Category: Top News

IBL News | New York

YouTube integrated this month for its Shorts creators Google DeepMind’s latest text-to-video model generator, Veo 2.

Veo 2, which is Google’s response to OpenAI’s Sora, allows users to generate AI backgrounds for their Shorts through a feature called Dream Screen. [See an example below]

To use Veo 2 in YouTube Shorts, creators can open the Shorts camera, select Green Screen, and then navigate to Dream Screen, where they can input a text prompt to generate a video.

YouTube uses a watermark tool called SynthID to indicate that videos are generated using AI.

YouTube is also launching another capability powered by Veo 2, which allows users to generate standalone video clips via text prompts that can be added to any Shorts.

To create a clip to add to any Short, users can open the Shorts camera, tap Add, and Create at the top. After inputting their prompt, they select their image, tap Create video, and choose their desired length.

These features were available in the U.S., Canada, Australia, and New Zealand. YouTube plans to expand access later.

IBL News | New York

Researchers at Stanford and the University of Washington said in a paper released this month that they were able to train an AI reasoning model called s1, which performed similarly to OpenAI’s o1 and DeepSeek’s R1 on math and coding.

The s1 model, along with the data and code, is available on GitHub. According to the researchers, its training costs less than $50 in cloud computing credits.

This team started with an off-the-shelf base model and then fine-tuned it through distillation, a process for extracting the “reasoning” capabilities from another AI model by training on its answers.

The model was distilled from Gemini 2.0 Flash Thinking Experimental, offered for free via the Google AI Studio platform.

Distillation is the same approach Berkeley researchers used to create an AI reasoning model for around $450 last month.

OpenAI has accused DeepSeek of improperly harvesting data from its API for model distillation.

Distillation is a suitable method for cheaply re-creating an AI model’s capabilities, but it doesn’t create new AI models.

The s1 paper suggested that reasoning models can be distilled with a relatively small dataset using supervised fine-tuning (SFT), in which an AI model is explicitly instructed to mimic certain behaviors in a dataset.

More specifically, s1 was based on a small, free AI model from Alibaba-owned Chinese AI lab Qwen. To train s1, the researchers created a dataset of just 1,000 carefully curated questions paired with answers to those questions and the “thinking” process behind each answer from Google’s Gemini 2.0 Flash Thinking Experimental.

After training s1, which took less than 30 minutes using 16 Nvidia H100 GPUs, s1 achieved strong performance on specific AI benchmarks.

Per the paper, researchers used a nifty trick to get s1 to double-check its work and extend its “thinking” time: They told it to wait. Adding the word “wait” during s1’s reasoning helped the model arrive at slightly more accurate answers,

Experts said that s1 raises fundamental questions about the commoditization of AI models.

IBL News | New York

GitHub announced Vision for Copilot, an AI feature that allows users to attach a screenshot of a webpage, diagram, or photo to the chat and generate their code or modified interface.

Another feature of GitHub Copilot is called “Next Edit Suggestions.” It predicts what the user might want to do next.

Now, Copilot packs a new “agent mode” that identifies all the files relevant to the changes a developer is trying to make.

GitHub CEO Thomas Dohmke explained, “Copilot does more of the work to figure out the intent that you had with your original request and then tries to solve that.”

IBL News | New York

Meta introduced LlamaCon, a developer conference that will take place on April 29.

Meta announced it will host its first development conference dedicated to generative AI, LlamaCon, on April 29.

The company said it would share the latest open-source AI developments designed to help programmers build apps and products there.

“It follows an unprecedented growth and momentum of our open-source Llama collection of models and tools.”

No further details were provided. The company’s annual developer conference, Meta Connect, will be held in September.

Meta claims hundreds of millions of downloads of the model and at least 25 partners hosting Llama, including Nvidia, Databricks, Groq, Dell, and Snowflake.

Some of them have built tools that run models at lower latencies.

Meta recently said it would spend as much as $80 billion on AI-related projects this year, including AI hires, new AI data centers, and several Llama models with “reasoning” and natively multimodal capabilities.

IBL News | New York

Researchers at ByteDance, owner of TikTok, demoed a model named OmniHuman-1 that generates realistically convincing deepfake videos. However, the Chinese company has yet to release the system.

To prove the quality of OmniHuman-1, the Chinese company released examples of a fictional Taylor Swift performance, a TED Talk that never took place, and a deepfake Einstein lecture, among others [click on the picture to watch videos].

ByteDance researchers said that OmniHuman-1 only needs a single reference image and audio, such as speech or vocals, to generate a clip of arbitrary length.

They said the output video’s aspect ratio and the subject’s “body proportion” — i.e., how much of their body is shown in the fake footage— are adjustable.

Trained on 19,000 hours of video content from undisclosed sources, OmniHuman-1 can edit existing videos—even modifying a person’s movements.

IBL News | New York

Elon Musk’s xAI company released its latest flagship AI model, Grok 3, yesterday as an alternative to OpenAI’s GPT-4o and Google’s Gemini. Musk also said that xAI plans to open-source Grok 2 in the coming months.

xAI claimed that Grok 3 beats GPT-4o on benchmarks.

“Grok 3 is an order of magnitude more capable than Grok 2,” Musk said during a livestreamed presentation on Monday. “It’s a maximally truth-seeking AI, even if that truth is sometimes at odds with what is politically correct.”

A smaller version of Grok 3, Grok 3 mini, was also introduced.

