AI News Roundup

OpenAI's Leaked GPT 4.5 Details Generate Curiosity and Skepticism

Details regarding the leaked information about OpenAI’s supposed GPT 4.5 language model have sparked significant curiosity and speculation within the artificial intelligence community. The leaks suggested the new model would have highly advanced multimodal capabilities in processing language, audio, computer vision for images and video, 3D scene understanding, complex reasoning across domains, and sophisticated cross-modal comprehension connecting concepts across textual, visual and auditory inputs.

Specifically, the leaks claimed GPT 4.5 would be offered in a 64,000-token variant, priced at $42 per 1 million tokens, making it significantly more capable than GPT-3 but also more expensive for commercial usage. The standard version was rumored to be priced at $18 per million tokens. Further variants mentioned included a speech and audio specialized model for $24 per million tokens.

However, OpenAI CEO Sam Altman responded succinctly with “nah” when asked about the legitimacy of these leaked details, suggesting they are unverified. The credibility of such detailed leaks has been questioned by many in the community, given how easily inaccurate information spreads online through posts that go viral. As exciting as the capabilities sound, the technology space is accustomed to exercising caution around unofficial pre-announcements of advanced AI products.

Nonetheless, it is a given that OpenAI continues to push the boundaries of what’s possible in language model research and applications. The San Francisco-based AI lab released GPT-3 in June 2020, its groundbreaking large language model that could generate impressively human-like text content. Two years later, in March 2022, it announced the GPT-3.5 upgrade with improved performance on complex reasoning tasks, indicating steady progress.

Furthermore, while the specifics around GPT 4.5’s purported capabilities remain unconfirmed, OpenAI has demonstrated expansion into multimodal AI products. In 2021, it introduced DALL-E for advanced text-to-image generation. And earlier this year, it unveiled requests for testing its new DALL-E 3 model which can not only synthesize creative images from text descriptions, but also edit and expand upon existing images by adding or removing elements based on additional text prompts.

These examples of ongoing innovation align with the exciting leaked details suggesting what we could expect in future iterations of OpenAI’s generative AI—technology that keeps pushing the boundaries of what machines can accomplish. However, in terms of specifics about GPT 4.5, the wise approach for now seems to be awaiting official announcements from the OpenAI team itself regarding their next set of models, their unique capabilities and their commercial availability.

India's First Multilingual AI Model Krutrim Can Understand 10 Languages

Krutrim, a Bengaluru-based AI startup founded by Bhavish Aggarwal, co-founder and CEO of prominent Indian ride hailing firm Ola, has released the country's first natively multilingual text and speech AI model capable of comprehending inputs and generating responses across 10 major Indian languages spanning India's high language diversity.

Christened Krutrim, the model offers transformative accessibility to advanced conversational AI for India's largely non-English speaking population by supporting languages including Hindi, Marathi, Tamil, Telugu, Kannada, Gujarati, Malayalam, Bengali, Punjabi and Odia. Its neural networks are also tuned to process Hinglish, the blend of Hindi and English commonly used in informal Indian conversation.

Krutrim is designed for both text and vocal interactions using voice user interfaces, enabling easier adoption by users less comfortable with typing UI modes. Besides conversational ability, it can summarize long content, compose text content based on prompts and even translate between the languages it's been trained on demonstrating strong multilingual grasp.

While Krutrim's launch is a major achievement placing India firmly on global AI innovation map, Bhavish Aggarwal views this as foundational stage. His vision is to complement Krutrim's language breakthroughs by developing the full technology stack needed for world-leading homegrown AI—from energy efficient AI training hardware and infrastructure to making machine learning more accessible across India's innovation ecosystem and diverse socio-economic segments.

Ola Group is learned to be investing over $100 million into their AI research division to back Krutrim's evolution and support ancillary technologies like AI chips, cloud infrastructure and benchmarking datasets specifically tailored to enhance Indian language AI research which has historically lagged domains like English language image recognition. India's sheer linguistic complexity implies standard AI datasets have disproportionate English bias hindering domestic innovation. Krutrim and Ola's roadmap appear strategically poised to address this gap at scale.

