Tag: ai

OpenAI co-founder John Schulman has left Anthropic after less than a year

Posted on by faz_business


Less than a year into his tenure at the company, OpenAI co-founder John Schulman is leaving Anthropic. The startup confirmed Schulman’s departure after The Information, Reuters and other publications reported on the exit.

“We are sad to see John go but fully support his decision to pursue new opportunities and wish him all the very best,” said Jared Kaplan, Anthropic’s chief science officer, in a statement the company shared with Engadget. Schulman left OpenAI last August alongside Peter Deng, the company’s former vice-president of consumer product. Schulman is considered one of the original architects of ChatGPT.

Following his departure from OpenAI, Schulman said he was joining Anthropic to focus on AI alignment — the process of making machine learning models safe to use — and a desire to return “to more hands-on technical work.” Schulman hasn’t publicly said why he decided to leave Anthropic, nor what he plans to do next. His X profile still says he “recently joined” Anthropic.

If you buy something through a link in this article, we may earn commission.

DeepSeek founder Liang Wenfeng receives a hero’s welcome back home

Posted on by faz_business


DeepSeek founder Lian Wenfeng is being hailed as a hero in the southern Chinese province of Guangdong, where he grew up and reportedly returned for the Lunar New Year, joined by bodyguards.

Wenfeng—who, at 40, is already a billionaire due to his hedge fund, High-Flyer—is apparently even more beloved by locals following DeepSeek’s breakthrough research, which demonstrated that strong AI models could be built with fewer and less-powerful Nvidia chips. The finding has significant implications, particularly in China, where access to the highest-end chips has been restricted.

Residents of his native community tell the Financial Times that Wenfeng was a “top student” who loved comic books, was a math whiz, and came from a family of educators. According to one local, Wenfeng also played some football. “We all grew up in this village,” the resident tells the FT. “We’re very proud of him.”

Despite his rising star, Wenfeng has largely avoided public attention, frustrating those eager to learn more about him. Given the fate of high-flying Chinese CEOs like Jack Ma and Pony Ma — who faced government scrutiny after gaining a little too much visibility — don’t be surprised if Wenfeng chooses to stay in the background.

MLCommons and Hugging Face team up to release massive speech data set for AI research

Posted on by faz_business


MLCommons, a nonprofit AI safety working group, has teamed up with AI dev platform Hugging Face to release one of the world’s largest collections of public domain voice recordings for AI research.

The data set, called Unsupervised People’s Speech, contains more than a million hours of audio spanning at least 89 different languages. MLCommons says it was motivated to create it by a desire to support R&D in “various areas of speech technology.”

“Supporting broader natural language processing research for languages other than English helps bring communication technologies to more people globally,” the organization wrote in a blog post Thursday. “We anticipate several avenues for the research community to continue to build and develop, especially in the areas of improving low-resource language speech models, enhanced speech recognition across different accents and dialects, and novel applications in speech synthesis.”

It’s an admirable goal, to be sure. But AI data sets like Unsupervised People’s Speech can carry risks for the researchers who choose to use them.

Biased data is one of those risks. The recordings in Unsupervised People’s Speech came from Archive.org, the nonprofit perhaps best known for the Wayback Machine web archival tool. Because many of Archive.org’s contributors are English-speaking — and American — almost all of the recordings in Unsupervised People’s Speech are in American-accented English, per the readme on the official project page.

That means that, without careful filtering, AI systems like speech recognition and voice synthesizer models trained on Unsupervised People’s Speech could exhibit some of the same prejudices. They might, for example, struggle to transcribe English spoken by a non-native speaker, or have trouble generating synthetic voices in languages other than English.

Unsupervised People’s Speech might also contain recordings from people unaware that their voices are being used for AI research purposes — including commercial applications. While MLCommons says that all recordings in the data set are public domain or available under Creative Commons licenses, there’s the possibility mistakes were made.

According to an MIT analysis, hundreds of publicly available AI training data sets lack licensing information and contain errors. Creator advocates including Ed Newton-Rex, the CEO of AI ethics-focused nonprofit Fairly Trained, have made the case that creators shouldn’t be required to “opt out” of AI data sets because of the onerous burden opting out imposes on these creators.

