The month of July has been a whirlwind of AI developments, from new and improving chatbots and AI assistants, through to new abilities for image generation tools and the rapid maturation of AI-generated video. Here's a quick review of the key news and insights from the intersecting worlds of AI and marketing in July 2023, which will be instrumental for marketing professionals looking to stay ahead of the curve.
The information in this blog post is derived from the thought-provoking discussions featured in episode 17, episode 18 and episode 19 of the "Artificially Intelligent Marketing" podcast, hosted by our CEO Paul Avery and his industry colleague Martin Broadhurst.
If you’d prefer, you can listen to episodes 17, 18 and 19 below, or subscribe on your favourite podcast platform.
OpenAI's ChatGPT User Growth and GPT-4 API Availability
OpenAI's language model, ChatGPT, has experienced some turbulence in its user growth. Data from SimilarWeb shows a decrease in the number of users from May to June, potentially indicating a stagnation or even a decrease in user engagement. In the midst of these changes, OpenAI has prioritised paying customers for access to its latest language model, GPT-4, through its API. Developers waiting for access to this cutting-edge technology should expect to have it by the end of July.
ChatGPT Internet Access Suspension
OpenAI made the decision to temporarily suspend ChatGPT's internet access following misuse of the tool to bypass paywalls and gain unauthorised access to gated content. This development has raised concerns about how AI tools can be used unethically and the potential impacts on information accessibility. It's unclear when or how OpenAI plans to reinstate this feature.
OpenAI's Automated Alignment Researcher
In a forward-thinking move, OpenAI has revealed plans to develop a human-level automated alignment researcher. The project is an ambitious step toward ensuring that superintelligent AI aligns with human values, and could significantly reduce dependence on the process of reinforcement learning through human feedback, which has limitations in terms of scalability and accuracy.
OpenAI's Code Interpreter
OpenAI's Code Interpreter, now made available for everyone, has a broad range of functionalities. It's capable of analysing and visualising data, modeling predictive scenarios, cleaning data, generating synthetic data, and even creating Python code. One feature that caught our attention is its unexpected ability to count the number of faces in an image, which makes us wonder what other capabilities it may have. The power of Code Interpreter has led to speculation that it might be an early version of GPT-4.5.
ChatGPT Performance Scrutiny
ChatGPT's performance has been scrutinised in a research paper published by academics at Stanford and Berkeley. They observed a significant decrease in the model's ability to perform tasks such as solving math problems and generating code over time. The most striking example was its ability to identify prime numbers, which plummeted from a success rate of 97.6% in March to just 2.4% in June.
ChatGPT Custom Instructions and Message Limits
OpenAI has introduced updates to ChatGPT to enhance its usability. Users now have the ability to send 50 messages every 3 hours (up from 25). They can also now set custom instructions, which allows them to provide context about themselves and more details on how they want ChatGPT to respond, making interactions more personalised and effective. This feature was highly appreciated by users, especially for its capability to produce better outputs without the need for proving long, contextual prompts at the start of every conversation.
Anthropic releases Claude 2, LLaMa 2 made available for commercial use
Claude 2, a new language model, has demonstrated performance almost on par with ChatGPT-4 in various standard benchmark tests, making it a serious alternative to OpenAI's tool (especially considering that Claude-2 is currently free to use). Claude-2 has also been integrated into Jasper, an AI copywriting tool. Jasper's team highlighted Claude 2's up-to-date knowledge training as a major advantage (it was trained on data up to and including January 2023, compared to mid-2021 for GPT-4).
Meanwhile, tech giant Meta has also rolled out updates to its AI offering, releasing a commercial version of its open-source language model, LLaMA 2, making it an interesting competitor to ChatGPT. The open source nature of LLaMa 2 means that companies and hobbyists can build on top of the platform to expand its abilities or customise it for a specific use case. As a rapid example of this, developers have already increased LLaMa 2's default context window from 4k tokens to 32k, significantly improving it's ability to handle a large amount of context in the initial prompt (e.g. a whole document to summarise).
