The major LLM players, as well as third-party developers using LLM APIs, will get us to genuinely AGI-ish levels during 2024. <\/p>\n
Depending on your definition of AGI, artificial general intelligence. <\/p>\n
Multimodal (visual etc) AI will be part of it, but the NLU (natural language understanding) provided by LLMs will be and always was the key to AGI. <\/p>\n
NLU is not just about content creation. NLU is actually bottled intelligence. When you solve NLU, you get intelligence as part of that. How else can the LLM write sensibly if its not intelligent? Even during 2023 the most interesting use of LLMs was in intelligent decision making, not content generation. <\/p>\n
To be useful, AGI need not be sentient, super intelligent or even necessarily multi-modal. <\/p>\n
Thats why its close. <\/p>\n
Early AGI will be quite simple. <\/p>\n
The key step required by a minimal AGI is simply resilient independence in not entirely controlled environments, whether cyber or real world. <\/p>\n
It neednt pass every test either. Engineering is iterative. Even though the aim of AGI is in principle complete independence, well initially be happy to see AGIs that can solve some of the problems of unconstrained environments. <\/p>\n
AGI does indeed need long and short term memory and goal creation and task facilitation skills. <\/p>\n
None of these capabilities are anything as difficult to add to an LLM as the 70 year grand challenge that LLMs already solved, namely shockingly good NLU. <\/p>\n
The addition of short and long term memory is being attacked from multiple angles: increased prompt sizes and RAG (retrieval augmented generation) for example. <\/p>\n
And LLMs, with the right app flow, are ready for decision making roles important for resilience or for goal creation and task facilitation. <\/p>\n
What well see during 2024 is a plethora of AGI-ish apps operating in the cyber world in text and multimodal form factors. <\/p>\n
And then in bots and androids including Teslas Optimus project from mid-2024. <\/p>\n
Most of these early attempts will be prescriptively designed, iterative ad hoc solutions to mimic human-like behaviour. <\/p>\n
<\/p>\n
Here is the original post:<\/p>\n
AGI predictions for 2024. The major LLM players, as well as | by Paul Pallaghy, PhD | Dec, 2023 - Medium<\/a><\/p>\n","protected":false},"excerpt":{"rendered":" The major LLM players, as well as third-party developers using LLM APIs, will get us to genuinely AGI-ish levels during 2024. Depending on your definition of AGI, artificial general intelligence Continue reading