AI News Roundup – Anthropic releases “hybrid reasoning” model, protests against UK AI copyright measures, Huawei improves AI chip production, and more

To help you stay on top of the latest news, our AI practice group has compiled a roundup of the developments we are following.

• • WIRED reports that Anthropic has released version 3.7 of its Claude AI system. Claude 3.7 Sonnet, according to the company, is the world’s first “Hybrid Reasoning” AI model, which allows users to control the amount of time and resources the model uses when generating outputs. To accomplish this, the model has two modes – a standard mode, similar to past iterations of Claude, and an “extended thinking” mode intended for reasoning tasks. Additionally, users can specify a “budget” for query: “you can tell Claude to think for no more than N tokens, for any value of N up to its output limit of 128K tokens,” the company said, allowing for trade-offs between speed and quality of response. The new release follows a trend in AI systems focused on reasoning, which enables models to think through and thus better address complex problems related to mathematics, science and code generation. Indeed, Anthropic claims Claude 3.7 Sonnet outperforms competitor models such as OpenAI’s o1, DeepSeek’s R1 and Grok 3 in software engineering and agentic tasks. Claude 3.7 Sonnet is now available for all users of Anthropic’s products.

• • Over 1,000 music artists have released a silent album in protest of the United Kingdom’s proposed copyright rules relating to AI training, according to The New York Times. The group of artists, which includes Annie Lennox, Billy Ocean, Hans Zimmer and Kate Bush, recorded empty musical studios for the album, which the artists say would be the effect if the U.K.’s proposed copyright reforms go ahead. The proposed measures would open up copyrighted works to AI companies to train unless the authors affirmatively opt out, which “shifts the burden of controlling your works onto the rights holder,” according to an organizer of the album. Several critics of the proposed measures say that many artists may not know if their works are being used to train AI models, and that copyrighted works can often spread quickly online beyond the control of their creators. The relationship between copyright law and AI training has been fraught, with numerous lawsuits ongoing in the U.S., U.K. and other countries over the issue, though no country has yet to introduce formal rules on the topic. The U.K. government’s consultation on the proposed copyright rules ended this past week, and action from the government in Parliament is expected in the coming months.

• • The Financial Times reports that Huawei has ramped up its production of AI-focused semiconductors, improving China’s access to the vital technology. Huawei’s “yield” on its latest chips has reached nearly 40%, referring to the percentage of functional chips out of the total produced, a marked improvement over a yield of 20% from a year ago. Huawei’s line of AI-focused chips, under the brand name Ascend, is now profitable for the first time, and bolsters efforts to support China’s rapidly growing AI industry, despite U.S. bans on the export of advanced AI chips from companies such as Nvidia and TSMC to the country. The company plans to manufacture 100,000 Ascend 910C chips and 300,000 Ascend 910B chips this year, even though the latter has experienced issues with model training due to memory and connectivity issues.

• • OpenAI has released GPT-4.5, the most recent update to the AI system that kicked off the AI revolution in the autumn of 2022, according to Bloomberg. In a blog post, OpenAI described the model as its “largest and best model for chat yet,” and claims the model is less prone to hallucinations than previous GPT models. The model, known as Orion internally, had a troubled development cycle – it repeatedly fell short of performance benchmarks and expectations last year. OpenAI used a technique known as “post-training” to incorporate human feedback for outputs generated by the model. OpenAI CEO Sam Altman said on X (formerly Twitter) that the new model would be the last model that did not feature advanced reasoning capabilities, like OpenAI’s o1 model. Future models from the company are likely to combine the two into a single model. GPT-4.5 is currently available as a “research preview” to a limited group of developers for testing before a wider rollout expected this year.

• • AI-generated photos are warping customer expectations, especially at small businesses, according to The Washington Post. Several aesthetic-focused businesses, including wedding dress designers, hair salons, and plastic surgeons, have encountered a rising number of customers demanding looks sourced from AI-generated images as “inspo.” However, AI images can often include impossible components; one wedding dress image encountered contained no sleeves and no back, and “would have violated the laws of physics,” according to one designer. Such images are often sourced from generative AI systems such as Midjourney or OpenAI’s DALL-E and spread rapidly on social media platforms such as Pinterest and Instagram. According to research, over half of marketers use generative AI, including in social media ads which do not disclose whether the images are AI-produced, which could contribute to the alteration of customer expectations seen in recent months. Some professionals attempt to push back against the AI trends as unrealistic and promoting unhealthy choices: one Brooklyn hairstylist tells clients who present her with AI images to “stop bringing it in…[w]e’re trying to do creative work, and AI is just pushing perfection.”