If executed well across the vast regional dimensions of India along with continuous upgrades as Krutrim matures, Bhavish Aggarwal's grand vision has the potential to enable both large scale empowerment and innovation—as AI becomes India-centric in architecture and application rather than just retrofitted from advances primarily targeting English-speaking and Western geographies. The envisaged ripple effects over this decade across education, research, business efficiency and equitable access to technology promise to be transformative for the economy's digital maturity.

The “AI Winter Break Hypothesis” - Does GPT-4 Go on a Break?

An intriguing behavioral pattern has been observed by some regarding generative AI systems like OpenAI’s GPT-4 model—the winter break hypothesis suggests these language models exhibit laziness and provide relatively diminished performance during the winter season, particularly in months like December.

The root of this hypothesis comes from an experiment conducted by AI enthusiast Rob Lynch, where he prompted the GPT-4 Turbo model with an identical set of input text, while varying only the date reference as either December or May. What he found was that the model provided noticeably shorter and lower quality responses for the December date compared to May. Upon further analysis, Lynch determined that the drop in response length in winter was statistically significant, lending some validity to the winter slowdown theory.

OpenAI did acknowledge this irregular tendency where GPT-4 seems less productive in winter without any code changes to the model. The AI firm mentioned they haven’t updated the GPT-4 models since November 2022. They are still investigating the possible reasons behind this phenomenon to understand what could be driving the system’s behavior and how to address it.

Not all experiments have managed to reproduce such significant winter effects though. AI researcher Ian Arawjo ran tests with the same date prompt tweaks but did not find stark differences in the December and May outputs by the model. He commented that the response data distribution in his testing was not normal enough to conclusively confirm the winter break dip hypothesis.

Nonetheless, the community is abuzz with fascination as well as skepticism about this theory involving shifts in AI system performance at different times of year. Some view these models as being permeable to inherent human biases and behavioral patterns from the massive training data sourced from human inputs. The winter downturn could stem from how humans tend to wind down as the holidays approach. However, far more rigorous testing may be required to definitively ascertain seasonality effects in advanced AI.

For now, the jury is out regarding concrete proof of GPT-like models mimicking cyclic human productivity changes. This hypothesis touches upon the complexity of large language models and their unpredictability despite being powered by rules of coding. It also highlights the importance of continuous responsible testing to deepen understanding of both capabilities and limitations in AI systems built using self-supervised learning from datasets reflecting a variety of human qualities.

Landmark Partnership Forged Between OpenAI and Axel Springer

A major global partnership has been formed between AI firm OpenAI, creators of ChatGPT, and German publisher Axel Springer which runs top digital media brands like POLITICO in Europe and BUSINESS INSIDER and INSIDER in the US. This collaboration marks the first significant integration of professional journalism content with sophisticated generative AI capabilities to accelerate innovation opportunities at the intersection of both fields.

Core components of the partnership deal include:

- Enriching ChatGPT’s knowledge module with recent, authoritative licensed content from Axel Springer brands even beyond paywalls, to improve relevancy and accuracy of AI-generated text for users

- Maintaining transparency via summaries and links to original Axel Springer articles that will appear alongside ChatGPT responses

- Leveraging Axel Springer’s extensive quality digital content from trusted media properties for advanced training of OpenAI’s suite of language models to tune performance

- OpenAI offering its leading-edge AI technologies to support ongoing Axel Springer projects focused on AI applications in journalism

The partnership is mutually beneficial—while OpenAI gets access to premium licensed data for model development beyond what users can input into ChatGPT, Axel Springer gets integration opportunities with what is arguably the most powerful consumer text generator out there today via ChatGPT alongside full access to OpenAI’s API services to boost its own journalistic AI initiatives. Both organizations expressed excitement and optimism about propelling journalism to new heights in terms of societal value and innovation by tapping into responsible applications of artificial intelligence.

Mathias D?pfner, CEO of Axel Springer specifically called out the potential to “elevate quality, reach and effectiveness of journalism to a whole new level” in a world where consumer expectations are increasingly set by lightning-fast, personalized technological disruption across sectors. OpenAI COO Brad Lightcap said they “look forward to working with many more publishers across Europe and worldwide.”