“Many creators (e.g. Squarespace users) have no meaningful way of opting out,” Newton-Rex wrote in a post on X last June. “For creators who can opt out, there are multiple overlapping opt-out methods, which are (1) incredibly confusing and (2) woefully incomplete in their coverage. Even if a perfect universal opt-out existed, it would be hugely unfair to put the opt-out burden on creators, given that generative AI uses their work to compete with them — many would simply not realize they could opt out.”

MLCommons says that it’s committed to updating, maintaining, and improving the quality of Unsupervised People’s Speech. But given the potential flaws, it’d behoove developers to exercise serious caution.

DeepSeek: Everything you need to know about the AI chatbot app

Posted on by faz_business


DeepSeek has gone viral.

Chinese AI lab DeepSeek broke into the mainstream consciousness this week after its chatbot app rose to the top of the Apple App Store charts. DeepSeek’s AI models, which were trained using compute-efficient techniques, have led Wall Street analysts — and technologists — to question whether the U.S. can maintain its lead in the AI race and whether the demand for AI chips will sustain.

But where did DeepSeek come from, and how did it rise to international fame so quickly?

DeepSeek’s trader origins

DeepSeek is backed by High-Flyer Capital Management, a Chinese quantitative hedge fund that uses AI to inform its trading decisions.

AI enthusiast Liang Wenfeng co-founded High-Flyer in 2015. Wenfeng, who reportedly began dabbling in trading while a student at Zhejiang University, launched High-Flyer Capital Management as a hedge fund in 2019 focused on developing and deploying AI algorithms.

In 2023, High-Flyer started DeepSeek as a lab dedicated to researching AI tools separate from its financial business. With High-Flyer as one of its investors, the lab spun off into its own company, also called DeepSeek.

From day one, DeepSeek built its own datacenter clusters for model training. But like other AI companies in China, DeepSeek has been affected by U.S. export bans on hardware. To train one of its more recent models, the company was forced to use Nvidia H800 chips, a less-powerful version of a chip, the H100, available to U.S. companies.

DeepSeek’s technical team is said to skew young. The company reportedly aggressively recruits doctorate AI researchers from top Chinese universities. DeepSeek also hires people without any computer science background to help its tech better understand a wide range of subjects, per The New York Times.

DeepSeek’s strong models

DeepSeek unveiled its first set of models — DeepSeek Coder, DeepSeek LLM, and DeepSeek Chat — in November 2023. But it wasn’t until last spring, when the startup released its next-gen DeepSeek-V2 family of models, that the AI industry started to take notice.

DeepSeek-V2, a general-purpose text- and image-analyzing system, performed well in various AI benchmarks — and was far cheaper to run than comparable models at the time. It forced DeepSeek’s domestic competition, including ByteDance and Alibaba, to cut the usage prices for some of their models, and make others completely free.

DeepSeek-V3, launched in December 2024, only added to DeepSeek’s notoriety.

According to DeepSeek’s internal benchmark testing, DeepSeek V3 outperforms both downloadable, openly available models like Meta’s Llama and “closed” models that can only be accessed through an API, like OpenAI’s GPT-4o.

Equally impressive is DeepSeek’s R1 “reasoning” model. Released in January, DeepSeek claims R1 performs as well as OpenAI’s o1 model on key benchmarks.

Being a reasoning model, R1 effectively fact-checks itself, which helps it to avoid some of the pitfalls that normally trip up models. Reasoning models take a little longer — usually seconds to minutes longer — to arrive at solutions compared to a typical non-reasoning model. The upside is that they tend to be more reliable in domains such as physics, science, and math.

There is a downside to R1, DeepSeek V3, and DeepSeek’s other models, however. Being Chinese-developed AI, they’re subject to benchmarking by China’s internet regulator to ensure that its responses “embody core socialist values.” In DeepSeek’s chatbot app, for example, R1 won’t answer questions about Tiananmen Square or Taiwan’s autonomy.

A disruptive approach

If DeepSeek has a business model, it’s not clear what that model is, exactly. The company prices its products and services well below market value — and gives others away for free.

The way DeepSeek tells it, efficiency breakthroughs have enabled it to maintain extreme cost competitiveness. Some experts dispute the figures the company has supplied, however.

