Welcome back to quick,
Signal President Mredith Whittaker said he pose a “deep” security risk agents, speaking at the SXSW 2025 conference in Austin.
(Photo by Susanne Cordeiro/AFP through Getty Images)
For all the hype they have received, that agents – software that can browse online, operate many applications on your device, and perform tasks in your name – are “haunted” from the true risks of safety and intimacy, said signal president Meredith Whittaker at the SXSW conference this week. Whittaker said these systems were like placing your “brain in a jar”, as agents should be given access to sensitive data such as credit card information or phone contacts and also require access to use applications in your name. There are currently no coded models to send and receive those data, which must be processed in the cloud, she said.
Now let’s get into the titles.
Huge appearance
Subsequently Deepseek made waves in JanuaryAnother Chinese beginner of that called Butterfly Effect bubble on social media over the weekend By demonstrating her a He agent called manusfor which he claims he can perform tasks as resumptions of review, seeking property and analyzing shares, mainly on their own. The system it is reported to be Built on top of existing patterns of he Developed by other companies such as anthropic based in San Francisco, rather than trained from scratch. It doesn’t seem like there is a lot of substance for hype, as Many people who have tried the vehicle have reported issues with simple tasks Like ordering fast food or booking a flight.
Chip wars
Social media giant Meta has begun testing a new chip inside the housespecially designed to train models he, in an attempt to get away from his trust in the Nvidia device, Reuters reported. Meta, whose family of the models called Llama is said to have spent 600 million monthly users last year, is far from the only player to create his own silicon. Openai, Microsoft and Google have invested all billions in their chips.
Dilemma
Easy is easy to subtract the massive amounts of human Work that goes to develop artificial intelligence models. And while the models themselves become better and more capable of advanced tasks, so do the qualifications of those who train them. $ 14 billion worth Data Labeling Company’s scale, which helps companies like open train models, Google and Meta, is looking for more and more domain experts – including people with doctorate – with JSC Instead of transferring and hiring contractors abroad, stale reported. Change is part of CEO’s new philosophy “First America” and billionaire Alex Wang under Trump administration.
The one -week deal
Lila’s sciences, an initial construction based on Massachusetts one The he’s system for scientific discovery, has raised $ 200 million in seed funding, New York Times reported. Technology has been used to run experiments and create innovation in different fields.
Also obvious: Servicenow has won the enterprise enterprise of Moveworks for about $ 3 billion, Reuters reported.
Leak
Tesla It should no longer be seen as a manufacturer of electric cars. it’s A company he – if you trust Ceo Elon Musk. His faith is associated with a unique database: Petabytes of video harvested by company cars As tesla clients log Millions of moving miles worldwide.
In theory, all this data in the real world It is exactly what Tesla needs to train her cars to operate without any human help, a goal that is essential to Musk’s vision for the future of Tesla. But there is a problem: These data are not necessarily as useful as Musk claims. Some of them are not useful at all.
Building a car that can drive a car, as well as a man, is a challenge in a dramatic way different than building a chatbot processing of natural language like chatgt, which was trained in billions of words written by the Internet. While the goal with chatgpt and competitive systems like Xai’s Grok is to use model recognition to provide reliable information and answers to questions, results often fall in embarrassing ways. But If the one who controls a screw vehicle, people can die.
Driving a car is a very different proposal with many more variables – driving conditions, weather, construction, change of traffic models, how other vehicles move. The successful clash of all those variables, and being ready to respond to unexpected developments, is at the core of the autonomous car. His endless video training of people moving down highways does not do much to help him learn how to handle what it takes most: cases of ends that cause clashes or other dangerous scenarios.
Yann Lecun, the chief of scientists of his drawbacks and a computer science professor at the University of New York, is also not convinced that Tesla’s data give her a competitive advantage.
“The impact of the data is generally overrated: while getting more data, the performance improves, but there are reduced returns,” he said. “A duplication of data volume brings marginal improvements that are still far from human credibility.” Even with massive amounts of data, no company has developed so -called level 5 autonomy, the point at which a vehicle can drive itself in all the circumstances a person can.
“However Anydo 17-year-old can learn to drive in about 20 hours of practice,“Said Lecun.” This tells you that he is missing something great in their ability to understand the world and learn from limited amounts of data or rehearsals. “
Large amounts of camera data are useful but She doesn’t immediately make Tesla a market leader of him. “Having access to unique data sources is certainly a kind of advantage,” said computer scientist Alex Ratner, Snorkel’s CEO, who makes software to help automate raw data.
Read the full story in stale.
Model behavior
it searching engines Fail wretched in the citation of news, According to a new study by Columbia Journalism Review (CJR), who climbed experts from 10 case articles in chatbots and asked them to identify the publisher, article and date of publication. CJR ran sixteen hundred questions in eight systems he including confusion, chatgpt, twins and others. Grok was particularly badreplying 94% of questions incorrectly. Even when they quote resources, people are not really clicking. Research engines to run 96% less traffic than the old school, no Google, as we reported last week.