humanoid form and transform entire military units. On the other hand, despite being in the media's spotlight, agents will be limited to single, uninteresting tasks and will fail to meet expectations overall.
🔮 I would be a fortune teller if: the number of manufacturing robots doubled in 2025, 10,000+ humanoid robots were ordered, there were >3 cities where robotaxis overcome conventional car ride-hailing platforms, and there were multiple articles about how hyperscalers’ AI agents failed to meet expectations.
⚔️ Who is going to win the AI Cold War between the US and China?
An AI Cold War between China and the US broke out in late 2023 and continued throughout 2024. The US is currently on the lead, having stronger AI science, in both software (foundational models) and hardware (GPUs), as well as larger-scale cloud infrastructure. On the other hand, China holds the keys of the AI supply chain, and has strong research talent and IP creation.
We believe China will catch up with the US in AI science in 2025. Also, we see Trump as an experienced businessman and believe his stance towards China is a negotiation tactic. China needs the US as much as the US needs China. Despite the fact that the AI Cold War will get worse before it gets better; it wouldn’t surprise us if there was a path to peace by the end of the year.
🔮 I would be a fortune teller if: Chinese frontier labs launched new foundational models ahead of US competitors, official discussions about a US/China treaty commenced.
📖 In what direction is AI regulation heading?
2024 was the year that the EU was yet again a “pioneer” of Tech regulation in the Western World and introduced the EU AI Act, some US states voted small AI-related laws, whilst the home of AI innovation, California, blocked them, and the UK-based journalists were proud of their country staying away from regulation to allow innovation to flourish.
We expect the new US President to stick to his campaign promise and avoid federal AI regulations. However, we believe that the main AI infrastructure companies will start becoming subject to AI regulation in 2025, as both California (Democrats won both Congress and House) and the UK will introduce new AI laws.
🔮 I would be a fortune teller if: the state of California and the UK introduced strict AI regulation.
🧠 Are we going to achieve AGI in 2025?
2024 was the year where Transformers’ limitations were exposed in terms of logical reasoning, memory, reliability, etc, and evolved, as well as completely new architectures, emerged.
Note: We don’t think it's constructive to have a debate about how far we are from Artificial General Intelligence (AGI) and whether we will achieve it this year or not. If the keyword here is “General”, then current AI is generic (multi-purpose, multi-domain) enough. It makes more sense to talk about superintelligence, i.e. measuring AI against human intelligence and marking the point that it exceeds it.
We believe 2025 will be the year that reminds everyone that we are still far from superintelligence. Nevertheless, the focus of frontier AI labs, both in the US and China, has clearly shifted from language understanding and reuse of the Internet’s knowledge to logical reasoning and advanced intelligence. We expect great achievements here.
🔮 I would be a fortune teller if: Multiple frontier AI models beat all humans in logical reasoning benchmarks. AI journalism shifted from AGI to superintelligence.
🗒️ Side note: A leading indicator here is the shift in the benchmarks used by frontier labs like OpenAI from one focused on language understanding, trivia-like questions and basic arithmetic reasoning (MMLU, GPQA and GSM8K) to benchmarks that measure Math Olympiad-level of mathematical reasoning, end-to-end coding and logical puzzles that humans can’t solve (MMLUPro and AIME, Codeforces and SWE-bench Verified, ARC-AGI and Konwiski prizes).
⚙️ What will happen in the AI hardware market in 2025? What does this mean for Nvidia?
As predicted by the Artificial Investor, 2024 was the year that supply and demand for AI hardware became balanced again, perhaps even with supply exceeding demand, and as a result GPU rental prices fell. We think 2025 will be a year that looks more like 2023, with supply proving insufficient and prices going up again.
After a 2-year race of scaling training data model size, which led to spending a lot of money in AI hardware to pre-train models, OpenAI showed the world a second scaling dimension in late 2024: inference. This means that if you design an AI model appropriately and you let it run a lot of calculations and “think” when the users ask it questions, then it can provide answers to logical problems that are as intelligent as the ones of the best human mathematicians. This makes models very hardware-hungry while they are being used by clients (b). As the world is getting excited about this, we expect demand for AI hardware to soar again, leading to higher prices and reduced availability in 2025. Good news for Nvidia again.
🔮 I would be a fortune teller if: GPU rental prices went back up and near 2023 levels. Nvidia’s share price nearly doubled.
💡 What will be the next “big thing” in AI?
Evidently most people are currently talking about AI agents and this will continue to take front stage for most AI hyperscalers. However, we are more interested in capturing some of the emerging AI topics. We expect AI innovators to increasingly talk about: i) explainability of AI models, ii) data curation for better quality of training data, iii) solving the AI model “hallucination problem”, and iv) building systems around AI models, such as agent integration and Enterprise scaling systems.