Microsoft Making Waves with New Office 365 AI Pricing Model
Microsoft recently announced plans to start charging $30 per user per month for access to its new Copilot AI features in Office products like Word, Excel, and PowerPoint. By charging specifically for AI capabilities on top of existing Office license costs, Microsoft is establishing a new precedent where AI is treated as a premium add-on service. This pricing strategy raises many questions about the future of productivity software licensing and revenue models. It remains to be seen if other tech companies will follow Microsoft's lead and charge extra for AI access. However, Microsoft appears to be betting that AI will become a major profit driver despite potential customer pushback. The $30 AI fee could significantly increase costs for enterprise customers in particular. Either way, the tech world will be watching closely to see how this pricing experiment impacts the proliferation of AI in mainstream business software.
AI in Hollywood
AI has made its mark even in Hollywood, with writers and actors going on strike over issues related to AI's impact on their profession (among other things). The issue highlights the broad reach of AI technology and its potential to disrupt many industries.
AI in Scientific Discoveries
A study published in Nature Human Behaviour presented evidence of how AI could assist and expand scientific discoveries. By leveraging AI to predict and generate "alien" hypotheses that are unlikely to be considered by humans, researchers can explore areas and possibilities that might not otherwise be considered, opening up new avenues for scientific breakthroughs.
"GPT Author" Creates Novels From Simple Text Prompts
GPT Author, a project on GitHub, is putting a new twist on novel creation. By utilising GPT-4 and Stable Diffusion API calls, GPT Author is a automated, multi-step process that can generate an entire original fantasy novel from a user's initial prompt. The AI then produces an EPUB file compatible with e-book readers. The process is astonishingly efficient, costing as little as $4 to produce a 15-chapter novel in just a few minutes.
LLMs for Business Understanding and Content Production
More and more marketers are recognising the potential of LLMs to gain a better understanding of their businesses, leverage their content, and utilise internal information. Such tools could be used for a variety of purposes, such as enhancing company-specific content production and marketing efforts, creating specialised customer service chatbots, or developing internal knowledge products. As these models continue to evolve, product managers and marketers will need to keep an eye on this space.
ClipDrop's Stable Doodle and Midjourney's Updates
In the realm of AI image generation, Stability.ai's ClipDrop launched Stable Doodle, a tool that uses sketches as a part of the image generation prompt. This novel approach brings a new level of user interaction to image generation. Meanwhile, Midjourney introduced a "Panning" feature in its image generation tool. This feature allows users to extend images in all directions and even change the prompt during the extension process, providing a dynamic and interactive image creation experience. At the same time, Midjourney users are finding creative uses for the new zoom tool. Users are creating infinite zoom outs and combining them with Adobe After Effects to produce intriguing video content. This emerging trend is a testament to the evolving digital creativity landscape.
Meta recently announced CM3Leon, a new image generation model that can generate images from text prompts and vice versa. One of the more intriguing features of CM3Leon is its ability to modify existing images at specific locations within an image, opening up exciting possibilities for creative applications that are not easily delivered using other image generation tools like Midjourney and DALLE-2.
AI in E-Commerce and Web Development
In the e-commerce sector, Shopify launched an AI-driven support agent called Sidekick. The agent is designed to provide business insights and help with bulk actions, making e-commerce management smoother and more efficient. Meanwhile, Wix, a popular web development platform, announced the integration of OpenAI LLMs in its AI website builder. The new integration will allow users to generate professional-quality websites from simple text prompts.
AI Risk Management
In an important stride toward responsible AI use, leading companies in the AI sector, including Amazon, Google, IBM, Microsoft, NVIDIA, Oracle, and Salesforce, have committed to implement safeguards to manage risks associated with AI technology. This collective move sets a precedent for the entire industry in prioritising safety and ethics in AI. Of particular interest is the commitment to watermark AI-generated content (images, video, audio etc.) to help audiences tell the difference between real content and fake news. It's not clear how this will be achieved in a reliable way or when such systems would be rolled out, so stay tuned for more updates here.
