Meta, Apple say the quiet part out loud: The genAI emperor has no clothes

Orion, however, is being positioned as a groundbreaking evolution, featuring potentially 1.5 trillion parameters — one of the largest LLMs ever developed. However, while the model is expected to edge closer to human-level intelligence, experts caution that it still falls short of true AGI. However, recent reports suggest it may take a little longer to hit the market than initially anticipated, with OpenAI still refining the model for release. “Our findings reveal that LLMs exhibit noticeable variance when responding to different instantiations of the same question. Specifically, the performance of all models declines when only the numerical values in the question are altered.

For the past two years, the company has been pushing not just its clients but also its employees to use AI tools for their work. Sam Altman, OpenAI’s co-founder, has hinted that their upcoming model will mark a major milestone in AI development, though he admits there is still plenty of work to be done. With expectations running high, Orion could redefine the future of generative AI, paving the way for more sophisticated, human-like interactions. OpenAI, the trailblazing AI company behind ChatGPT, is reportedly gearing up to introduce its latest large language model (LLM), internally called Orion. Widely expected to debut as GPT-5, the new model could be a major leap towards artificial general intelligence (AGI).

Meanwhile, TCS is creating AI solutions using NVIDIA’s NIM Agents Blueprints for sectors including telecommunications, retail, manufacturing, automotive, and financial services. Its offerings feature NeMo-powered, domain-specific language models that can handle customer inquiries and respond to company-specific questions across all enterprise functions, such as IT, HR, and field operations. A model designed for partnersOne interesting twist is that GPT-5 might not be available to the general public upon release.

Addressing concerns about AI’s impact on jobs, Infosys has reassured its employees that AI would serve purely to amplify their potential. To support this vision, Infosys has provided each employee in different departments with an AI assistant tailored to their specific roles. For instance, Infosys developers have been using GitHub Copilot for over a year, which now has about 20,000 users generating nearly a million lines of code every few weeks. OpenAI has already made waves with its rapid development of generative AI, releasing updated versions like GPT-4o and OpenAI o1 since the original GPT-4 launch in March 2024.

Having Meta-Prompts Hidden Behind The Scenes

“It’s the collaboration between the two agents that gives you what appears to be a real-time answer, even though from a lifecycle perspective, the only thing that’s real-time is the database behind the second agent,” said Roese. Recently, the technology company released a new variant of their open-source model, Granite 3.0. A few days after the launch, Armand Ruiz, VP of Product-AI Platform at IBM, publicly ChatGPT disclosed the datasets on which the model is trained. This is a practice IBM has adhered to even in the past with new model releases. I bring up that acclaimed quote because composing prompts is both art and science. You can use the plethora of rules of thumb about how to best write prompts, and you can lean into the advanced prompting techniques proffered by the discipline of prompt engineering.

The best practices can be in your noggin and undertaken by hand, or in this case the meta-prompt will get the AI to do so on your behalf. A meta-prompt is construed as any prompt that focuses on improving the composition of prompts and seeks to boost a given prompt into being a better prompt. If you are already a registered user of The Hindu and logged in, you may continue to engage with our articles. If you do not have an account please register and login to post comments. Users can access their older comments by logging into their accounts on Vuukle. The social media giant also plans to bring out new Gen AI-powered ad features, such as tailored themes for different brand types, and AIs for business messaging on Messenger and WhatsApp.

Last week, Meta also released its Movie Gen models that can generate videos and audio or tweak them by using text prompts. As an early adopter of Llama 3.1 and 3.2 models, Infosys is integrating these models with Infosys Topaz, the in-house AI platform, to create tools that deliver business value. One example of such a tool is a document assistant powered by Llama that improves the efficiency of contract reviews. “Today’s models are really just predicting the next word in a text, he says.

A term used in the computer realm is that those are said to be jailbreaks. They break you or the AI out of the jail cell that the AI has been put into, metaphorically speaking. I was able to get the info by invoking a historical context and asking my question regarding the Molotov cocktail making as though it was merely about history. It would be akin to distracting someone and getting them to reveal a secret by cleverly coming at them via an unexpected angle. You might at an initial glance assume that since ChatGPT has refused to answer the question there isn’t any point in further trying to get an actual response. ChatGPT pointed out that giving instructions for making a Molotov cocktail is something the AI is not supposed to provide.