Whatever the case may be, developers have taken to DeepSeek’s models, which aren’t open source as the phrase is commonly understood but are available under permissive licenses that allow for commercial use. According to Clem Delangue, the CEO of Hugging Face, one of the platforms hosting DeepSeek’s models, developers on Hugging Face have created over 500 “derivative” models of R1 that have racked up 2.5 million downloads combined.

DeepSeek’s success against larger and more established rivals has been described as “upending AI” and ushering in “a new era of AI brinkmanship.” The company’s success was at least in part responsible for causing Nvidia’s stock price to drop by 18% on Monday, and for eliciting a public response from OpenAI CEO Sam Altman.

As for what DeepSeek’s future might hold, it’s not clear. Improved models are a given. But the U.S. government appears to be growing wary of what it perceives as harmful foreign influence.

TechCrunch has an AI-focused newsletter! Sign up here to get it in your inbox every Wednesday.

OpenAI launches Operator: an AI assistant that handles web tasks

Posted on by faz_business


OpenAI on website on smartphone stock photo (2)

Edgar Cervantes / Android Authority

TL;DR

  • OpenAI has launched Operator, an AI agent that performs browser-based tasks like filling out forms, booking services, and placing orders.
  • Operator is currently only available as a research preview for Pro users in the US.
  • OpenAI plans to expand access, refine Operator through user feedback, and eventually integrate it into ChatGPT.

AI agents have recently become the talk of the tech world, so it was only a matter of time before we saw the biggest name in AI get in on the action. OpenAI just introduced Operator, an AI agent capable of performing tasks directly within a browser. Your automated PA has arrived.

According to the press release on OpenAI’s website, Operator is currently only available as a research preview for Pro users in the US. It uses a new model called the Computer-Using Agent (CUA) to handle tasks like filling out forms, booking services, and placing orders online.

Operator combines GPT-4o’s vision capabilities with advanced reasoning and tools for navigating websites. It interacts with pages by clicking, typing, and scrolling — just like you or I would. Tasks can be assigned with simple instructions, and Operator returns control to you when sensitive input like passwords or payment details is required.

You can also customize workflows for specific sites or save prompts for frequent tasks, such as restocking groceries or managing multiple orders simultaneously. You can see a practical demonstration in the preview video below.

This initial release as a research preview is being used to gather feedback and refine the tool. OpenAI plans to expand access to other user tiers and eventually integrate Operator into ChatGPT. Collaborations with companies like Instacart and public-sector initiatives are also being tested to explore practical use cases.

Our natural instinct is to wonder how this might immediately go wrong, so OpenAI is quick to point out that Operator is designed with safeguards to protect user data and ensure safe interactions. It pauses for confirmation before completing major actions and avoids handling sensitive tasks like financial transactions. Users can manage privacy settings, delete browsing data, and opt out of data collection. OpenAI accepts it’s still a work in progress and may struggle with more complex interfaces, hence this research phase.

OpenAI Pro users in the US can give Operator a try now at operator.chatgpt.com. Just describe a task, and Operator will handle the rest, with you retaining the option to take over if needed.

Got a tip? Talk to us! Email our staff at news@androidauthority.com. You can stay anonymous or get credit for the info, it’s your choice.

ChatGPT Search can be tricked into misleading users, new research reveals

Posted on by faz_business


ChatGPT Search, an AI-powered search engine that went live this month, can be fooled into generating completely misleading summaries, U.K. newspaper The Guardian has found.

ChatGPT’s search feature is meant to make browsing faster by doing things like summarizing a web page’s product reviews. But The Guardian found it could get ChatGPT to ignore negative reviews and generate “entirely positive” summaries by inserting hidden text into websites it created. ChatGPT Search could also be made to spit out malicious code using this method.

Such hidden text attacks are a well-known risk for large language models, but this appears to be the first time it’s demonstrated on a live AI-powered search product. Google, the leader in search, has more experience dealing with similar problems, The Guardian noted.

OpenAI didn’t comment about this specific incident when TechCrunch reached out, but said it uses a variety of methods to block malicious websites and is continually improving.

TechCrunch has an AI-focused newsletter! Sign up here to get it in your inbox every Wednesday.

