Welcome back to The Prompt,
AI search startup Perplexity wants to acquire TikTok and “rebuild” its algorithm “without creating a monopoly,” the company said in a blog post. The company proposed merging the Perplexity’s search capabilities with TikTok’s video recommendation algorithm. The company, reportedly raising funding at a $18 billion valuation, also wants to make TikTok’s For You feed more transparent by open sourcing it. It also claims it would develop and store these systems in American data centers. But the line to buy TikTok is long–with powerful companies like Microsoft and Oracle currently at the front.
Now let’s get into the headlines.
POLITICS
Under the Trump administration, top AI companies have changed their tunes from being pro-regulation and asking for more safety checks on their frontier models to lobbying for fewer restrictions on the advancement of AI, according to The New York Times. In a bid to influence the new government’s policies on AI, Google, Meta, OpenAI and Microsoft have lobbied that their use of copyrighted works and publicly available data to train models is legal and falls under fair use. Meanwhile, chip makers are also encouraging the Trump administration to rethink policies that curb chip exports to several countries–including China, according to Bloomberg.
ETHICS + LAW
AI models are being tasked with more consequential work, but they still struggle with prejudices. A Wired investigation that evaluated 250 videos generated by OpenAI’s video generation tool Sora, found that it amplifies biases and sexist and racist stereotypes. For example, when asked to generate videos of pilots and CEOs, Sora depicted men, while prompts for videos showing flight attendants and receptionists would produce videos that largely featured women.
DATA DILEMMAS
To compete with OpenAI’s ChatGPT, Meta wanted to train its Llama 3 large language model, and do it fast. So the social media giant turned to Library Genesis, a shadow library of pirated books that contains 81 million research papers and 7.5 million books, according to unsealed documents that are part of a copyright lawsuit against Meta filed by a group of authors. CEO Mark Zuckerberg is said to have signed off on the use of pirated books to train the company’s AI, according to The Atlantic, even though multiple employees acknowledged that using the torrent books posed a legal risk.
AI DEAL OF THE WEEK
Distributed Spectrum, a startup building AI-powered sensors to help the military detect radio frequencies from adversaries, has raised $25 million in funding from Conviction and Shield Capital, Forbes reported. Founded by three Harvard students, the startup already has $7 million in contracts from the Department of Defense.
DEEP DIVE
DeepSeek released an upgrade to its large language model this week, featuring “significant improvements” over its predecessor as the China-based startup appears to escalate its rivalry with OpenAI and other U.S. artificial intelligence firms, after an earlier release rattled global tech stocks.
DeepSeek launched an upgrade to its V3 large language model on Hugging Face on Tuesday, which the startup claims includes advancements in reasoning and coding capabilities over its earlier V3 model. The update included improvements across several benchmark tests for the language model, DeepSeek said, in addition to upgrades to front-end web development, Chinese writing proficiency—with new features like “interactive rewriting” and Chinese search capabilities like “enhanced report analysis.”
DeepSeek is an AI startup founded by Chinese entrepreneur Liang Wenfeng in 2023. Liang started accumulating thousands of Nvidia graphics processors for an unnamed AI project in 2021, just before the Biden administration restricted trade of those chips to China. He called for China to “gradually transition” to being a contributor in the AI industry, “rather than continuing to ride on the coattails of others.” DeepSeek has claimed its products are more efficient and cost less to train and develop compared to similar products made by OpenAI and Meta. The company said training one of its latest models cost about $5.6 million, far less than the $100 million to $1 billion an AI executive estimated for costs to build a model.
A bipartisan bill banning Deepseek from federal devices was introduced in February, after a report linked the company’s chatbot to a banned Chinese state-run telecommunications company. It’s not immediately clear whether the bill will be supported by other lawmakers. The Trump administration is also mulling a ban over national security concerns, a person familiar with the matter told Reuters.
Some U.S. agencies have already restricted access to DeepSeek, including the U.S. Navy, the Defense Department, the Commerce Department and NASA, among others. The House’s Chief Administrative Officer reportedly claimed “threat actors are already exploiting DeepSeek to deliver malicious software and infect devices,” as some congressional offices have been warned not to use DeepSeek.
Read the full story on Forbes.
WEEKLY DEMO
OpenAI is giving ChatGPT new image generation powers. The company claims its GPT-4o model is better at following complicated prompts and producing images without garbled text, potentially making it useful to create diagrams, logos and infographics.
In other news: OpenAI has released its first research studies on how using ChatGPT affects people’s emotional well being based on data from 40 million interactions and interviews of some 4000 people. Among other things, it found that female participants who interacted with ChatGPT were less likely to socialize with people than their male counterparts and that those who used ChatGPT’s voice mode with a gender not their own tended to report higher levels of loneliness.
MODEL BEHAVIOR
AI might be getting better at coding and solving college-level math problems. But when it comes to chemistry, it’s still behind. When asked to illustrate the chemical structure for caffeine, AI chatbots including xAI’s Grok, Google’s Gemini and OpenAI’s ChatGPT failed miserably. In some cases, the AI model just couldn’t create molecular structures, in others it labelled chemicals incorrectly and in one case, Google’s Gemini made up a nonexistent symbol out of thin air. “Grok supplied one image that was more like a spiderweb than anything resembling caffeine,” Rowan Walrath at Chemical and Engineering News wrote.