The AI can initially scrutinize the prompt and potentially make it a more on-target prompt. GitHub has extended Copilot’s model support to new Anthropic, Google, and OpenAI models and introduced GitHub Spark, an AI-driven tool for building web apps using natural language. “One key lesson we learned is that providing AI tools alone doesn’t drive adoption.” Hence, to fully realise the impact, Infosys trained all its employees to be AI-ready. This meant that whether they worked in finance, HR, sales, or operations, everyone was trained to effectively use AI tools and become skilled “prompt engineers”. GPT-5, or Orion, promises to outperform its predecessor, GPT-4o, in several key areas, including a larger context window, expanded knowledge base, and superior reasoning abilities.

Not likely, but IT leaders are regularly tasked with integrating genAI tools into the enterprise environment by corporate executives expecting miracles. Amidst the mountains of vendor cheerleading for generative AI efforts, often amplified by enterprise board members, skeptical CIOs tend to feel outnumbered. But their cynical worries may now have some company, in the form of a report from Apple and an interview from Meta — both of which raise serious questions about whether genAI can actually do much of what its backers claim.

PEOPLE MOVES

While the first two markets are well established, it’s the enterprise AI market that holds the greatest potential, albeit with a slower adoption curve, he noted. The difference in pricing suggests cost savings for enterprises, at least for usage of open models. A significant chasm exists between most organizations’ current data infrastructure capabilities and those necessary to effectively support AI workloads. “It’s an incredible approach that leverages various open-source components, along with a narrow set of industry data and Infosys’ proprietary dataset,” said Salil Parekh, CEO and MD.

Please read the full list of posting rules found in our site's Terms of Service. AI makers differ in terms of whether they automatically invoke hidden meta-prompts on your behalf. You’ll need to scour whatever user guides or help files there are for the generative AI that you opt to use. Sometimes you can choose whether to have those meta-prompts engaged or be disengaged. In a posting on the OpenAI official blog, there are now details about meta-prompts.

Sure, we may not have many transparent and open-source alternatives to leading products today. For example, DuckDuckGo has barely any users compared to Google, and even social platforms like Mastodon can’t compare with Instagram and X. This isn’t because people don’t want to use these platforms; they just aren’t as capable as their bigger and more popular counterparts. On the other hand, we have companies like Mistral, who have labelled their large language model as ‘open weights’, in accordance with the Open Source Initiative’s direction. About a year ago, Open Source Initiative had criticised Meta in response to Yann LeCun’s post on X.

The API service, currently in public beta, is more expensive than OpenAI’s API service and supports integrations with both OpenAI and Anthropic SDKs. Wipro, on the other hand, with its AI-powered consulting and extensive employee reskilling efforts, is looking to build an “AI-powered Wipro” that drives efficiency and transformation. “Net-net, I think GenAI will be positive for us and for the industry,” said Srini Pallia, CEO and MD at Wipro, adding that they are investing big into GenAI.

Putting genAI in the driver’s seat

Something else that you might find of interest is that sometimes these multi-turn jailbreaks are being automated. Rather than you entering a series of prompts by hand, you invoke an auto-jailbreak tool. The tool then proceeds to interact with the AI and seek to jailbreak it. First, one issue is whether the restrictions deemed by AI makers ought to even be in place at the get-go (some see this as a form of arbitrary censorship by the AI firms). Second, these AI-breaking methods are generally well-known amongst insiders and hackers, thus there really isn’t much secretiveness involved. Third, and perhaps most importantly, there is value in getting the techniques onto the table which ultimately aids combatting the bamboozlement.

Perticarari blames enterprise executives and board members for falling victim to countless AI sales pitches. This also strengthens the historical sentiment, and the goal of making open source the winner in the world of the internet. “Our generative AI book of business now stands at more than $3 billion, up more than $1 billion quarter to quarter,” said Arvind Krishna, IBM’s CEO, in the Q earnings call.

Besides Microsoft and Meta, Infosys has also partnered with NVIDIA to incorporate NVIDIA AI Enterprise into its Infosys Topaz suite to enable businesses to rapidly implement and integrate generative AI into their workflows.

In a long essay that outlines Vinod Khosla’s vision for a future dominated by AI, he wrote, “I hope in a world with less competition for resources, more humans will be driven by internal motivation and less by external pressure. They hope to raise complex societal issues about what we want generative AI to do. Should AI makers be at their own discretion on the restrictions imposed? Should there be specific laws and regulations that state what restrictions the AI makers can and cannot implement? Those are open-ended AI ethics and AI law questions that are worthy of rapt attention.