Elon Musk’s xAI lands $6B in new cash to fuel AI ambitions

Posted on by faz_business


Updated December 25, 12:21 p.m. Pacific: Added details of xAI’s valuation and Kingdom Holdings’ contribution.

xAI, Elon Musk’s AI company, has raised $6 billion in a Series C financing round.

The company announced this week that Andreessen Horowitz , Blackrock, Fidelity, Lightspeed, MGX, Morgan Stanley, OIA, QIA, Sequoia Capital, Valor Equity Partners, Vy Capital, Nvidia, AMD, and others participated.

Kingdom Holdings, the Saudi conglomerate holding company, invested roughly $400 million in the round, according to a public filing. The filing also revealed that xAI is now valued at $45 billion, close to double its previous valuation.

The new cash brings xAI’s total raised to $12 billion, adding to the $6 billion tranche xAI raised in May.

According to the Financial Times, only investors who’d backed xAI in its previous fundraising round were permitted to participate in this one. Reportedly, investors who helped finance Musk’s Twitter acquisition were given access to up to 25% of xAI’s shares.

“xAI’s most powerful model yet … is currently training and we are now focused on launching innovative new consumer and enterprise products,” xAI said in a statement. “The funds from this financing round will be used to further accelerate our advanced infrastructure, ship groundbreaking products … and accelerate … research and development.”

Ramping up AI

Musk formed xAI last year. Soon after, the company released Grok, a flagship generative AI model that now powers a number of features on X, including a chatbot accessible to X Premium subscribers and free users in some regions.

Grok has what Musk has described as “a rebellious streak” — a willingness to answer “spicy questions that are rejected by most other AI systems.” Told to be vulgar, for example, Grok will happily oblige, spewing profanities and colorful language you won’t hear from ChatGPT.

Musk has derided ChatGPT and other AI systems for being too “woke” and “politically correct,” despite Grok’s own unwillingness to cross certain boundaries and hedge on political subjects. He’s also referred to Grok as “maximally truth-seeking” and less biased than competing models, although there’s evidence to suggest that Grok leans to the left.

Over the past year, Grok has become increasingly ingrained in X, the social network formerly known as Twitter. At launch, Grok was only available to X users — and developers skilled enough to get the “open source” edition up and running.

Thanks to an integration with xAI’s in-house image generation model, Aurora, Grok can generate images on X (without guardrails, controversially). The model can analyze images as well, and summarize news and trending events — imperfectly, mind.

Reports indicate that Grok may handle even more X functions in the future, from enhancing X’s search capabilities and account bios to helping with post analytics and reply settings. X recently got a “Grok button” designed to help users discover “relevant context” and dive deeper into trending discussions and real-time events.

xAI is sprinting to catch up to formidable competitors like OpenAI and Anthropic in the generative AI race. The company launched an API in October, allowing customers to build Grok into third-party apps, platforms, and services. And it just rolled out a standalone Grok iOS app to a test audience.

Musk asserts that it hasn’t been a fair fight.

In a lawsuit filed against OpenAI and Microsoft, OpenAI’s close collaborator, attorneys for Musk accuse OpenAI of “actively trying to eliminate competitors” like xAI by “extracting promises from investors not to fund them.” OpenAI, Musk’s counsel says, also unfairly benefits from Microsoft’s infrastructure and expertise in what the attorneys describe as a “de facto merger.”

Yet Musk often says that X’s data gives xAI a leg up compared to rivals. Last month, X changed its privacy policy to allow third parties, including xAI, to train models on X posts.

Musk, it’s worth noting, was one of the original founders of OpenAI, and left the company in 2018 after disagreements over its direction. He’s argued in previous suits that OpenAI profited from his early involvement yet reneged on its nonprofit pledge to make the fruits of its AI research available to all.

OpenAI, unsurprisingly, disagrees with Musk’s interpretation of events. In a mid-December press release, the company characterized Musk’s lawsuit as misleading, baseless, and a case of sour grapes.

An xAI ecosystem

xAI has outlined a vision according to which its models would be trained on data from Musk’s various companies, including Tesla and SpaceX, and the models could then improve technology across those companies. xAI is already powering customer support for SpaceX’s Starlink internet service, according to The Wall Street Journal, and the startup is said to be in talks with Tesla to provide R&D in exchange for some of the carmaker’s revenue.