New AI Initiatives from Elon Musk and Apple
Elon Musk, always at the forefront of tech innovation, has announced the creation of xAI, a company focused on creating a “good AGI” (Artificial General Intelligence). Rumors are also circulating that Apple is investing in new AI products to compete with major players like OpenAI and Google, which suggests that, rather than sitting back and watching the explosion of AI-driven tools, the teams at Apple have been beavering away in the background... it will be interesting to see what they release to the public (and when).
Microsoft Brings AI Shopping Tools To Bing And Edge
In a move to make shopping online even more seamless, Microsoft is integrating new AI-powered shopping tools into Bing and its Edge browser. The tools include an automatic buying guide generator, AI-generated review summaries, and a Price Match tool, all designed to streamline the shopping experience. Interestingly, the buying guide feature utilises Bing's GPT-powered AI to list and compare product specifications, potentially impacting websites that focus on creating buying guides (and E-Commerce in general).
AI-generated Tweets More Convincing Than Real People
A recent study revealed that people tend to trust tweets generated by GPT-3 more than those written by humans. The research involved a survey where participants from several countries, including the UK, Australia, Canada, the US, and Ireland, were asked to identify the source (human or GPT-3) of a tweet and its validity. The study, published in Science Advances, concluded that GPT-3 generated content was "indistinguishable" from organic content.
Generative AI Could Boost the UK Economy
According to a report by KPMG, generative AI could add an additional £31 billion to the UK's GDP annually. This translates into a 1.2% boost in productivity. The report also mentioned that about 2.5% of tasks could be performed by generative AI, impacting 40% of jobs in some capacity.
OpenAI Planning to Turn ChatGPT into a Work Assistant
OpenAI is reportedly planning to transform ChatGPT into a "supersmart personal assistant for work". This could potentially revolutionise how tasks such as drafting emails or documents are handled in the workplace, as the AI assistant would be equipped with an in-depth knowledge of individual employees and their work.
OpenAI's Expansion to London
OpenAI, an AI research lab originally established in San Francisco, is breaking new ground by opening its first expansion office in London. This move signifies a major step towards OpenAI's global expansion efforts. It's a strategic move, considering London's status as a hub of technological innovation and access to diverse talent.
Google DeepMind’s Next Algorithm
Google DeepMind’s CEO, Demis Hassabis, has big plans for its new Gemini system. In an interview with Wired Magazine, Hassabis revealed that the company is working on combining language model technology with techniques used in AlphaGo to create a system more powerful that ChatGPT-4. The aim is to enhance the system's capabilities to solve complex problems and provide new functionalities like planning. This could have significant implications for a variety of sectors, including gaming, healthcare, and business.
DragGAN Now Available on Hugging Face
Remember DragGAN, which was mentioned in a previous episode of the podcast? It's the AI model that can adjust images simply by dragging them. Well, it's now available on Hugging Face, a popular platform for natural language processing models. With this release, more developers can explore and utilise the power of DragGAN for their projects.
Insilico Medicine Enters Phase 2 Clinical Trials
Insilico Medicine, a prominent member of NVIDIA Inception, has achieved a significant milestone by entering Phase 2 clinical trials with a drug candidate discovered using its AI platform. This achievement underscores the potential of AI in expediting drug discovery, having reached the first phase of clinical trials in just two and a half years after starting the project, a process that typically takes up to six years and costs more than $400 million when using traditional methods.
Amazon Builds LLM Marketplace
Amazon is reportedly taking a unique approach to AI development. Instead of directly competing with Microsoft and Google to build the next ChatGPT, Amazon is reportedly working on compiling the best large language models (LLMs) in one location, essentially creating a marketplace. The initiative aims to offer a convenient access point to foundational models from companies like Stability, Falcon, and Anthropic via its SageMaker JumpStart marketplace. If successful, this could revolutionise the way businesses choose and leverage AI models for their operations.