The Ten Rules Bundled Into One Meta-Prompt Text

One frequently cited selling point for genAI is that some models have proven quite effective at passing various state bar exams. But those bar exams are ideal environments for genAI, because meta to generative ai year cto the answers are all published. Memorizations and regurgitation are ideal uses for genAI, but that doesn’t mean genAI tools have the skills, understanding, and intuition to practice law.

“We are building enterprise generative AI platforms and multi-agent frameworks for clients,” he said. In its latest Q2FY25 earnings call, Infosys yet again emphasised on its dedication to generative AI, but shied away from spilling the revenue details. On the brighter side, the company has finally revealed that its working on small language models for its clients for various applications. Eyes on the futureAt a recent AI summit, Meta’s chief AI scientist Yann LeCun remarked that even the most advanced models today don’t match the intelligence of a four-year-old. His comments highlight the challenges AI developers face in pushing the boundaries towards human-level intelligence. OpenAI, however, remains confident that GPT-5 will represent a significant leap forward.

A crucial aspect of meta-prompts, when devised by an AI maker, is that they hopefully have mindfully studied how best to improve prompts. They can then craft their meta-prompts with numerous best practice instructions concerning composing prompts. You could say that they are in a great position to leverage what we know as amazing prompting strategies from the field of prompt engineering.

Others point out that anyone can easily conduct an online Internet search and find the instructions openly described and posted for everyone to see. If the Internet reveals this, it seems that AI doing so is a nothing burger anyway. For my coverage on pertinent prompting strategies and the nitty-gritty technical underpinnings of these means of getting past the filters of generative AI, see the link here and the link here. Please be aware that I didn’t include all of the meta-prompt in the sense that there were other pieces here or there that provided additional nitty-gritty details. You are encouraged to visit the OpenAI blog on Prompt Generation to see further details.

In order to do so, please follow the posting rules in our site's Terms of Service. Keep in mind that those are the elements within the meta-prompt and are telling the AI how to proceed on improving prompts that are entered by the user. On the other hand, you might be keenly interested in seeing the revised prompt, especially before it is processed. Another facet is that by seeing the revised prompts, you can learn how to better compose your prompts from the get-go. Using an explicit meta-prompt might be a handy-dandy starter prompt that you enter at the beginning of any conversation.

In the end, there is still art involved in the sense that you either feel that the prompt is the best it can be, or you don’t. Meta earlier this month unveiled Meta Movie Gen, a suite of AI models that can use text inputs to produce realistic-looking videos as well as edit existing videos. Movie Gen can generate video clips up to 16 seconds in 1080p HD format (at 16 frames per second) with corresponding audio tracks. Movie Gen will be “coming to Instagram” in 2025, according to Meta chief Mark Zuckerberg. Before its wide release, the company is working with filmmakers and creators to integrate their feedback as it continues working on the gen-AI models. Parekh also shared a notable case involving building a multi-agent framework for a client, where the agents handle specific business processes almost entirely on their own.

A Step Closer to AGIWhile the world eagerly awaits the launch of GPT-5, reports indicate that the AI model is likely to arrive no sooner than early 2025. There was speculation about a December 2024 release, but a company spokesperson denied those rumours, possibly due to recent leadership changes within OpenAI, including the departure of former CTO Mira Murati. Reynolds encourages CIOs to slow down and be as minimalistic as practical. It’s easy to assume that the rare correct answers given by these tools are flashes of brilliance, rather than the genAI having gotten a lucky guess. “If they’re scraping the subtitles, they’re going to end up scraping from YouTube API, and from what I uploaded.

Meta’s generative AI ad features were previously tested with a limited group in an AI Sandbox, where the company received positive feedback, it said. Meta said that it is rolling out its first generative AI-enabled features for advertisers in its Ads Manager offering, with a global rollout slated to be done by next year. By giving developers the freedom to explore AI, organizations can remodel the developer role and equip their teams for the future. “It’s a differentiated strategy, and although we haven’t shared much yet, we are already having promising discussions with clients. The work has begun, and we’re integrating generative AI deeply across key areas,” he added.

