For the general AI user of today, he is fine if he is able to do his mundane tasks without any effort he becomes super happy with AI. He will latch onto any LLM model which has an acceptable brand and people around him are using it. It has become a search of a different kind. An AI tool bringing in a marked difference in your work, qualitatively and quantitatively, is still not in the vast majority’s mind. An AI assistant which can do wonders with your knowledge, expertise and acumen and you being able to guide the assistant endlessly in newer and innovative tasks, out of your ordinary work, will take a long time to settle down in us.
How many of us really need AI assistants for the purposes they are being created in another question, which no one is ready to answer. How many of us know the differentiators of different models? How many of us need multimodal capabilities? How many of us have started doing their 8 works of work everyday, in a day? If that is not the case, the enterprise level integration is also on the minimalist side. While mediocrity is the harsh reality of today’s existence, the product givers / researchers / developers / creators need to show us the exact differences to prove that their effort has been worthwhile and the product is better than similar products, though even by a whisker.
In this backdrop is Meta’s most recent claim. All of us are aware of the issues the LLM models sidestep, as they are trained to do so. These are called guardrails. What are the guardrails, and how the model trained for that purpose and in which manner and to what precision, no one is really able to provide an answer. Meta’s recent claim is that its latest models answer more ‘contentious’ questions than the last version. How commendable or worthwhile, no one knows? Topics can be political, and more often than not, these would be the topics, which one ought to know.
It is being claimed that the latest family of Meta AI models. “can wade into more contentious territory than its predecessor,” which given the current situation of this process, does not mean much. It might just be a crude beginning. LIama 4 is also “dramatically more balanced” in the prompts it refuses. There is no denying the fact that AI models have for long struggled with bias, with Musk labeling ChatGPT as “woke.” Too much dodging the prompt might lead to missing out important context and for sure leaves the user annoyed. LIama 4 is less likely to dodge hot-button questions. LIama refused to answer 7% of politically or socially charged prompts, LIama 4 turns down less than 2%.
LLMs HAVE BECOME LIKE ANY OTHER BUSINESS PRODUCT WHICH NEED TO BE MARKETED.
Sanjay Sahay
Have a nice evening.