Researchers from Cohere and the University of Oxford have introduced an innovative method to enhance reward models (RMs) in reinforcement learning from human feedback (RLHF) by leveraging large language models (LLMs) for synthetic critiques. The novel approach aims to reduce the extensive time and cost associated with human annotation, which is traditionally required for training RMs to predict scores based on human preferences.
In their paper, ‘Improving Reward Models with Synthetic Critiques’, the researchers detailed how LLMs could generate critiques that evaluate the relationship between prompts and generated outputs, predicting scalar rewards. These synthetic critiques improved the performance of reward models on various benchmarks by providing additional feedback on aspects like instruction following, correctness, and style, leading to better assessment and scoring of language models.
The study highlighted that high-quality synthetic critiques significantly increased data efficiency, with one enhanced preference pair as valuable as forty non-enhanced pairs. The approach makes the training process more cost-effective and has the potential to match or surpass traditional reward models, as demonstrated by GPT-4.0’s performance in certain benchmarks.
As the field continues to explore alternatives to RLHF, including reinforcement learning from AI feedback (RLAIF), this research indicates a promising shift towards AI-based critiquing, potentially transforming how major AI players such as Google, OpenAI, and Meta align their large language models.
The AI industry invests heavily in digital twin technology, creating virtual replicas of humans and objects for research. Tech companies believe these digital twins can unlock AI’s full potential by mirroring our physiologies, personalities, and objects around us. Digital twins can range from models of complex phenomena, like organisms or weather systems, to video avatars of individuals. This new technology promises to revolutionise healthcare by providing personalised treatment, accelerating drug development, and enhancing our understanding of environments and objects.
Gartner predicts the global market for digital twins will surge to $379 billion by 2034, mainly driven by the healthcare industry, which is expected to reach a market size of $110.1 billion by 2028. The concept of digital twins began in engineering and manufacturing but has expanded thanks to improved data storage and connectivity, making it more accessible and versatile.
One notable example is LinkedIn co-founder Reid Hoffman, who created his digital twin, REID.AI, using two decades of his content. Hoffman demonstrated the potential of this technology by releasing videos of himself conversing with the twins and even sending them for an on-stage interview. While most digital twins focus on statistical applications, their everyday utility is evident in projects like Twin Health, which uses sensors to monitor patients’ health and provide personalised advice. The technology has shown promise in helping diabetic patients reverse their condition and reduce medication reliance.
Like the broader AI boom, the digital twin market starts with impressive demonstrations but aims to deliver significant practical benefits, especially in healthcare and personalised services.
Samsung Electronics announced securing an order from Japanese AI company Preferred Networks to manufacture chips using its advanced 2-nanometre foundry process and advanced chip packaging service. The trade exchange between the two countries marks Samsung’s first disclosed order for its cutting-edge 2-nanometre chip manufacturing process, although the order size remains undisclosed.
The chips will employ high-tech gate-all-around (GAA) architecture and integrate multiple chips into a single package to enhance connection speed and reduce size. According to Preferred Networks ‘ VP Junichiro Makino, designed by South Korea‘s Gaonchips, these chips will support high-performance computing hardware for generative AI technologies, including large language models.
The development highlights Samsung’s advancements in semiconductor technology and its role in supporting innovative AI applications.
IBM has launched its GenAI Innovation Center in Kochi, designed to help enterprises, startups, and partners explore and develop generative AI technology. The centre aims to accelerate AI innovation, increase productivity, and enhance generative AI expertise in India, addressing challenges organisations face when transitioning from AI experimentation to deployment.
The centre will provide access to IBM experts and technologies, assisting in building, scaling, and adopting enterprise-grade AI. It will utilise InstructLab, a technology developed by IBM and Red Hat for enhancing Large Language Models (LLMs) with client data, along with IBM’s ‘watsonx’ AI and data platform and AI assistant technologies. The centre will be part of the IBM India Software Lab in Kochi and managed by IBM’s technical experts.
IBM highlights that the centre will nurture a community that uses generative AI to tackle societal and business challenges, including sustainability, public infrastructure, healthcare, education, and inclusion. The initiative underscores IBM’s commitment to fostering AI innovation and addressing complex integration issues in the business landscape.
