Amazon’s CTO on how developers can ride the AI-powered coding wave

Amazon CTO Werner Vogels told Fortune at the UN's AI for Good Summit that software engineers must become 'Renaissance developers' who combine deep technical skills with the ability to review AI-generated code, as AI coding tools like Claude Code transform the industry. The summit also featured UN chief António Guterres calling for global AI regulation, particularly on autonomous weapons, and debates over U.S. export controls on AI models.

Welcome to Eye on AI. Beatrice Nolan here. In today’s issue: - Amazon’s CTO on the AI coding revolution. - All the news from AI for Good. - SpaceXAI launches Grok 4.5. - And OpenAI says a key benchmark is broken. I’ve been on the ground in Geneva this week at the UN’s AI for Good https://aiforgood.itu.int/ Summit. The Summit is meant to bring together governments, industry leaders, researchers, and civil society to explore how AI can be harnessed to address global challenges and advance the UN’s Sustainable Development Goals. The conversations on stage ranged from how to tackle a growing AI divide between the global north and the global south to technical solutions for mitigating AI risks https://fortune.com/2026/01/15/ai-godfather-yoshua-bengio-changes-view-on-ai-risks-sees-fix-becomes-optimistic-lawzero-board-of-advisors/ like models engaging in deception and sycophancy. Before the Summit, global leaders gathered at the UN’s Global Dialogue on AI, the first intergovernmental AI governance summit to bring all 193 UN member states together to discuss potential international rules for controlling the technology. The UN chief António Guterres used the event to appeal for worldwide AI regulation, particularly when it comes to lethal autonomous weapons https://fortune.com/2026/02/27/pentagon-brands-anthropic-ceo-dario-amodei-a-liar-with-a-god-complex-as-deadline-looms-over-ai-use-in-weapons-and-surveillance/ , as AI technology increasingly shifts from civilian use to the battlefield. It’s not the first time the UN has attempted to put some controls on autonomous weapon use—Guterres first raised the alarm over lethal autonomous weapons in a 2023 policy paper, calling for a legally binding treaty banning “killer robots” by 2026. That deadline has now passed with no treaty in place. Also a recurring issue on the table at the Summit: creatives asking for a seat at it. Björn Ulvaeus, best known as co-founder and member of Swedish pop group ABBA , kicked off the proceedings on Tuesday, arguing that AI wouldn’t exist without creatives like him. Several speakers followed with various appeals for tech companies to recognize that contribution. On policy and governance: Salesforce’s Marc Benioff and Microsoft’s Brad Smith pushed back on the idea that recent export controls imposed on Anthropic’s Fable Five model were explicitly aimed at blocking foreign nationals from American AI. Both Benioff and Smith said the U.S. government was trying to address real-world national security concerns, rather than trying to deprive foreign nationals of the model. That said, the U.S. government’s actions have caused panic across Europe, where politicians worry they’re losing control over a fundamental technology. I also spoke with Amazon’s chief technology officer, Werner Vogels, who laid out his roadmap for software engineers trying to navigate through the AI boom currently disrupting their industry. The rise of the “renaissance developer” Software engineering is going through its most dramatic transformation in years. At the center of the shift are AI coding tools like Claude Code that can generate software with natural language prompts, reducing the need for engineers to write it line by line. Also known as “vibe-coding,” the process also allows non-engineers and novices to spin up prototypes in minutes, although this has been done with varying degrees of success. According to Vogels, these new capabilities are making the ability to successfully review and fact-check code more important than ever. Someone still has to catch what the model gets wrong, he told Fortune , especially when regulated industries or safety-critical systems are involved. “You can’t say to the regulator, oh, AI made a mistake,” he said. “That doesn’t work like that.” How to be successful in this new era of software engineering? Vogels thinks that means becoming what he calls a “Renaissance developer”—his term for engineers who combine deep technical expertise with broad, cross-disciplinary curiosity, the way Leonardo