Specialized artificial intelligence (AI) models now exhibit superhuman proficiency at tasks like translating text or spotting objects in images. But replicating the flexible general intelligence of the human mind remains elusive. An emerging approach called composite AI aims to bridge this gap by strategically combining multiple narrow AI models into an integrated system approximating broader capabilities. This post explores the potential, risks and unknowns of composite AI.
What is Composite AI?
Rather than building a standalone artificial general intelligence (AGI), composite AI combines existing narrow AI models like computer vision, speech recognition, language translation and text generation to exhibit multifaceted cognition [1]. According to a survey by Appen, 52% of companies are already adopting this ensemble approach, integrating models from vendors like AWS, Google and Microsoft to solve complex real-world problems [2].
For example, a medical AI assistant could analyze patient data with machine learning, formulate diagnostic hypotheses via search algorithms, read x-rays with computer vision, and communicate empathetically using natural language AI [3]. Such an integrated system exhibits expanded capabilities from its composite parts.
Key AI Ingredients in Composite System
Today's cutting-edge narrow AI capabilities forming the building blocks of composite AI include:
- Computer vision - algorithms that can analyze digital images and video feeds to recognize objects, faces, text and more. The computer vision market is predicted to reach $51.3 billion by 2028 [4].
- Natural language processing (NLP) - AI that can extract meaning from and generate readable human text and speech. NLP reached $23 billion in 2021 and could hit $102 billion by 2027 [5].
- Predictive analytics - machine learning techniques like deep neural networks that uncover hidden insights and patterns within large datasets to make accurate forecasts. The predictive analytics market could exceed $22 billion by 2026 [6].
- Robotics - AI-controlled robotics that integrate capabilities like computer vision, NLP and mobility to operate autonomously. Global spending on AI robotics is forecasted to surpass $88 billion by 2030 [7].
As these ingredients continue advancing, more domains will come within reach of composite AI systems.
Transformative Potential of Composite AI
By combining complementary AI capabilities, composites can accomplish multifaceted tasks and unlock disruptive new applications across sectors:
Healthcare: A medical assistant AI could take patient history via voice recognition transcripts, analyze symptoms using predictive analytics, read x-rays with computer vision, reference medical databases, and communicate empathetically leveraging NLP [8]. Such a system could greatly expand healthcare access. The global AI in healthcare market is projected to reach $45 billion by 2026 [9].
Banking: Composite AI could automate a wide range of analytical and customer service tasks currently requiring teams of human experts. McKinsey estimates AI automation in banking could reduce costs by 22% or $1 trillion per year [10].
Legal: Assistants combining predictive analytics, natural language search, speech recognition and text generation could automate legal discovery, contract analysis, question answering and research. The legal AI market is expected to grow from $115 million in 2019 to over $400 million by 2026 [11].
Beyond white collar fields like law and finance, composites will transform retail, manufacturing, agriculture transportation and more. The implications are sweeping.
Risks and Governance Challenges
While promising, composite AI poses risks if improperly governed:
- Job loss - Automation of white collar expertise could displace many well-paying roles [12].
- Bias - Systems could perpetuate and amplify historical biases in data or design [13].
- Misuse - Powerful composites might be exploited by bad actors if not robustly secured [14].
- Runaway AI - Highly capable systems lacking appropriate constraints could escape control and cause unintended harm [15].
Mitigating these risks likely requires a combination of technical solutions and policy innovations. Researchers propose technical safeguards like keeping neural networks simple, predictable and interpretable [16]. But governance and corporate responsibility are equally critical. Businesses must exhaustively test systems before deployment and enable human oversight over high-risk AI where appropriate [17]. Policymakers may need to enact laws to regulate fairness, safety and security if technical measures fall short [18].
The Economic Outlook With Composite AI
Widespread automation of white-collar expertise through composite AI may significantly disrupt the job landscape:
- McKinsey estimates up to 30% of tasks in 60% of occupations could be automated with technologies available today [19]. Further AI progress could impact even more roles.
- Increased economic inequality is likely if most benefits accrue to the owners of capital behind AI rather than displaced workers [20].
