4
Fig.2. AsummaryofAGIcompetencies.
tion,andlearningcapabilitiesofhumanintelligence[82].AGI for gathering raw data from the physical environment via the
aims to produce AI with greater generalizability across tasks embedded sensors and transmitting this data to central repos-
and situations. Key capabilities for AGI include: itories or other connected devices. However, the bandwidth
1) Transfer learning: Applying knowledge gained in one limitations and communication in the dynamic environment
domain to new domains [83]. of IoT networks [86] can hinder the timely transfer of data to
2) Multimodal understanding: Integrating and contextual- centralized AGI processing nodes.
izing data from diverse sensors (text, audio, video, etc.) The integration of AGI with IoT devices also magnifies
[75], [83]. challenges surrounding data privacy and security. The dis-
3) Reasoning: Understanding causes and effects to explain tributed nature of IoT devices and the data-hungry nature
events and make predictions [84]. of AGI amplify vulnerabilities, necessitating robust measures
4) Self-supervised learning: Discovering higher-level con- forauthentication,encryption,andaccesscontrol.IoTdevices
ceptsandrepresentationsbyexploringenvironments[82], gather diverse and often sensitive data, which are required by
[78]. AI models for training and inference. However, deploying AI
5) Interactivity: Modern generative AGI models are capa- models on IoT devices that respect data privacy presents a
bleofdynamicandinteractiveengagementswithhumans formidable challenge. Models trained on personal or sensitive
and environmental inputs, bringing us closer to the goal data, e.g., the LLMs fine-tuned with users’ personal behavior
of truly interactive agents [78]. patterns or habits, may inadvertently expose confidential in-
6) Few-shot learning: Learning new skills and apply to formation during inference, raising privacy, ethical, and legal
unseen tasks rapidly from limited examples [85]. issues.
Another challenge lies in ensuring the lifespan and adapt-
AGI will enable the creation of AI agents that can dynami-
abilityofAGIovertime.IoTdevicesoftenfindtheirwayinto
cally process and analyze multifaceted IoT data, understand
infrastructure and devices with extended lifespans. Though
context, and make decisions accordingly. This will unlock
AGI could achieve much better generalizability compared to
morepowerful,flexibleapplicationsratherthancurrentnarrow
the conventional AI application that mainly targets solving
AI solutions.
specific problems, the integration of AGI within IoT devices
necessitates foresight into or adapting to the evolving deploy-
F. Challenges
ment scenarios over time.
IoTdevicesaredestinedtobethemaincarrierforfutureAI
applications.However,theconfluenceofAGIandIoTdevices
II. AIFOUNDATIONFORIOT
is still facing many challenges.
A. IoT-Centric Foundation Model for Achieving AGI
AGI models, such as the LLMs, exhibit immense computa-
tional demands, while IoT devices are often characterized by As mentioned previously, the emergence of IoT has led
theirlimitedcomputingpower(e.g.,computationresourceand to a broad spectrum of applications across various domains,
memory size). This is a natural challenge when accommodat- including power grids, healthcare, and smart cities, among
ing resource-intensive AGI. IoT devices also have a limited others. Despite their diverse forms, these IoT-related appli-
energy budget since most of them are designed to be battery- cations can be mathematically abstracted as a function f(·)
powered. However, intensive computation of AI models could that establishes a mapping between input features x and the
drain the battery quickly. resulting output y. With x representing the measured vitals
Other than executing the computation of AI models locally, for wearables, the output y could be a binary indicator for
the computation can also be offloaded to the centralized data heart attack in smart healthcare. Building on data modality,
center. In this paradigm, IoT devices are mainly responsible the prevalent approaches for IoT data analysis can be broadly