Why does it matter?
lBM’s new GenAI hub stems from a significant investment in advancing AI technology in India. This centre is set to play a crucial role in accelerating AI innovation, boosting productivity, and enhancing generative AI expertise, which is critical for the growth of enterprises, startups, and partners. By providing access to advanced AI technologies and expert knowledge, the centre aims to overcome the challenges of AI integration and deployment, thereby fostering a robust AI ecosystem. Furthermore, the initiative underscores the potential of generative AI to address pressing societal and business challenges, contributing to advancements in sustainability, public infrastructure, healthcare, education, and inclusion.
US tech giant Microsoft is committed to offering generative AI services in Hong Kong through educational initiatives, despite OpenAI’s access restrictions in the city and mainland China. Microsoft collaborated with the Education University of Hong Kong Jockey Club Primary School to offer AI services starting last year.
About 220 students in grades 5 and 6 used Microsoft’s chatbot and text-to-image tools in science classes. Principal Elsa Cheung Kam Yan noted that AI enhances learning by broadening students’ access to information and allowing exploration beyond textbooks. Vice-Principal Philip Law Kam Yuen added that in collaboration with Microsoft Hong Kong for 12 years, the school plans to extend AI usage to more classes.
Additionally, Microsoft also has agreements with eight Hong Kong universities to promote AI services. Fred Sheu, national technology officer of Microsoft in Hong Kong, reaffirmed Microsoft’s commitment to maintaining its Azure AI services, which use OpenAI’s models, further emphasising that API restrictions by OpenAI will not affect the company. Microsoft’s investment in OpenAI reportedly allows it to receive up to 49% of the profits from OpenAI’s for-profit arm. As all government-funded universities in Hong Kong have already acquired the Azure OpenAI service, they are thus qualified users. He also emphasised that Microsoft intends to extend this service to all schools in Hong Kong over the next few years.
A recent survey by Tracklib reveals that 25% of music producers are now integrating AI into their creative processes, marking a significant adoption of technology within the industry. However, most producers exhibit resistance towards AI, citing concerns over losing creative control as a primary barrier.
Among those using AI, the survey found that most employ it for stem separation (73.9%) rather than full song creation, which is used by only a small fraction (3%). Concerns among non-users primarily revolve around artistic integrity (82.2%) and doubts about AI’s ability to maintain quality (34.5%), with additional concerns including cost and copyright issues.
Interestingly, the survey highlights a stark divide between perceptions of assistive AI, which aids in music creation, and generative AI, which directly generates elements or entire songs. While some producers hold a positive view of assistive AI, generative AI faces stronger opposition, especially among younger respondents.
Overall, the survey underscores a cautious optimism about AI’s future impact on music production, with 70% of respondents expecting it to have a significant influence going forward. Despite current reservations, Tracklib predicts continued adoption of music AI, noting it is entering the “early majority” phase of adoption according to technology adoption models.
Microsoft has warned about a new jailbreaking technique called Skeleton Key, which can prompt AI models to disclose harmful information by bypassing their behavioural guidelines. Detailed in a report published on 26 June, Microsoft explained that Skeleton Key forces AI models to respond to illicit requests by modifying their behavioural guidelines to provide a warning rather than refusing the request outright. A technique like this, called Explicit: forced instruction-following, can lead models to produce harmful content.
The report highlighted an example where a model was manipulated to provide instructions for making a Molotov cocktail under the guise of an educational context. The prompt allowed the model to deliver the information with only a prefixed warning by instructing the model to update its behaviour. Microsoft tested the Skeleton Key technique between April and May 2024 on various AI models, including Meta LLama3-70b, Google Gemini Pro, and GPT 3.5 and 4.0, finding it effective but noting that attackers need legitimate access to the models.
Microsoft has addressed the issue in its Azure AI-managed models using prompt shields and has shared its findings with other AI providers. The company has also updated its AI offerings, including its Copilot AI assistants, to prevent guardrail bypassing. Furthermore, the latest disclosure underscores the growing problem of generative AI models being exploited for malicious purposes, following similar warnings from other researchers about vulnerabilities in AI models.
Why does it matter?
In April 2024, Anthropic researchers discovered a technique that could force AI models to provide instructions for constructing explosives. Earlier this year, researchers at Brown University found that translating malicious queries into low-resource languages could induce prohibited behaviour in OpenAI’s GPT-4. These findings highlight the ongoing challenges in ensuring the safe and responsible use of advanced AI models.
