Weekly review #3
Miko Pawlikowski ???
Follow for breakthrough ideas in tech, business & performance | Keynote speaker | Co-founder SREday.com
Here's everything you missed this week in the world of AI.
Friday 2024/05/31
?? Google is working on fixing AI Overview (5 min)
After a few days filled with discussion and some memes, Liz Reid, Head of Google Search, stated that the company is working to fix AI Overview and hopefully stop it from giving odd responses. This includes tweaking their AI to detect satire and sarcasm, excluding health topics from AI-generated responses, and preventing the system from searching for answers on social media and forum posts.
Nonetheless, Reid defended Google, claiming that virality plays a big part in the Overview situation, as plenty of the weird responses were to questions no one would have thought to ask the search engine until users noticed what was happening.
?? Claude engages with external data (1 min)
Anthropic has announced that Tool Use is now generally available. This feature allows Claude to interact autonomously with external data, APIs, and tools. Focusing mostly on enterprise solutions, Anthropic aims to help businesses integrate AI into their processes to streamline and automate tasks.
Tool Use is available for the entire Claude 3 model lineup, and several companies have already tried its power with repetitive and data-heavy tasks. It is a feature useful in a wide range of fields, from education to financial services.
?? Siri gets an AI revamp with iOS 18 (1 min)
One of Apple's biggest changes focused on generative AI will be dedicated to Siri. According to reports, the upcoming iOS 18 update will allow the assistant to take full control of iPhones, granting it access to analyse device activity, manage Apple apps through voice commands, and eventually handle multiple tasks.
As we mentioned before, we should not expect a revolutionary AI transformation from Apple, but rather a practical approach aimed at optimising the user experience and unlocking new levels of productivity.
??A big tech group that includes Google, Intel, Microsoft, and Meta, among others, has come together to develop an interconnected standard for AI acceleration chips in data centers. This initiative, Ultra Accelerator Link, plans to create new alternatives and reduce reliance on Nvidia's proprietary NVLink technology.
??Early research provides insight into how AI could reshape EV charging. By leveraging AI to analyse charging behaviours, researchers have outlined some of the main challenges drivers encounter with their EVs, such as power inconsistency and low power quality, which can ultimately damage charging equipment.
Thursday
???? Mistral AI introduces its first coding model (5 min)
As part of its mission to democratise access to AI and provide developers with new tools for code creation, Mistral AI has announced Codestral, its first code model. Codestral boasts a context window of 32,000 tokens and 22B parameters. Trained on a dataset of over 80 programming languages, it aims to assist developers in writing and interacting with code, minimising errors and reducing time spent coding.
Codestral is an open-weight model licensed under Mistral's new AI Non-Production License, making it suitable for research and testing.
OpenAI has disclosed another set of deals. Vox Media and The Atlantic have agreed to provide the company with access to their content to train AI models. In return, they will be credited as sources if ChatGPT happens to cite content from them as publishers. Also, both companies will leverage OpenAI's technology to enhance user experiences.
Licensing deals are currently very important to OpenAI, primarily for keeping ChatGPT up-to-date, and also for its potential plans to release a search engine.
???? Google announces AI expansion in Malaysia (1 min)
The demand for AI is prompting Google to further expand its efforts in Southeast Asia. The company has announced plans to invest $2 billion in Malaysia to build a data center and cloud region. This move has the potential to increase the country's GDP and support 26,500 jobs.
Google is not the only company making strategic moves globally. Microsoft has also taken similar steps as the need for more cloud computing services and data centers continues to grow rapidly alongside the development of AI technologies.
??Enterprises looking to integrate ChatGPT into their processes may have a new pathway: the multinational professional service consultancy PwC offers the option to purchase the ChatGPT Enterprise licence, previously available only directly from OpenAI. The two companies have partnered to expand the availability of AI technology for businesses.
??As announced in May, free ChatGPT users can now access features that were initially available only through a paid subscription, such as custom GPTs, chart creation, data analysis, and file uploads. However, while they can use custom GPTs, they won't be able to create their own.
Wednesday
?? OpenAI is training its next frontier model (2 min)
OpenAI has been the subject of much discussion recently, and not all of it has been good. Yet, the company is redirecting this attention to another topic with the announcement of the beginning of training for its next frontier model. Not much insight has been disclosed, so it is yet to be known if this model is going to be GPT-5 or something even more advanced.
On the other hand, the company also announced the formation of a "Safety and Security Committee" after the disbandment of its "Superalignment" team. This committee will oversee their projects and operations, assessing risk management and responsible AI development. It is led by a group of current OpenAI members, such as Bret Taylor, Adam D'Angelo, Nicole Seligman, and last but not least, Sam Altman.
A group of researchers from Microsoft and Beihang University have created a new technique for fine-tuning LLMs that is more cost-effective. MoRA improves traditional methods by focusing on updating only the necessary subset of parameters instead of all the model parameters.
MoRA, a type of "parameter-efficient fine-tuning," addresses the limitations developers often encounter with the LoRA (low-rank adaptation) approach by using square matrices to enhance the model's learning capabilities. Also, a special compression/decompression function allows MoRA to work with different model sizes.