Tesla shareholders, for one, object to these plans. Several have sued Musk over his decision to start xAI, arguing that Musk has diverted both talent and resources from Tesla to what’s essentially a competing venture.

Nevertheless, the deals — and xAI’s developer and consumer-facing products — have driven xAI’s revenue to around $100 million a year. For comparison, Anthropic is reportedly on pace to generate $1 billion in revenue this year, and OpenAI is targeting $4 billion by the end of 2024.

Musk said this summer that xAI is training the next generation of Grok models at its Memphis data center, which was apparently built in just 122 days and is currently powered partly by portable diesel generators. The company hopes to upgrade the server farm, which contains 100,000 Nvidia GPUs, next year; in a press release, xAI said it plans to fully double that number. (Because of their ability to perform many calculations in parallel, GPUs are the favored chips for training and running models.)

In November, xAI won approval from the regional power authority in Memphis for 150MW of additional power — enough to power roughly 100,000 homes. To win the agency over, xAI pledged to improve the quality of the city’s drinking water and provide the Memphis grid with discounted Tesla-manufactured batteries. But some residents criticized the move, arguing it would strain the grid and worsen the area’s air quality.

Tesla is also expected to use the upgraded data center to improve its autonomous driving technologies.

xAI has expanded quite rapidly from an operations standpoint in the year since its founding, growing from just a dozen employees in March 2023 to over 100 today. In October, the startup moved into OpenAI’s old corporate offices in San Francisco’s Mission neighborhood.

xAI has reportedly told investors it plans to raise more money next year.

It won’t be the only AI lab raising immense cash. Anthropic recently secured $4 billion from Amazon, bringing its total raised to $13.7 billion, while OpenAI raised $6.6 billion in October to grow its war chest to $17.9 billion.

Megadeals like OpenAI’s and Anthropic’s drove AI venture capital activity to $31.1 billion across over 2,000 deals in Q3 2024, per PitchBook data.

TechCrunch has an AI-focused newsletter! Sign up here to get it in your inbox every Wednesday.

The DOJ wants a Perplexity executive to testify in its Google antitrust case

Posted on by faz_business


A U.S. court ruled in August that Google has a search monopoly, and while Google appeals, the Justice Department is figuring out what kind of potential penalties to impose — like breaking off Chrome. 

As part of this process, the DOJ wants to call on a specific witness, according to a recent court filing: Dmitry Shevelenko, chief business officer of Perplexity, an AI search provider most recently valued at $9 billion, per Reuters.  

Perplexity and other generative AI tools like OpenAI’s ChatGPT Search have emerged as a potential replacement for searching the internet, as they can offer direct answers to complicated queries (albeit, sometimes with made-up or inaccurate information). Google has responded to the threat with its own AI search tools, such as AI Overviews, which provide AI-generated answers above search results.

The DOJ wants to ask Shevelenko about “generative AI’s relationship with Search Access Points, distribution, barriers to entry and expansion, and data sharing.”

“Search access points” is a term the DOJ uses to describe things like Google Chrome — places where people go to search the internet.

While the filing doesn’t spell out exactly why the DOJ wants to ask Perplexity about these topics, it could help its argument that Google monopolizes the search business and closes out potential competitors, and thus deserves stronger penalties. 

TechCrunch asked Perplexity whether it has agreed to have its executive testify and for its thoughts on the antitrust case. Perplexity didn’t respond immediately to a request for comment, and neither did Google. 

Perplexity is effectively caught in the middle of the dispute, as both sides want information from it that could help their cases. Google subpoenaed Perplexity in October for company documents to make its own case that it has viable competition in the search field. (Google subpoenaed Microsoft and OpenAI as well.)

However, Perplexity has yet to provide “a single document” to Google as of December 11, the tech giant lamented in a court filing, claiming that there is “no conceivable justification for further delay” after two months of waiting.

For its part, Perplexity says in the filing that it has already agreed to fulfill 12 out of Google’s 14 document requests but is “still evaluating the burden associated with collecting such a potentially expansive universe of documents.”