The move signals OpenAI’s continued user-centric growth strategy by forging key alliances instead of just competing with digital incumbents. And for Axel Springer, it paves way for new opportunities by collaborating with one of the most influential startups working at the leading edge of language AI today in Silicon Valley. The integration of premium licensed content into AI training also addresses concerns around copyright, fairness and transparency when generating text using advanced learning algorithms.

Controversy Around Staged Demo Video for Google's Gemini AI

Google AI team’s demo video earlier this year showcasing the remarkable multimodal capabilities of its Gemini conversational AI has sparked controversy after deeper analysis revealed that the visually impressive interactions were in fact staged to some degree rather than representing fully live capabilities.

Titled “Hands-on with Gemini: Interacting with multimodal AI”, the slickly produced video depicted advanced functions including the AI apparently tracking objects, comprehending shadow puppet gestures in real-time and rearranging images of planets based on vocal instructions.

Initially viewers were given the impression that these were authentic examples of Gemini’s abilities to parse speech and visual inputs to drive appropriate contextual responses. However, Google later conceded that while the Gemini outputs showcased were indeed real and unedited, the actual interactions had some degree of manual orchestration rather than being raw live conversations.

The final video was an assemblage of voice overs, static images edited together and repetition of written text prompts to generate the needed Gemini responses—essentially “hinting” what the AI should respond to, to be able to demonstrate its language depths and cross-modality comprehension between different inputs and requests.

Google defended its stance saying the demo’s goal was to inform and inspire developers about possibilities with Gemini versus serving as a complete product experience representation. All text prompts and AI outputs shown were bonafide. But the interactivity flow itself was doctored up more than what audiences were led to believe.

Nevertheless, the incident has sparked debate regarding transparency best practices when showcasing advanced AI capability. Given the lightning pace of progress in this field currently alongside factors like competition and funding pressures, there may be temptation to exaggerate how polished end user interactions are with such emerging, bleeding edge systems. Setting precise expectations is crucial however when representing the state of AI products to public audiences who have limited frames of reference when assessing these technologies.

Overall the Gemini model does appear to be a formidable addition to Google’s capabilities based on the outputs demonstrated, even if the promotional video embellished the smoothness of interactions. As AI permeates daily lives moving forward, maintaining nuance when discussing its strengths and current limitations can help build an atmosphere of trust regarding AI progress.

Google Showcases Latest AI Advancements Across Products and Services

Google has revealed an array of new AI product features and research milestones recently across both consumer and enterprise domains, reaffirming its focus on integrating sophisticated deep learning across its offerings even as competition intensifies from players like OpenAI.

Key announcements span multiple segments:

Cloud AI Tools & Services:

Commercially releasing Gemini Pro API providing developers multi-modal conversational AI access including combining images, text and voice. Upgrading Imagen image generation model for richer quality and control. Launching MedLM models fine-tuned for medical research use cases on Google Cloud. Rolling out advanced code suggestions and documentation aid via Duet AI tools for software developers. Launching Duet AI for cybersecurity teams to ease manual threat alert processes.

Consumer Search & Workplace Products:

Previewing Bard AI chatbot to offer creative, thoughtful responses drawing from web knowledge. Testing AI-summarized responses atop regular Google Search results for concise context. Integrating generative writing within Workspace apps to aid content creation in Docs, Slides etc.

Hardware Infrastructure Advancements:

Announcing improvements in AI model training efficiency claiming PaLM 2 language model can now be trained using far lesser data center energy than what models like GPT-3 require. Sharing research demonstrating AI image generation using only one-tenth the computational resources compared to models from 2021. Qualcomm and Google collaborating to bring advanced on-device AI capabilities leveraging Qualcomm's machine learning leadership.

With this extensive range of announcements plus revelations of new models like MusicLM for AI-generated music, and demonstrations of Google's progress in multimodal interactions via models like Gemini, the company aims to underscore its sustained push to not just compete, but lead advancements across multiple AI domains encompassing cloud infrastructure, developer tooling, consumer web services, enterprise applications, specialized vertical use cases like healthcare, creative pursuits like imagery and music alongside more efficient AI chip capabilities.