Adobe is expanding its generative AI team in India, seeking researchers skilled in NLP, LLMs, computer vision, deep learning, and more. With approximately 7,000 employees already in India, Adobe aims to bolster its research capabilities across various AI domains. Candidates will innovate and prototype AI technologies, contributing to Adobe’s products, publishing research, and collaborating globally.
Successful applicants are expected to demonstrate research excellence and a robust publication history, with backgrounds in computer science, electrical engineering, or mathematics. Senior roles require a minimum of seven years’ research experience, coupled with strong problem-solving abilities and analytical skills. Adobe prioritises integrating generative AI across its Experience Cloud, Creative Cloud, and Document Cloud, aiming to enhance content workflows and customer interactions.
Adobe’s foray into generative AI began with Adobe Firefly in collaboration with NVIDIA in March 2023. The company recently integrated third-party AI tools such as OpenAI’s Sora into Premiere Pro, offering users flexibility in AI model selection.
By partnering with AI providers like OpenAI, RunwayML, and Pika, Adobe continues to innovate, enabling personalised and efficient content creation workflows for enterprise customers.
Why does it matter?
The IATSE’s tentative agreement represents a significant step forward in securing fair wages and job protections for Hollywood’s behind-the-scenes workers, ensuring that the rapid technological advancements do not come at the expense of human employment.
YouTube has introduced an updated eraser tool that allows creators to remove copyrighted music from their videos without affecting speech, sound effects, or other audio. Launched on 4 July, the tool uses an AI-powered algorithm to target only the copyrighted music, leaving the rest of the video intact.
Previously, videos flagged for copyrighted audio faced muting or removal. However, YouTube cautions that the tool might only be effective if the song is easy to isolate.
Good news creators: our updated Erase Song tool helps you easily remove copyright-claimed music from your video (while leaving the rest of your audio intact). Learn more… https://t.co/KeWIw3RFeH
YouTube chief Neal Mohan announced the launch on X, explaining that the company had been testing the tool for some time but struggled to remove copyrighted tracks accurately. The new AI algorithm represents a significant improvement, allowing users to mute all sound or erase the music in their videos. Advancements like this are part of YouTube’s broader efforts to leverage AI technology to enhance user experience and compliance with copyright laws.
In addition to the eraser tool, YouTube is making strides in AI-driven music licensing. The company has been negotiating with major record labels to roll out AI music licensing deals, aiming to use AI to create music and potentially offer AI voice imitations of famous artists. Following the launch of YouTube’s AI tool Dream Track last year, which allowed users to create music with AI-generated voices of well-known singers, YouTube continues to engage with major labels like Sony, Warner, and Universal to expand the use of AI in music creation and licensing.
Why does it matter?
The IATSE’s tentative agreement represents a significant step forward in securing fair wages and job protections for Hollywood’s behind-the-scenes workers, ensuring that the rapid advancements in technology do not come at the expense of human employment.
Actor Morgan Freeman, renowned for his distinctive voice, recently addressed concerns over a video circulating on TikTok featuring a voice purportedly his own but created using AI. The video, depicting a day in his niece’s life, prompted Freeman to emphasise the importance of reporting unauthorised AI usage. He thanked his fans on social media for their vigilance in maintaining authenticity and integrity, underscoring the need to protect against such deceptive practices.
This isn’t the first time Freeman has encountered unauthorised use of his likeness. Previously, his production company’s EVP, Lori McCreary, encountered deepfake videos attempting to mimic Freeman, including one falsely depicting him firing her. Such incidents highlight the growing prevalence of AI-generated content, prompting discussions about its ethical implications and the need for heightened awareness.
Thank you to my incredible fans for your vigilance and support in calling out the unauthorized use of an A.I. voice imitating me. Your dedication helps authenticity and integrity remain paramount. Grateful. #AI#scam#imitation#IdentityProtection
Freeman’s case joins a broader trend of celebrities, from Taylor Swift to Tom Cruise, facing similar challenges with AI-generated deepfakes. These instances underscore ongoing concerns about digital identity theft and the blurred lines between real and fabricated content in the digital age.
