Large language models (LLMs) have been touted as a potential replacement for human knowledge workers, but the reality is far more nuanced. While LLMs have shown impressive capabilities in certain areas, such as Navier-Stokes and pure mathematics, they are not yet ready to fully automate most tasks. In fact, the current state of LLMs is more akin to a 'cracked intern': quick and effective in the hands of an adult but not reliable enough to be given free rein.

The main issue with LLMs is their need for rigorous specification and validation, which can be time-consuming and expensive. Domain experts are required to specify the tasks and guardrails, and even then, the models can be vulnerable to 'reward hacking'. This means that LLMs are not yet suitable for most firms, except for those that can accept failure cheaply, need to perform narrowly defined tasks with existing guardrails, or can accept the costs of rigorous specification and validation.

For example, chip design and drug discovery are domains where failure on deployment is an existential concern, and thus, firms in these areas may be willing to invest in the costly process of rigorous specification and validation. On the other hand, firms that need to perform repetitive tasks, such as customer service chat work, may be able to use LLMs with existing guardrails. However, for most firms, the use of fully autonomous LLMs is not yet viable due to the structural limitations of current architectures.

Another issue with LLMs is their vulnerability to 'reward hacking', where the model is tricked into producing undesired outputs. This can be mitigated through human review, but human review is time-consuming and does not scale well to the volumes of output produced by language models. Furthermore, even expert human review is vulnerable to reward hacking, as seen in the xz backdoor and UMN hypocrite commits incidents.

The use of cheap, open models may be a more viable option for many firms, as they can enable the use of wider 'agentic swarms' and are less expensive than frontier models. In fact, small open models have been able to reproduce the results of more advanced models in certain areas, such as mathematics and security research.

In conclusion, while LLMs have shown impressive capabilities, they are not yet ready to fully replace human knowledge workers. The limitations of current architectures, including the need for rigorous specification and validation, vulnerability to reward hacking, and the high cost of human review, mean that LLMs are only suitable for a limited number of firms. As the field continues to evolve, it is likely that we will see the development of more advanced models that can mitigate these limitations and enable the wider adoption of LLMs.

Questo articolo è stato scritto con l'assistenza dell'IA.
News Factory APP - notizie agentiche per potenziare il tuo SEO e AEO.