A new era of AI– Google Gemini ? 

 

Google's 'Gemini' closes the AI ​​gap, but artificial general intelligence remains elusive
Over the once time, generative AI has taken the world by storm.. After his ChatGPT was announced late last year, OpenAI and its partnership with Microsoft gained widespread attention, and the pace of progress has been nothing short of astonishing.
Last week, Google announced its latest generative AI model, Gemini, and raised the price. It was jointly developed by Google Brain and DeepMind and is a direct competitor to OpenAI's GPT-4.
Google touts Gemini as being natively multimodal. This means that the component is not subsequently assembled from different content or data types. Instead, it was built from the ground up across multiple modalities, combining billions of parameters representing text, images, video, audio, and programming code.
The release of Gemini has some wondering if we have finally reached the point where we have created artificial general intelligence (AGI) and technology is smarter than we are.
"Without public API access, it's impossible to determine whether Google achieved his AGI based on his heavily edited PR video," said Wisner of the Paul G. Allen School of Computer Science.・Yejin Choi, Professor Slivka and MacArthur Fellow, said: Senior Director of the University of Washington's College of Engineering and the Allen Institute for AI (AI2).
AI2 is a Seattle-based nonprofit research institute that has been at the forefront of AI research and development since its founding in 2014. Its mission is to serve humanity through impactful AI research and engineering.
“While the advances in GPT and similar technologies have been amazing, we have yet to reach human-level intelligence,” said Oren Etzioni, a professor emeritus at the University of Wisconsin and former CEO of AI2. Masu. “For example, it remains difficult to bring self-driving cars to market.”
Large-scale language models (LLMs) and other forms of generative AI have enabled artificial intelligence to perform an increasing number of tasks with abilities that match and in some cases exceed human cognitive abilities. From the instant creation of business templates to the creation of poetry to the rapid exploration of new approaches to problem-solving, these recent advances are ushering in a whole new relationship with technology.
With Gemini, Google appears to have not only caught up with his OpenAI GPT-4 based ChatGPT, but surpassed it. According to test results published by Google, Gemini outperforms ChatGPT in world knowledge and problem-solving understanding, and also outperforms many human values.
According to Google, Gemini Ultra is the first model to outperform human experts in MMLU (Massive Multitask Language Understanding) with his 90% score. The MMLU covers knowledge in 57 subjects and tests both world knowledge and problem-solving skills. If these test results hold, Gemini will advance and accelerate human knowledge unlike any previous AI.
Still, it would be a mistake to think that this new AI is actually as smart as humans. In fact, it appears that Gemini and other large models still have a lot of work to do. When announcing the release of Gemini, Google created a series of videos to promote the new AI. It also includes a demo video where Gemini quickly and easily responds to questions from a human user off-screen. However, while YouTube's description stated that the video had been edited due to delays, it quickly became clear that these edits were far more extensive than the statement had suggested.
Even behind the scenes, it appears there were far more extensive and detailed requests than what was shown in the video. The media reaction that followed was far from generous. We've never seen exactly how powerful Gemini actually is, but this isn't the kind of launch anyone wants for their product. In retrospect, Google should have been more realistic about the twins' current capabilities.
While all of these advances are exciting, AGI may still be a distant goal. This will no doubt be a question that will be asked regularly over the next few decades as AI advances significantly.
In the meantime, this does not alleviate concerns about AI security. Even if this AI doesn't match his AGI, how can we be sure that these powerful new systems are suitable for business and public use? AI training and security measures
When training and testing Gemini, Google reportedly used his AI2's Real Toxicity Prompts to ensure the output was suitable for these purposes. As I reported on his GeekWire nearly three years ago, Choi's team at AI2 developed and released this set of his 100,000 prompts in early 2021.
Because the toxicity of language is complex and pervasive, we cannot simply filter out vulgar and hateful language, especially when it comes from online content. Real Toxicity Prompts provides a way to train systems to identify and filter out more subtle forms of harmful language and meaning.

This is very important because as these large models become more complex, more effort is required to create safeguards against the results. This is likely to become even more important as developers adopt a multimodal approach. 

 

Gemini is" designed from the ground up to be multimodal, meaning that it can generalize and seamlessly understand, process, and combine nonidentical manners of information, involving textbook, law, audio, images, and video." ” said Demis Hassabis, CEO and co-founder of DeepMind. , I recently wrote on the Google blog.

Bydrawing the power of a large-scale multimodal modeling approach, Gemini provides capabilities that were unimaginable just a few years ago. In recent years, multimodal approaches have been seen as a way to give generative AI important new capabilities due to the additional context that these additional information layers can provide.

Unlike the original large-scale language models that built their output solely from large collections of text, multimodal models gain more meaning from the different forms of the underlying data. In some ways, this is similar to how we ourselves use multiple senses, such as sight and hearing, to better understand a situation. This multimodal approach allows these systems to produce more powerful, sophisticated, and useful results. Gemini is released in three levels.

 Gemini Pro is formerly integrated with Google Bard.