Grok’s reasoning models were enhanced by a new feature in the Grok app called DeepSearch, xAI’s answer to tools like OpenAI’s deep research. DeepSearch scans the Internet and X to analyze information and deliver an abstract in response to a question.

Musk said some of the reasoning models’ “thoughts” were obscured in the Grok app to prevent distillation, a method AI model developers use to extract knowledge from other models. Recently, DeepSeek was accused of distilling OpenAI’s models to create its own.

https://t.co/hEfQ31gANQ

— xAI (@xai) February 18, 2025

IBL News | New York

Hugging Face cloned OpenAI’s agent “Open Deep Research” within 24 hours after its launch, and it was freely available to open-source developers.

This agent autonomously browses the web and creates research reports.

The code, posted publicly on GitHub, received great responses from developers. It also opened positions for engineers to help expand the project’s capabilities.

Hugging Face‘s researchers said, “While powerful LLMs are now freely available in open-source, OpenAI didn’t disclose much about the agentic framework underlying Deep Research. So we embarked on a 24-hour mission to reproduce their results and open-source the needed framework along the way!”

A core component of Hugging Face’s reproduction was the open-source “smolagents” library, which uses what they call “code agents” rather than JSON-based agents.

The developers behind Open Deep Research built off of the work of others, which shortens development times.

They used web browsing and text inspection tools borrowed from Microsoft Research’s Magnetic-One agent project in late 2024.

IBL News | New York

Linda McMahon, President Trump’s education secretary pick, had her confirmation hearing on Thursday.

The event came amid Trump’s statement that he’d like to immediately dismantle the Department of Education, which primarily facilitates the $1.7 trillion student-loan portfolio.

More than 40 million student loan borrowers hold federal loans, and the department’s Federal Student Aid office works with student loan servicers to manage borrowers’ loan repayment.

“I told Linda, ‘Linda, I hope you do a great job and put yourself out of a job.’ I want her to put herself out of a job,” Trump told reporters on February 4, adding on February 12 that he wanted the Department of Education to be closed “immediately.”

Q: “Why nominate Linda McMahon to be the Education Department Secretary if you’re going to get rid of the Education Department?”

President Donald Trump: “I told Linda, ‘Linda, I hope you do a great job and put yourself out of a job.’ I want her to put herself out of a job.” pic.twitter.com/PreB1Zx4p3

— CSPAN (@cspan) February 4, 2025

When announcing McMahon as his education secretary pick in November, Trump posted on Truth Social that one of his goals for her was to “send Education BACK TO THE STATES.”

Trump and some GOP lawmakers have said that education can be managed locally and that a federal agency isn’t needed.

Trump has also signed an executive order to “promote patriotic education” in public schools and eliminate ideologies that don’t align with his politics.

Eliminating any federal agency requires congressional approval, and while some Republican lawmakers have introduced legislation to eliminate the Department of Education, there hasn’t yet been enough support to make that happen.

During the hearing, McMahon, a former business executive who led the Small Business Administration during Trump’s first term, told ranking member Sen. Bernie Sanders (I-Vt.) that she would continue Pell Grants, which help the neediest students pay for college, and wished to see them expanded for short-term workforce programs.

She also told Sen. Susan Collins (R-Maine) that Title I funds for low-income school districts and IDEA funds for students with disabilities, both appropriated by Congress, would remain.

IBL News | New York

The OpenAI’s Board of Directors rejected yesterday the Elon Musk investment group’s unsolicited offer of $97.4 billion to gain control of the AI company.

In a statement, Bret Taylor, the chairman of the OpenAI board, said, “OpenAI is not for sale, and the board has unanimously rejected Mr. Musk’s latest attempt to disrupt his competition.” Bret Taylor was referring to Mr. Musk’s own AI company, xAI.

OpenAI sent a letter on Friday to Marc Toberoff, the lawyer representing Musk, saying that the offer was “not in the best interests of OpenAI’s mission,” which is to build artificial intelligence that benefits “all of humanity.”

Toberoff said in a statement sent to The New York Times: “This comes as no surprise, given that Altman and Board chair, Taylor, already rejected Musk’s $97 billion bid while stating they had not yet received it. But we are surprised to see the Board, which has strict fiduciary duties to carefully consider the bid in good faith on behalf of the charity, use the same kind of deflective double-talk Altman used in testifying to the Senate.”

Elon Musk also filed a lawsuit in federal court last year to block OpenAI’s restructuring plans.

This week, Robert Bonta, California’s attorney general, said that the state was scrutinizing OpenAI’s plan to shift to a for-profit structure.

xAI raised $6 billion in December, saying it would use the money to build infrastructure and accelerate research and development. BlackRock, Fidelity, Sequoia Capital, and other investors participated in the funding.

IBL News | New York

French AI start-Mistral significantly upgraded its web assistant interface, Le Chat, released a mobile app on iOS and Android and introduced a Pro tier for $14.99 monthly.

The company’s flagship models, Mistral Large, and the multimodal Pixtral Large, were available for commercial use through an API or cloud, such as Azure AI Studio, Amazon Bedrock, and Google’s Vertex AI.

It also released several open-weight models under the Apache 2.0 license.

Mistral is trying to position itself as a credible alternative to OpenAI, Anthropic’s Claude, Google Gemini, or Microsoft Copilot.

Mistral stated that Le Chat runs on “the fastest inference engines on the planet,” which can answer up to 1,000 words per second.

It also claims that it generates much better images than ChatGPT or Grok, as it relies on Black Forest Labs’ Flux Ultra, one of the leading image-generation models.