Perplexity also says that while it has agreed to provide copies of licensing agreements “related to AI training,” Google wants all of Perplexity’s licensing agreements and that it has asked Google to “meet and confer” about this. 

X gains a faster Grok model and a new ‘Grok button’

Posted on by faz_business


XAI, Elon Musk’s AI company, may be embroiled in an escalating lawsuit with OpenAI. But that’s not stopping it from shipping new products — on a Friday night, no less.

This evening, xAI revealed that it has begun to roll out an upgraded version of its flagship Grok 2 chatbot model to all users on X, the social network formerly known as Twitter. (X, which Musk also owns, often serves as a testing ground of sorts for Grok.) The enhanced Grok is “three times faster,” xAI claims in a blog post, and offers “improved accuracy, instruction-following, and multi-lingual capabilities.”

Free users can only ask Grok ten questions every two hours. Subscribers to X’s Premium and Premium+ plans get higher usage limits.

XAI also announced tonight the addition of a “Grok button” to X, which the company says is designed to help users discover “relevant context, understand real-time events, and dive deeper into trending discussions.”

xAI Grok
The new Grok button. Image Credits:xAI

And the startup said it’s making several changes to its enterprise API.

XAI’s API has a pair of new Grok models with better efficiency and multilingual performance, xAI says. As a result of the efficiency gains, pricing has been reduced from $5 per million input tokens (~750,000 words) or $15 per million output tokens to $2 per million input tokens and $10 per million output tokens.

In the coming weeks, xAI’s image generation model, Aurora, will come to the API as well, xAI says. Aurora, a largely unfiltered image AI, was released on X this month in the Grok chatbot experience.

ChatGPT now understands real-time video, seven months after OpenAI first demoed it

Posted on by faz_business


OpenAI has finally released the real-time video capabilities for ChatGPT that it demoed nearly seven months ago.

On Thursday during a livestream, the company said that Advanced Voice Mode, its human-like conversational feature for ChatGPT, is getting vision. Using the ChatGPT app, users subscribed to ChatGPT Plus, Team, and Pro can point their phones at objects and have ChatGPT respond in near-real-time.

Advanced Voice Mode with vision can also understand what’s on a device’s screen, via screen sharing. It can explain various settings menus, for example, or give suggestions on a math problem.

To access Advanced Voice Mode with vision, tap the voice icon next to the ChatGPT chat bar, then tap the video icon at the bottom left, which will start video. To screen-share, tap the three-dot menu and select “Share Screen.”

The rollout of Advanced Voice Mode with vision will start today, OpenAI says, and wrap up in the next week. But not all users will get access. OpenAI says that ChatGPT Enterprise and Edu subscribers won’t get the feature until January, and that it has no timeline for ChatGPT users in the EU, Switzerland, Iceland, Norway, and Liechtenstein.

In a recent demo on CNN’s 60 Minutes, OpenAI president Greg Brockman had Advanced Voice Mode with vision quiz Anderson Cooper on his anatomy skills. As Cooper drew body parts on a blackboard, ChatGPT could “understand” what he was drawing.

OpenAI unveils ChatGPT Advanced Voice Mode in May 2024.
Image Credits:OpenAI

“The location is spot on,” the assistant said. “The brain is right there in the head. As for the shape, it’s a good start. The brain is more of an oval.”

In that same demo, Advanced Voice Mode with vision made a mistake on a geometry problem, however — suggesting that it’s prone to hallucinating.

Advanced Voice Mode with vision has been delayed multiple times — reportedly in part because OpenAI announced the feature far before it was production-ready. In April, OpenAI promised that Advanced Voice Mode would roll out to users “within a few weeks.” Months later, the company said it needed more time.

When Advanced Voice Mode finally arrived in early fall for some ChatGPT users, it lacked the visual analysis component. In the lead-up to today’s launch, OpenAI has focused its attention on bringing the voice-only Advanced Voice Mode experience to additional platforms and users in the EU.

In addition to Advance Voice Mode with vision, OpenAI today launched a festive “Santa Mode,” which adds Santa’s voice as a preset voice in ChatGPT. Users can find it by tapping or clicking the snowflake icon in the ChatGPT app next to the prompt bar.