2024 AGI Outlook: Trends, Predictions, and a Surprise Bonus

1. Virtual Humans and Virtual Worlds

The author observes that 2023’s most successful consumer AI products, such as ChatGPT and Character.ai, introduced the first wave of AI virtual humans. Early excitement gave way to fatigue because of three recurring flaws: chaotic memory, lack of proactive plot driving, and short‑term user engagement.

These flaws stem from large‑model capabilities: imagination and "pleasing ability" develop before logical reasoning. Product stickiness therefore relies on emotional bonding rather than pure technical superiority, allowing a window for incremental upgrades.

One mitigation example is an "external notebook" that summarizes chat history, but it cannot solve the fundamental memory‑forgetting mechanism of LLMs, and fine‑tuning the summarizer is costly.

1.2 Multimodal Explosion

The author frames modalities as follows: text engages the brain, voice engages the heart, and visuals engage the kidneys (i.e., dopamine pathways). Adding audio/video to social/entertainment products yields a qualitative leap.

Since late 2023, companies such as Runway, Pika, Meta, and Google have released video‑generation tools, culminating in OpenAI’s Sora. The author cautions against over‑estimating short‑term maturity (e.g., immediate commercial products) while under‑estimating long‑term impact (e.g., disruptive applications).

Sora’s current limitations include unstable image quality, low speed, and high cost. By comparing to the DALL‑E timeline (product launch in early 2022 → commercial impact after ~1.5 years), the author predicts that 2024 will see AI video technology mature enough for commercial use, with AI‑3D breakthroughs appearing in late 2024 or 2025.

Audio‑driven AI assistants are already mature for many business scenarios; multimodal understanding will enable emotionally expressive dialogue by 2024, making audio even more critical than video for AI‑chat companions.

1.3 Virtual Humans vs. Virtual Worlds

Within a 3‑5 year horizon, virtual humans capable of emulating emotions and independent personalities are likely, but not in 2024. Two core technical gaps remain:

Memory: the need for selective forgetting and context‑triggered recall, which is still a black‑box problem.

Lack of a "human model" and insufficient individual‑level data, meaning current LLMs simulate personalities rather than embody them.

Nevertheless, AI can replace low‑skill visual influencers (e.g., livestream hosts) in 2024‑25, while high‑skill creators retain an advantage.

AR/VR content costs are dropping; Apple’s Vision Pro exemplifies renewed optimism, though large‑scale adoption will likely follow breakthroughs in VR gaming rather than enterprise use.

2. AI‑Native Business Models

AI should not merely be grafted onto existing workflows; true AI‑native companies redesign business models around AI capabilities, akin to how electricity spurred entirely new industries rather than merely powering horse‑drawn carriages.

Current AI‑native examples are scarce: OpenAI/Google/MS official products, Character.ai (top‑10 AI‑chat), domestic "Miao Duck Camera", and seasonal AI‑girlfriend bots.

Future AI‑native directions include:

Universal language translation (text‑to‑text and code‑to‑code).

Enhanced imagination and creativity via virtual humans and worlds.

AI‑to‑AI tool collaboration, forming multi‑agent pipelines for complex tasks.

Massive micro‑decision making (e.g., high‑frequency trading, recommendation engines).

Human‑AI cooperation, requiring a "human model" for personalized interaction.

3. "To AI" Business Opportunities

3.1 Synthetic Data

Synthetic data—AI‑generated data used to train other models—offers a scalable way to augment training sets. Two tiers exist: large‑volume, moderate‑quality data for pre‑training, and high‑quality, domain‑specific data for fine‑tuning.

3.2 Model Market / Platforms

Hugging Face (HF) exemplifies a model marketplace that will become essential when AI agents call each other’s models. The author notes the risk of a closed‑source oligopoly but sees HF as a catalyst for open‑source AGI development.

3.3 Model Engineering Platforms

As data volumes grow, efficiency of training, inference concurrency, and cost become critical. The author lists four focus areas:

Data‑throughput efficiency (e.g., vector databases tuned for LLM access patterns).

Platform stability (reducing long‑running job failures).

Inference cost (expected to rise as user bases expand in 2024).

Inference speed (sub‑second response needed for recommendation, search, advertising, and gaming).

3.4 Firmware & Hardware Co‑Optimization

Joint hardware‑software optimization, especially firmware ecosystems like NVIDIA’s CUDA, presents opportunities for small firms to collaborate with chip giants. Geopolitical tensions may force Chinese companies to develop domestic firmware solutions.

3.5 Model Safety

Models inherit attack surfaces similar to traditional IT systems. New threat vectors include AI‑driven attacks, prompt injection, and the need for AI‑assisted defense mechanisms.

3.6 Privacy

Privacy is framed as a power issue; most users will not pay for privacy, and platforms lack incentives, making pure privacy solutions a pseudo‑problem.

4. Edge AI and 24/7 Hardware

Several smartphone/PC vendors plan to embed small models on‑device, but true offline large‑model capability remains a hype. The most promising edge use case is continuous data collection, exemplified by the "AI Pin" (a chest‑mounted camera + mic) that harvests multimodal data for future model training.

Two possible industry trajectories are outlined:

Centralized large‑model services combined with edge data collectors (Plan‑A).

Personal‑owned models cooperating with central services (Plan‑B).

5. Human Model and Embodied Intelligence

Current LLMs are "world models" trained on tiny slices of many users. A true "human model" would require massive, diverse personal data and would enable AGI to learn from a single individual’s perspective.

Embodied intelligence—AI equipped with sensors such as gyroscopes, pressure, and tactile feedback—is essential for physical robot control. The author expects significant progress in 2024‑25 but does not anticipate fully functional human models within that window.

6. Data Production Balance

The author predicts that AI‑generated data will soon exceed all human‑produced data, ushering an "AI era" where authenticity becomes a scarce commodity.

7. AI Demand: Energy, Compute, Robotics

Energy: Controllable nuclear fusion is the only technology with the potential to increase global energy supply by orders of magnitude. AI‑assisted plasma prediction (Princeton Plasma Physics Laboratory, 2024) demonstrated a 300 ms pre‑disruption forecast.

Compute: Progress continues via 3D stacking, graphene, quantum computing, and high‑temperature superconductors. The author notes that quantum computing remains farther from commercial use than fusion.

Robotics: General‑purpose robots (drones, autonomous vehicles, manipulators) will likely outpace humanoid forms as the primary physical embodiment of AGI.

2024 AGI Opportunity Map

Fine‑grained control of images and ultra‑short video (expressions, precise motions).

Generative short video with style transfer (anime first, real‑person later).

Emotionally expressive AI voice synthesis reaching maturity.

Fully AI‑driven virtual influencers capable of live streaming commerce.

Milestone AI NPCs enabling new game production pipelines.

AI companions with noticeable memory improvements and multimodal interaction.

Real‑time AI‑generated content in social media and advertising.

AI agents delivering solid office‑assistant experiences.

Emerging AI business models: synthetic data, model‑engineering platforms, model safety services.

Wearable, always‑on AI hardware (most will fail, but the market will experiment).

Chinese AI reaching or surpassing GPT‑4; US releases GPT‑5; sovereign AI initiatives appear.

Huawei Ascend ecosystem maturing; domestic inference chips begin to replace imports.

Deep‑fake, fraud, and AI‑driven attacks entering public consciousness, outpacing regulation.