https://fortune.com/company/leonardo/ da Vinci’s studies of anatomy and bird flight fed into his engineering and inventions. Vogels describes it as a “T-shaped” model: deep in one domain, but broad enough to understand the systems and people that domain serves. Even with his own engineers, Vogel says he advises them to take one afternoon a week away from their normal workload to read a paper or test a new tool. Is AI actually killing entry-level jobs? There has been widespread concern about the rise of AI coding tools narrowing the number of employment opportunities for entry-level workers. However, according to Vogels, the anxiety around displacing junior engineers is primarily noise. “Every day is a new model, every day is a new system,” he said, adding that the pace of announcements—and geopolitical battles over which country’s models lead—leaves even him “confused at times.” His advice for junior engineers is to build skills beyond programming itself. When hiring, Vogels said he now weighs collaboration and teamwork over raw technical fluency—things like whether a candidate has worked on an open-source project or has an example of working well inside a team. Programming languages, he said, can be picked up in a month or two once someone knows how to learn. With that, here’s more AI news. Beatrice Nolan beatrice.nolan@fortune.com mailto:beatrice.nolan@fortune.com @beafreyanolan https://x.com/beafreyanolan FORTUNE ON AI Microsoft’s Brad Smith on Washington’s AI policy: ‘Regulation without transparent or complete rules’ https://fortune.com/2026/07/09/microsoft-brad-smith-washingtons-ai-policy-regulation-without-transparent-rules/ — Beatrice Nolan AI IN THE NEWS OpenAI's new voice models aim for more natural conversations. OpenAI has launched two new voice models, GPT-Live-1 and a smaller mini version, designed to sound more human and handle conversation more smoothly. Unlike the previous setup that stitched together separate transcription, language, and speech-generation systems, the new models can listen and speak simultaneously, letting users interrupt naturally. The mini model will become the new default in ChatGPT's voice mode, with paid subscribers getting the full version. Both can tap GPT-5.5 for reasoning and search mid-conversation, stay silent while tracking context, and even surface visual information—a feature also being pursued by startups like Monogram. OpenAI executives suggested voice could become a primary interface for complex, long-running agentic tasks, as rivals including Apple and Amazon pursue similar conversational upgrades. Some are specualting the new GPT-Live models may help to power the new AI device OpenAI currently has under development with former Apple designer Jony Ive. Read more in TechCrunch. White House denies giving OpenAI a "green light" for GPT-5.6 release. Confusion continued this week over who actually holds the authority to approve new AI model releases in the U.S. Axios reported that OpenAI's decision to publicly launch its GPT-5.6 model—released in three tiers, Sol, Terra, and Luna—followed a "green light" from the Trump administration. Altman also confirmed the model, https://x.com/sama/status/2074709023807664454?s=20 Sol, was being released on Thursday. But the White House later denied this, telling the outlet that no such approval is required or granted and that release timing rests entirely with companies. This stance aligns with a June executive order that explicitly ruled out mandatory government preclearance for new models, the White House said. But, OpenAI had previously said the Trump Administration had requested it limit the release of GPT-5.6 to government-approved customers after concerns were raised about the model's capability advances. The administration has shown it can intervene more aggressively when it chooses to: last month, the Department of Commerce forced Anthropic to pull its Fable 5 and Mythos 5 models offline over national security concerns, a move that drew criticism from cybersecurity and legal experts and sparked a global panic over sovereign AI. The muddle seems to reflect what many critics contend: that the U.S. is operating a licensing regime for frontier AI models but in a completely non-transparent way, in part because it does not want to admit for ideological reasons that it is, in fact, operating a de facto licensing regime. Read more in Axios. https://www.axios.com/2026/07/08/openai-gpt-trump-ban-lifted SpaceXAI launches Grok 4.5. Pitched by CEO Elon Musk as "Opus-class," SpaceXAI has released Grok 4.5, the company's