?? Jan Leike turns to Anthropic (2 min)
领英推荐
Jan Leike, a former member of OpenAI's superalignment team, made headlines recently, primarily due to his criticism of the company's leadership. He noted that he no longer agreed with how they handled priorities. Leike has now joined Anthropic.
The news was shared by Leike himself through a post on X, where he announced his intention to continue his work monitoring the safety and security of AI, while also calling for new members to join his journey. This marks a significant step for Anthropic, a longstanding contender against OpenAI... As for taking the lead, it remains to be seen.
??A team from the University of Washington developed a system called "Target Speech Hearing" that leverages AI to allow a user wearing headphones to sync with the person they're looking at while talking. The system then "enrols" the person who's speaking and cancels every other sound, focusing solely on their voice, even if the headphone-wearing user moves.
??Apple's upcoming WWDC has officially been scheduled. The company announced that the keynote will start at 10 a.m. PDT and will be live-streamed on its website. After months of anticipation, we're finally going to learn about Apple's AI plans, which are expected to include practical features.
Tuesday
xAI is still young, but there's a lot in store for its future. Elon Musk's AI startup recently announced that they have secured a significant $6 billion in funding, which will be used to improve infrastructure, launch its first products to market, and advance its research efforts.
As we mentioned yesterday, xAI has some catching up to do if it wants to be on the same page as OpenAI, Meta, Microsoft, and Google. Yet, there's a possibility that it could end up surprising even the most sceptical. We've seen Grok's capabilities, and they're not falling short so this funding boost may be a turning point for the startup.
?? MIT study refines linear hypothesis in AI (2 min)
A team of researchers at MIT has discovered a novel way to understand how language models work. Previously, it was believed one-dimensional representations powered them, but there's more than meets the eye. Multi-dimensional features are also involved in their activation space, which alters the models' performance.
This research has been tested on Mistral 7B and Llama 3 8B models, where one of the multi-dimensional features found is based on circular representations. These cyclic patterns could be utilised for calendar-related tasks, such as determining the day of the week for a given date.
AI agents will be pretty important in future AI applications, and "TransAgents" is a good example of how they'll do that. Developed by researchers from Monash University, the University of Macau, and Tencent AI Lab, this innovation simulates a translation agency by assigning AI agents different roles with detailed profiles. Once the team is assembled, the group of "TransAgents" excels at literary translation.
The simulation includes every role found in a real translation agency, from editors to proofreaders, and it aims to help reduce translation costs. However, for now, it requires close supervision, so it's mostly a handy tool.
??Stephen Wolfram, well-known for his computer algebra and theoretical physics work, has shared his thoughts on GenAI. He claims that this technology is not a threat to humanity but rather a further development of automation that complements human action and elevates it to another level. Our AI expectations shouldn't be unrealistic, according to him.
??An increasing number of fields benefit from AI algorithms. One of these is agriculture, where researchers are leveraging AI-based tools to help breeders and farmers address challenges such as degrading soil, diseases, and climate change, improving the quality of their crops.
Monday
Google's AI search feature, "AI Overview," has been in the works since last year, and the company proudly rolled it out recently after arduous testing and tweaks. However, it has not been warmly received by users, who are currently complaining on social media and uploading screenshots of the search engine giving odd answers, prompting Google to rush to fix the issues.
The deployment of this feature has been anything but smooth for the company. Although it sounds like a promising integration for the search engine, many believe in the principle of "if it isn't broken, don't fix it." Yet, Google insists that the weird answers are mostly for uncommon queries, and they are taking action to address the matter, claiming that AI Overview has actually provided valuable information to users.
?? Apple’s WWDC is just days away (2 min)
As the countdown begins for one of the most anticipated developer conferences annually, there's a word to describe Apple's approach to AI in this one, and according to Mark Gurman, the word is "practical." The integration of AI into Apple's devices is mainly focused on allowing chatbots to run smoothly on their operating systems, with everything pointing to OpenAI being the first to join the party with ChatGPT. Apple is also reportedly in talks with Anthropic and Google.
Other rumoured AI features include transcribing voice memos, auto-generated emojis, writing suggestions for emails, and enhancements to Spotlight search.
?? Elon Musk reveals xAI’s supercomputer plans (1 min)
Elon Musk is pushing to get xAI ahead of competitors. The CEO announced to investors that the startup was working to build a supercomputer to power Grok's new version. There's not much detail on the matter, besides the plan to have the supercomputer running by fall 2025. Additionally, Oracle will be in charge of developing the computer, which will be powered by Nvidia's H100 GPUs.
xAI is still not at the same level as rivals Google or OpenAI, and Grok, its chatbot, is currently in the early stages. Yet, it promises to evolve into a notable participant to watch as the AI race continues.
??♂?Microsoft's CEO, Satya Nadella, has mixed feelings about the term "AI." During an interview with Bloomberg Television, Nadella explained that we shouldn't attribute human qualities to AI and it must be treated for what it is: a tool. He even proposed a new term, "different intelligence."
??TikTok is stepping up its AI game. The social media video platform has announced the integration of AI into ads through a creative suite that will help creators craft engaging content with simple prompts. This upcoming feature will be called "TikTok Symphony.”