Mar 09 2024, 08:34
Google's 'unreliable' Gemini underestimated LLMs.will Meta, Microsoft do better?
Google Gemini's recent stumble has put the spotlight on Microsoft and Meta. Experts feel the 'bake-fast, break-fast' approach of rushing LLMs out the door to win the 'first mover' title might grab headlines but ends up serving audiences a half-baked dish of errors and biases. Are these complex language models trickier to master than Big Tech initially anticipated?
Google's stars do not seem to be aligned with Gemini, literally. In astrology, Gemini is the third sign of the zodiac, but for Google it's a first mega step to make its presence felt in the still-evolving world of artificial intelligence (AI).
As things turned out, Google is now grappling with critical flaws in Gemini, its flagship LLM (large language model).
Following a series of embarrassing incidents - from racist image generation to labelling the Indian Prime Minister a "fascist" - questions are swirling about the future of the project, and even the leadership of CEO Sundar Pichai.
"We definitely messed up on image generation," admitted a subdued Sergey Brin, Google's co-founder, during a recent Gemini Hackathon at AGI House in San Francisco on March 3. This admission comes hot on the heels of Pichai's February 29 memo where he admits, "We got it all wrong".
When Google, the world's leading tech giant, stumbles with its flagship LLM, it raises a crucial question: Are these complex language models trickier to master than Big Tech initially anticipated?
Understanding language goes far beyond vocabulary and grammar. Humans took thousands of years to develop this skill, and even masters do make mistakes. Expecting Al models to instantly grasp the nuances of context, sarcasm, culture, and intent is an unfair ask. We simply haven't built the algorithms sophisticated enough to capture these subtleties.
Well, OpenAI started working on ChatGPT in 2015, investing years into training it, before launching it in November 2022. But Google, fearing to be left behind, launched Bard within four months in March 2023. Where did it leave scope and time for the model to undergo extensive training?
The pressure to be first is pushing companies to rush LLMs out the door. As Brian Green of Santa Clara University observes, "LLMs are being rushed... due to the accelerating competition." This often means users encounter bugs and companies scramble to fix them on the fly. While this allows for faster bug fixes, "the risk is that their incomplete product makes them look foolish".
Gemini and ChatGPT, trained on massive real-world text data, can inherit their biases and limitations. This translates to a risk of preserving stereotypes, generating offensive language, or even failing to understand certain identities and cultures. While tech companies strive for data curation and diverse representation, achieving true inclusivity remains a continuous struggle.
"For years, Big Tech, following what I half-jokingly term the Silicon Valley Code of Ethics - move fast, break things, apologise later-has prioritised speed over everything. This approach, albeit effective for less impactful applications like email or social media, becomes problematic with Al's broader capabilities," says Jibu Elias, an Al ethicist and the chief architect and research and content head of INDIAai - The national AI portal of the Indian
government.
Further complicating matters, training data itself can be biased, incomplete, or contradictory. LLMS need to filter and analyse this data effectively to avoid perpetuating biases or generating factual errors. However, current models lack transparency in their reasoning, raising concerns about hidden biases and potential errors.
Users want accurate, respectful, and diverse outputs tailored to their specific needs. But the complexity of this task, coupled with the fierce competition, might be leading tech giants to underestimate the R&D and long-term commitment needed to truly crack the LLM code.
"The rapid advancement of Al in India faces challenges not only of multilingualism but also of fostering a genuine culture of innovation," says Elias, adding that, the Indian tech landscape, known for its service-centric approach, frequently trails behind global peers in pioneering breakthroughs.
The allure of diverse user markets like India is evident in the Big Tech's rush to incorporat
local languages into their Al models. To name a few-Microsoft's Jugalbandi project with A14Bharat and the India Development Center (IDC), Google's Vaani project with the
Bengaluru-based Al and Robotics Technology Park (ARTPARK) and the Indian Institute of Science (ISC) and most recently Meta joining forces with Sarvam Al a generative AL.
source: et
Mar 09 2024, 09:04
- Whatsapp
- Facebook
- Linkedin
- Google Plus
0