Stability AI Unveils SDXL 1.0 for image generation
Stability AI has launched SDXL 1.0, a groundbreaking open model for text-to-image generation. This model, boasting a whopping 10.1 billion parameters, can generate high-quality images in any art style, including photorealism. Its advanced language interpretation capabilities set it apart, allowing it to distinguish between similar terms with different meanings. The quality of SDXL 1.0 is impressive, rivalling that of Midjourney. The developers claim it has the power to generate legible text in images, making it a first for AI-driven image generation tools (although, in our experience, it is extremely hit and miss). You can test it out now using ClipDrop, DreamStudio or PlaygroundAI.
Rewind Launches Personalized AI App for iPhone
Rewind, an AI-driven app, has launched a new personalised app for iPhone users. Powered by OpenAI’s GPT-4 model, Rewind acts as a personal AI time-traveller, allowing users to record, store, and rewind their work. This news builds upon Rewind's existing app for Mac.
Amazon Introduces Agents for Bedrock
Amazon has announced the launch of Agents for Bedrock at the AWS Summit New York. The new feature allows customers to create conversational agents that can deliver personalized, up-to-date answers and execute actions.
Cohere Announces Coral, a Knowledge Assistant for Enterprises
Cohere has announced Coral, a knowledge assistant designed to enhance the productivity of strategic teams within enterprises. Coral can find answers across your company’s documents and data sources, providing responses backed with citations, ensuring the information provided is verifiable. The system currently integrates with over 100 other tools, such as CRMs, project management software, knowledge bases etc. Importantly, Cohere cannot access your company’s proprietary info, making this a secure option for large companies concerned about what other players such as OpenAI may be doing with the info they provide to tools like ChatGPT.
ChatGPT Now Available on Android
ChatGPT, the AI-powered chatbot, is now available on Android in several countries, including the US and UK. The app allows users to dictate their prompt (a nice additional feature that is not available on the desktop browser version), as well as sync chats across devices.
FraudGPT: A New Threat in Cybersecurity
FraudGPT is a new generative AI tool being advertised on the Dark web and Telegram channels that offers capabilities to cybercriminals to help them improve and scale the production of malicious code, phishing campaigns etc. The tool has already got over 3,000 confirmed sales and reviews. We’ll all need to be even more vigilant when it comes to phishing campaigns from here on out.
ElevenLabs Releases New Voices
ElevenLabs has released new voices created in collaboration with industry professionals. These voices offer a wider range of delivery styles, accents, and improved audio quality. They even include options that vary from whispering through to screaming.
Runway GEN2 Image to Video Made Widely Available
Runway has made additional elements of its GEN2 image to video tool more widely available. GEN-2 is a tool that allows users to create 4-second videos from text or images. Now that you can more easily use images to drive the prompts, the results are impressive, providing a clear improvement over text to video, as you can more easily control the production of the video using a starting image.
Stability AI Releases Two New Open-Source Large Language Models
Stability AI has released two new open-source Large Language Models, FreeWilly1 and FreeWilly2, developed by its CarperAI lab. These models excel in reasoning and understanding linguistic subtleties and have been validated through various benchmarks. Despite being trained on a smaller dataset compared to previous models, the FreeWilly models demonstrate exceptional performance.
Discover More with the Artificially Intelligent Marketing Podcast
This blog offers just a glimpse into the rich insights shared by Paul and Martin in their captivating podcast episodes. We of course recommend exploring the full episodes, available below or on your preferred podcast platform, for a comprehensive understanding of the fascinating world of AI in marketing. You can also delve deeper into the archive of past episodes, where you'll find a treasure trove of knowledge about AI's impact on the marketing landscape.
The creation of this blog post was made possible with the support of AI technology. The podcast audio was initially transcribed using AI, and GPT-4 (ChatGPT) assisted in summarising the transcript into easily digestible blog sections. Our BioStrata human team then meticulously reviewed and edited these sections to ensure clarity and coherence. Additionally, AI technology from Clipdrop played a pivotal role in crafting the blog image.