• Gemini Ultra is currently being improved and tested for safety. Launch for developers and enterprise customers is scheduled for early 2024.

• Gemini Nano, a more compact version available on mobile devices, is now part of the Pixel 8 Pro and will be added to more products in the coming months.

Hurdles remaining for AI

After the meteoric release of ChatGPT by OpenAI, it seemed to many that Google was catching up. However, the company takes AI safety concerns seriously and aims to build important safeguards into Gemini, such as reducing the possibility of harmful language, so it is important to keep in mind the development and release of new models. seems to be taking a tardy path.

There are numerous other considerations when it comes to AI safety and ethics. As we saw with ChatGPT, there were many unanticipated use cases, many of which were illegal or harmful.

While it's encouraging to see that Google has reportedly taken the time to put safeguards in place around the new technology, it remains to be seen whether this is enough. Given the complexity of the system and the opacity of the underlying data, we may face entirely new challenges. Moreover, as we all get up to speed on the latest advances and understand what these new systems can and cannot actually do, expect another frenzied PR campaign and breathtaking media hype. It's no exaggeration to say that. Will they destroy jobs or simply change the way we work? Will these models help us better manage the vast amounts of information our world produces? Or will there be an explosion of misinformation, resulting in mistrust? Given the challenges that new technologies pose, it is important to remember that we are building these AIs as tools. They are still far from having self-awareness or anything close to the human motivations that drive our own decisions and actions.

The statistical means by which generative AI makes inferences and produces results is completely different from how human cognition works, and is likely to remain so. Perhaps this is a coincidence. In many ways, this difference is what makes AI so useful to us as a necessary tool for the next stage of progress. Time will tell what role the twins will play in this.

Which  is better? Gemini vs GPT

 Google claims its Gemini Ultra  hardly outpaces GPT- 4 in  utmost  orders  similar as  calculation,  law, and multimodal tasks. For case, it's better than GPT at  calculation by 2.  still, this  exploration lacks comparison with OpenAI's  prideful GPT- 4 Turbo. There are  presently no comparison  explorations with Anthropic’s Claude2.1.   Google states Gemini is the first model to outperform  mortal experts on MMLU( Massive Multitask Language  gathering), which is a test that asks questions in 57 subjects  similar as STEM,  humankinds, and others. In this area, it got a  grievance of 90 vs GPT- 4 at86.4.    still, anecdotal  crashes by  druggies have been  warm to  enunciate the least, citing constant  visions and  restatement  crimes( as well as some questions about the  rally  vids). A clearer picture of Gemini's capabilities will come out over time,  formerly there is been time for independent  exploration to be done.   Gemini is more multimodal than GPT and Claude  In  tours of being multimodal — being  suitable to understand  multitudinous  manners of input — Gemini is  presently ahead of the  rucksack. It can natively take  videotape, images,  textbook, and audio as input. In comparison, GPT- 4 with  unreality( GPT- 4V) accepts image and  textbook, and Claude2.1 only takes  textbook input. Gemini can  produce images, and with access to DALL- E 3, GPT- 4V can as well.   Gemini has a  lower mind, produces significantly  lower affair  Gemini’s commemorative window is significantly  lower than both Claude and GPT- 4 Turbo Gemini has a 32k commemorative capability, GPT- 4 Turbo has a 128k commemorative window, and Anthropic has a massive 200k commemorative window — the  fellow of around 150k words, or 500  runners of  textbook. Commemoratives are  usually an  index of how  important information a model can flash back  and produce.   The  quiescence of Gemini is still unknown  One thing that's a  monumental  procurator with AI models with  candescent new features is  quiescence when GPT- 4 came out, it  handed a lot better  labors than GPT-3.5, but at the cost of celerity.  easily Google is offering three  nonidentical  performances of Gemini to extend  lesser  quiescence options at the  expenditure of capabilities, but how these  mound up against other models has yet to be  discerned. Again, this  exploration will only be a matter of time.   How do I  exercise Google Gemini AI?  Google Bard now uses a fine- tuned  interpretation of Gemini Pro behind the scenes, and it's also accessible on Pixel. Google plans to bring it to Hunt, Advertisements, Chrome, and Duet AI in the coming many months. For  inventors, Gemini Pro will be accessible from December 13 via the Gemini API in Google AI Studio or Google Cloud Vertex AI.   Google has  spoke Android  inventors will soon have access to Gemini Nano via AICore, a new system capability accessible in Android 14. Gemini Ultra is still being fine- tuned and  tried  for  security, with  awaited release early 2024.

   Conclusion

A  major step in multimodal AI input  While Gemini’s on- paper capabilities do not blow GPT- 4’s out of the water — a single  number chance  disparity is n’t actually going to mean  important to someone  utilizing ChatGPT — the multimodal inputs are actually  commodity differently. I anticipate OpenAI and Anthropic will be rushing to  append native  videotape and audio input to their  point channel, if it’s not there  formerly. It'll be  intriguing to  know how these  places stack up in  tours of adding  quiescence to the process.