by Welcome back to In the loopThe new two -week newsletter of the time over the world of AI. We publish rates as stories about time.com and as e -mails.
If you read this in your browser, you can have the next delivered directly to your inbox.
What to know:
Could the coding with AI slow down?
In recent years, AI has completely changed the world of software engineering. Writing your own code (at least from scratch to new) has become picturesque. With tools such as cursor and copilot, human developers can now mares the AI to write code for them. The human role now is to understand what the models ask for the best results and to iron the inevitable problems.
Conventional wisdom states that this has accelerated the software engineering considerably. But does it have it? A new study by METR, which was published last week, measures to what extent AI accelerates the work of experienced software developers. The results were very unexpected.
What the study showed – METR has measured the speed of 16 developers who worked on complex software projects with and without AI support. After completing their tasks, the developers estimated that access to Ki had accelerated their work on average by 20%. In fact, the measurements showed that the AI had slowed it by about 20%. The results were surprisingly surprised in the KI community. “I was pretty skeptical that it was worth doing this study because I thought that we would obviously see a significant speed,” wrote David Rein, an employee of Metr, in a contribution to X.
Why did that happen? – The simple technical answer seems to be: While today’s LLMs can code well, they are often not good enough to answer exactly what a developer wants and to answer perfectly in a recording. This means that you can need a lot back and forth, which may take longer than if you have only written the code yourself. But participants in the study also offered several other human hypotheses. “LLMS are a large dopamine link-up button that fills your problem,” wrote Quentin Anthony, one of the 16 coders who took part in the experiment. “Press the button that has a 1% chance of repairing everything? It is much more pleasant than the strenuous alternative.” (It is also easy to scroll social media while waiting for your LLM to generate an answer, he added.)
What it means for Ai – The authors of the study asked the readers not to generalize themselves too far from the results. On the one hand, the study only measures the effects of LLMS on experienced coders, not new ones that could benefit more from their help. And developers still learn how to get the best out of LLMS, which is relatively new tools with strange peculiarities. They found that other metr research show the duration of the software tasks that AI is able to double every seven months. This is not the case if today’s AI affects productivity that may not be the case tomorrow.
Whom to know:
Jensen Huang, CEO of Nvidia
Huang is in the news today after he has announced CNN that the US government does not “worry” about the possibility of the Chinese military to use the market leading AI chips that its company produces Nvidia. “You just can’t rely on it,” he said. “Of course it could be limited at any time.”
Pour away – Huang argued against guidelines in which the USA severely restricted the export of graphics processing units or GPUs to China in order to impair the military skills of Beijing and the AI progress. NVIDIA claims that these guidelines have only suggested China to build up their own competing chip supply chain while they violated US companies and thus the US economy.
Selfish argument – Huang would of course say that a company CEO has lost billions because it was blocked by the sale of his most advanced chips on the Chinese market. He tried to convince President Donald Trump of his positions in a recent meeting in the White House, reported Bloomberg.
Actually… The Chinese military uses Nvidia Chips, according to the Georgetown Center for Security and emerging technology, which analyzed 66,000 military purchasing files to come to this conclusion. A large black market has also corresponded to smuggling Nvidia chips to China since export controls were set up, the New York Times reported last year.
AI in action
The AI assistant of Anthropic, Claude, changes the way the company’s scientists keep pace with the thousands of scientific literature that were published every day in their area.
Instead of reading papers, many anthropic researchers simply invite them to Claude and chat with the assistant to distill the main results. “I changed my habits of how I read papers,” said Jan Leike, a senior alignment researcher at Anthropic, at the beginning of this year. “Where now, usually I only put them in Claude and question: Can you explain?”
To be clear, Leike adds, sometimes Claude misunderstood important things. “But if I just cried the paper, I will sometimes misunderstand important things,” says Leike. “I think the greater effect here is that I can read a lot more papers than before.” That, he says, has a positive effect on his productivity. “There is a lot of time when you read papers, just to find out whether the paper is relevant for what you try,” he says. “And this part is so fast 1752605637You can only concentrate on the papers that are actually important. “
What we read
The AGI fight by Microsoft and Openai is larger than a contract – by Steven Levy in Wired
Steven Levy goes deep into the “Agi” clause in the contract between Openaai and Microsoft, which could decide the fate of their billion dollar partnership. It is worth understanding how both sides think about the definition of Agi. They could get worse than Levy’s own description: “A technology that Sauron’s strength ring looks like a plastic plastic Doodad from Centore.”