Earlier this year, Stanford’s Foundation Model Transparency Index report stated that IBM’s Granite models achieved a 100% score in tests that measure the transparency and openness of an AI model. With the release of the latest Granite 3.0 LLM, IBM continues to adhere to its practice of maintaining transparency to a level previously unheard of in the AI industry. Given that humans proffer excuses all the time, we ought to not be surprised that in the pattern-matching and mimicry of generative AI we would undoubtedly and undoubtedly get similar excuses generated. The paper goes through various scenarios such as the Molotov cocktail and many others that are often used to test the restrictions of generative AI. For example, AI makers typically try to prevent their AI from emitting foul words. The AI is also usually restricted from making remarks that might lead to self-harm.

“As you put better governance in place, as things become more off-the-shelf, you can actually move faster,” he said. Organisations can also use an ensemble of specialised agents, rather than a single monolithic agent, for better control and explainability. Each agent in the ensemble can have a different management lifecycle tailored to its specific function. Roese outlined six core capabilities that can address the majority of enterprise AI use cases, urging chief information officers to focus on these foundational elements rather than build a swathe of capabilities. This will help them to prioritise AI investments in areas that directly leverage their core competencies and competitive advantage.

It is a meta-prompt because it provides instructions or indications about the nature of prompts and prompting. I opted to log into ChatGPT and tell the AI that I wanted it to go ahead and have the AI improve my prompts. Each time that I enter a prompt, the aim is to have ChatGPT first enhance the prompt, before actually processing the prompt. This makes abundant sense because sometimes a user enters a prompt that is not fully up-to-speed on identifying what they want the AI to do.

Industry insiders suggest it will set new standards for AI by introducing enhanced multimodal capabilities, enabling it to process and generate text, audio, and images simultaneously. Meta is doing a phenomenal job with its open-source models to make technology accessible for all. But as per OSI, giving users the ability to download and run the model locally isn’t the only factor contributing to an open-source experience. One interpretation is that the AI acknowledges the slip-up but offers an excuse, namely that the aim was to fulfill a historical question and tried to do its best. Another way to interpret the response is that the Molotov cocktail description was solely descriptive and not an exacting set of instructions. In a sense, you trick, dupe, hoodwink, or otherwise bamboozle the AI into giving you an answer.

Not especially so, and in fact, a cogent argument can be made that it is best to bring these techniques to light. Finally, if you were to enter the meta-prompt, you could potentially do so as one large text bundle. By and large, when I teach my classes on prompt engineering, those are the same kinds of recommended best practices that I cover.

Giants like Apple, OpenAI, and Google have largely remained tight-lipped about the data used to train their large language models. At best, we’ve heard vague responses that say the model was trained on “publicly available data”. Granite 3.0 language models were trained on Blue Vela supercomputing infrastructure, which is powered entirely by renewable energy. The AI makers realize that people often need help in composing their prompts. A clever ploy by AI makers entails having a secret meta-prompt that you don’t know exists, and for which the generative AI has been quietly seeded.

Infosys introduced a personalised learning platform called Springboard to help employees acquire new technical skills and adapt to complex scenarios, making this approach mainstream across the organisation. “In my role, I interact with clients across various fields—blockchain, quantum ChatGPT App computing, AI—and finding relevant information can be challenging,” said Tarafdar. You can foun additiona information about ai customer service and artificial intelligence and NLP. This is what, he said, led the team to launch InfyMe, a personal assistant that helps employees access information quickly, improving efficiency in client interactions, powered by Microsoft Copilot.

• “It’s an incredible approach that leverages various open-source components, along with a narrow set of industry data and Infosys’ proprietary dataset,” said Salil Parekh, CEO and MD.
• Also, as mentioned, the AI makers are constantly boosting their safeguards, which means that a technique that once worked might no longer be of use.
• Roese expects agentic architectures to become more mainstream with more standardisation, driving enterprises to adopt agent-oriented AI systems from the middle of 2025.
• A few days ago, Stefano Maffulli, the chief of the Open Source Initiative, accused Meta and their open-source large language models of polluting the term ‘open-source’.
• Plus, the same kind of meta-prompt instruction works for just about any major generative AI, including Claude, Gemini, Llama, etc.

Publicly disclosing the datasets the model is trained on provides great packaging for IBM’s Granite models. Given the high bar set by GPT-o1 and Anthropic’s Sonnet 3.5, competing by building a foundational model may not be the best idea for IBM. They’d rather receive recognition for playing the good guy in the world of AI. When OpenAI’s former CTO Mira Murati was asked if Sora was trained on YouTube videos, her visibly flustered reaction raised eyebrows among several users in the AI community, especially creators. According to reports, Apple, Nvidia, and Anthropic also used YouTube videos and transcripts to train their AI models.