first model since going public. SpaceXAI has positioned it as a versatile model for coding, office work, research, and writing, with claimed token efficiency roughly double that of rival models. Benchmarks released alongside the launch showed Grok competitive with—though not quite matching—top models from competitors. Elon Musk went further on X, calling it an "Opus-class model" and later comparing it specifically to Anthropic's Opus 4.7, though faster and cheaper. Grok 4.5 runs $2/$6 per million input/output tokens, versus $5/$25 for Opus 4.7 and up to $5/$30 for OpenAI's priciest GPT-5.6 tier, Sol, which is itself set for release this week after earlier delays. Read more in TechCrunch. Meta breaks ground on its first major Canadian data center. Meta is building its first large Canadian data center, a 1-gigawatt facility in Alberta's Sturgeon County, at a cost of around $9 billion and a two-to-three-year construction timeline. It's Meta's 33rd data center and part of its broader AI infrastructure buildout, with Alberta chosen for its abundant energy supply, favorable regulatory environment, and industrial zoning. The move comes as Meta simultaneously plans a new cloud computing business to sell excess capacity, even as investors remain skeptical of its roughly $145 billion capex forecast and the company's lag behind AI leaders OpenAI, Anthropic, and Google. Meta's stock is down about 9% this year. Read more in CNBC. https://www.cnbc.com/2026/07/08/meta-is-building-its-first-big-data-center-in-canada-amid-ai-push.html News outlets seek sanctions over alleged OpenAI evidence destruction. The New York Times and other publishers have filed a motion in Manhattan federal court asking a judge to sanction OpenAI, accusing the company of misleading the court about its ability to search training data and chat logs and deleting output records in a copyright. The plaintiffs include the Times , the Daily News , the Center for Investigative Reporting , the Intercept, and digital publisher and CNET parent Ziff Davis. The motion comes amid a broader copyright fight in which the Times and other publishers have sued OpenAI and Microsoft over the company's use of articles to train generative AI models. Read more in Variety. https://variety.com/2026/biz/news/new-york-times-news-outlets-accuse-openai-of-lying-lawsuit-1236805648/ EYE ON AI NUMBERS 30% That’s the rough percentage of tasks OpenAI says are broken https://openai.com/index/separating-signal-from-noise-coding-evaluations/ in a widely used test for measuring AI models' programming skills called SWE-Bench Pro. The finding may mean that one of the key assessments measuring what AI models can actually do is skewed. Test results like these inform release decisions, including how a model measures up against OpenAI’s Preparedness Framework safety criteria. Flawed tests can distort the real picture of an AI system’s capabilities. OpenAI pulled its endorsement of the coding benchmark and called for a more reliable option. SWE-Bench Pro was originally built in 2025 by researchers at Scale AI, the company that had been founded and run by Alexandr Wang, who runs Meta's Superintelligence Labs. It largely superseded SWE-Bench, an earlier coding benchmark, that had become "saturated," meaning too many AI models could ace the evaluation. “It’s just very disappointing that they took this long to find out,” Stuart Russel, a leading computer scientist and Professor of Computer Science at the University of California at Berkeley, told Fortune . “It’s just incorrectly formulated.” He said it speaks to a bigger issue about the benchmarks the industry is relying on to test AI capabilities. “Our benchmarking on everything is very, very suspect,” he said. “People talk about dataset contamination, where your training data ends up containing a lot of the test questions or something very similar to the test questions, and that’s a huge problem.” AI CALENDAR July 6-11: The International Conference on Machine Learning ICML , Seoul, South Korea. Aug. 4-6: Ai4 2026, Las Vegas. Nov. 16-17: Fortune 500 Innovation Forum, Detroit. Apply here https://conferences.fortune.com/event/fortune-500-innovation-forum-2026/HOME to attend. Dec. 6-12: Neural Information Processing Systems Neurips conference. Sydney, Australia. Dec. 7-8: Fortune Brainstorm AI, San Francisco. Apply here https://conferences.fortune.com/event/brainstorm-ai-2026/HOME to attend. Sign up for free https://www.fortune.com/newsletters/eye-on-ai?&itm source=fortune&itm medium=nl article tout&itm campaign=eye on ai .