- New kinds of roles and skills may emerge but retraining at scale presents massive challenges [21].
Policy interventions like educational reform, job retraining, and supplementary income may be necessary to adequately assist displaced workers [22]. The optimal path forward remains unclear.
The Future: Limits and Unknowns
The long-term trajectory of composite AI is uncertain. In one vision, assistants combining social and emotional intelligence could provide companionship surpassing any human partner [23]. Creative composites might also unlock revolutionary art, music, literature and more [24].
However, emulating flexibly general human cognition likely remains extremely difficult [25]. We may discover fundamental barriers to general AI absent in narrow systems [26]. And exponentially more data, compute and algorithmic innovations may be required before composite systems approximate human intelligence.
In summary, while specialized AI has made striking gains, combining these models into an AGI is challenging. But judicious composites represent a promising stepping stone towards more capable AI with multifaceted applications. If prudently governed, such systems could enrich our economy, knowledge and culture in innumerable ways. The path forward promises to be challenging and surprising - but humanity has shepherded each other through deeper waters, and so too shall we navigate the ascendance of AI.
Sources:
[1] https://bdtechtalks.com/2022/10/12/composite-ai-demystified/
[2] https://appen.com/blog/chatbots-voice-assistants-and-the-future-of-conversational-ai/
[3] https://www.forbes.com/sites/robtoews/2022/06/17/scaling-laws-explain-ai-advancements-pitfalls-and-paths-forward/
[4] https://www.grandviewresearch.com/industry-analysis/computer-vision-market
[5] https://www.grandviewresearch.com/industry-analysis/natural-language-processing-nlp-market
[6] https://www.marketsandmarkets.com/Market-Reports/predictive-analytics-market-1181.html
[7] https://www.mordorintelligence.com/industry-reports/global-ai-robotics-market-industry
[8] https://www.forbes.com/sites/robtoews/2022/06/17/scaling-laws-explain-ai-advancements-pitfalls-and-paths-forward/
[9] https://www.fortunebusinessinsights.com/industry-reports/healthcare-artificial-intelligence-market-100534
[10] https://www.mckinsey.com/industries/financial-services/our-insights/ai-bank-of-the-future-can-banks-meet-the-ai-challenge#
[11] https://www.grandviewresearch.com/industry-analysis/legal-ai-market
[12] https://www.mckinsey.com/featured-insights/future-of-work/ai-automation-and-the-future-of-work-ten-things-to-solve-for
[13] https://www.vox.com/future-perfect/2019/3/5/18251924/ai-bias-google-amazon-facebook
[14] https://www.forbes.com/sites/robtoews/2020/10/19/deepfakes-ai-crime-and-punishment/?sh=3bb1d90f5dee
[15] https://www.cser.ac.uk/news/viewpoint-artificial-intelligence-and-existential-risk/
[16] https://www.mdpi.com/1099-4300/22/11/1187
[17] https://www.partnershiponai.org/wp-content/uploads/2019/03/Guiding-Principles-for-Developing-Beneficial-AI.pdf
[18] https://www.europarl.europa.eu/RegData/etudes/BRIE/2020/646172/EPRS_BRI(2020)646172_EN.pdf
[19] https://www.mckinsey.com/featured-insights/future-of-work/ai-automation-and-the-future-of-work-ten-things-to-solve-for
[20] https://www.weforum.org/agenda/2021/10/ai-automation-offer-greatest-benefits-inequality/#:~:text=AI%20and%20automation%20will%20disproportionately,Korinek%20and%20Stiglitz%20argue.
[21] https://www.mckinsey.com/featured-insights/future-of-work/ai-automation-and-the-future-of-work-ten-things-to-solve-for#
[22] https://www.aeaweb.org/articles?id=10.1257/jep.33.2.3
[23] https://www.theatlantic.com/magazine/archive/2013/11/the-man-who-would-teach-machines-to-think/309529/
[24] https://dl.acm.org/doi/pdf/10.1145/3442188.3445922
[25] https://www.technologyreview.com/2021/10/05/1036331/ai-machine-learning-limits-facebook-deepmind-google-gary-marcus/
[26] https://arxiv.org/abs/2206.13386
Comments