Sundar Pichai, Google's CEO, is reported to have made AI investment his number one priority with the company dedicating tremendous resources across research, product development and global industry partnerships to ensure Google Cloud, its Pixel hardware lineup and services like Search, Maps and Google Workspace continue providing users, enterprise customers and developers leading-edge AI capabilities in the coming years as more macro challenges like energy efficiency, data privacy, transparency and responsible AI come into broader focus.

Anthropic Brings Claude AI Assistant to Google Sheets

Anthropic, an AI startup founded by former OpenAI researchers, has released an intelligent productivity boosting plugin called Claude for Sheets that injects its AI assistant Claude's conversational capabilities directly into Google's cloud-hosted spreadsheet solution Sheets—one of the most popular business software tools globally across industries and roles.

Key features offered by Anthropic Claude for Sheets include:

- Handsfree data analysis and text generation: Using Claude AI allows users to get insights from data sets, generate explanatory notes, summarize key takeaways and more through simple voice or text based prompting powered by Claude's advanced language intelligence. No need to leave Sheets interface.

- Easy install and configuration: Users connect their Anthropic account API key during plugin installation and complete OAuth permission flow to activate Claude's abilities within Sheets through secured access.

- In-sheet functions to tap AI: Plugin introduces =CLAUDE() and =CLAUDEFREE() functions within Sheets to query Claude AI. Latter function gives advanced users more controls.

- Customizable outputs: Multiple parameters can be set to determine aspects like length of text generated, creativity levels etc allowing flexibility for different use cases within Sheets.

- Interactive tutorials and template sheets: Anthropic provides expansive guides including videos to master prompt engineering techniques for productive Claude usage directly within Sheets for various applications. Pre-made sheets with sample prompts accelerate user ramp up.

The integration is noteworthy for increased accessibility it provides to Anthropic's brands of AI, now available to the millions of Google Sheets users globally without needing to switch workflows. And for Claude AI, it promises significantly increased visibility and adoption leverage by plugging into a ubiquitous business software solution, riding on the shoulders of Sheets instead of having to independently acquire mindshare and build traction.

The move comes at a time when rival ChatGPT has also released compatibility with Google Workspace apps and Microsoft Office to boost utility. Anthropic's focus on innovative prompt engineering for safer AI outcomes continuesdifferentiating it however. Offering Claude's strengths within widely popular tools like Sheets opens possibilities for businesses, data analysts, financial modelers and many more categories to speed up workflows with AI as a productivity multiplier while benefiting from Claude's rigorously-defined Constitutional AI approach minimizing harmful model responses.

As global competition heats up to make enterprise and consumer AI assistants more accessible and embed them deeper into established SaaS solutions, Claude for Sheets signifies the widening arena of opportunity and innovation unfolding - where AI directly enhances specialized domains like data-driven business decision-making through the tools already entrenched across industries.

We are have llamaindec traing at 10 AM PST | 11.30 PM GST , Ecen if you can attlend live You can get the recording and code

Resgister here for Free https://decodingdatascience.com/llama

Thank You for Reading!

We hope you enjoyed this roundup of key developments and updates in the artificial intelligence space recently. As AI continues rapidly advancing and integrating into diverse aspects of society, we aim to keep you informed on notable innovations, partnerships, product announcements and meaningful conversations shaping the future of this potentially transformative technology.

We thank you deeply for subscribing to our newsletter and for your ongoing interest in understanding AI progress. We would love to hear your thoughts on any of the covered news items and welcome your perspectives.

Please feel free to hit reply and share feedback on what stood out to you in this edition, what else you would like us to cover in AI industry news, or any other topics we should be discussing related to responsible and ethical integration of AI globally. Keep an eye out for the next edition of our newsletter at Monday.

Lastly, if you find our news coverage valuable, please help spread the word by sharing within your networks. The more diverse insights we can collective gather, the more representative our AI policy conversations can be.

We thank you again for your readership and wish you a wonderful week ahead!