Weekly analysis of AI & Emerging Tech news from an analyst's point of view.
1️⃣Meta’s chief AI Scientist Yann LeCun thinks AI can master physics rules just by watching videos
Details:
This research was conducted collaboratively by scientists from Meta FAIR, the University of Paris, and EHESS, demonstrating that AI systems can acquire intuitive physical knowledge through self-supervised learning without predefined rules. The research team employed a new method called Video Joint Embedding Predictive Architecture (V-JEPA), which operates more similarly to the human brain's information processing compared to generative AI models like OpenAI's Sora. In the study, the team adapted an evaluation method from developmental psychology known as "violation of expectation," originally used to test infants' understanding of physics. The efficiency of V-JEPA's learning is exemplary, as a small 115 million parameters model watched 128 hours of video to grasp basic physical concepts and became a master!
Analysis:
This is exciting and scary at the same time. This research shows how self-supervised learning is such that the models can "learn" much like a human would. Because it takes a self-supervised learning approach, V-JEPA is pre-trained entirely with unlabeled data. This is one of the best traits of the human brain, it can learn the "unknown unknowns" when it encounters them by analyzing them. This research challenges the fundamental assumption in many AI research projects, which works on the basis that AI systems require predefined "core knowledge" to understand physical laws. The findings of V-JEPA suggest that observational learning can help AI acquire such knowledge, similar to how infants, primates, and even young birds understand physics. The research indicates that V-JEPA can effectively recognize motion patterns and accurately identify physically unreasonable events, laying the groundwork for AI to truly understand the world. What is more scary is that it is done by a small model which doesn't require massive power or compute requirements, and from a very minimal data set of 128 hours of video.
Full details of this project can be seen here - https://ai.meta.com/blog/v-jepa-yann-lecun-ai-model-video-joint-embedding-predictive-architecture/
The full research paper can be downloaded here - https://ai.meta.com/research/publications/revisiting-feature-prediction-for-learning-visual-representations-from-video/
2️⃣Clouds of Suspicion Surround the Death of OpenAI Whistleblower
Details:
The autopsy report of OpenAI whistleblower Suchir Balaji was recently released. His family is skeptical about the conclusion that it was a suicide. They requested a strange hair sample found in his apartment to a private laboratory for testing. According to a report jointly released by the city coroner and police, investigators found no evidence or information at the scene to determine that Balaji's cause and manner of death were not suicide; the cause of death was ruled as a self-inflicted gunshot wound to the head. “The SFPD and OCME have thoroughly reviewed this matter,” police Chief Bill Scott and David Serrano Sewell, executive director of the Chief Medical Examiner’s Office, said in the letter.
“The OCME found no evidence or information to establish a cause and manner of death for Mr. Balaji other than a suicide by self-inflicted gunshot wound to the head ... SFPD conducted an independent investigation and based on the information SFPD reviewed, there is insufficient evidence to find Mr. Balaji's death was the result of a homicide.”
Analysis:
This is sad! This is especially concerning because Suchir promised to release more information soon. The non-response of the police department beyond the initial finding and the quick assessment that it was a suicide and nothing else gives them cause for concern.
While I am not the one to promote conspiracy theories like many others, the death sounds very mysterious, and the investigation feels rushed. Hopefully, they will conduct a thorough investigation and release full details. Especially given that he was a whistleblower saying that OpenAI was using stolen data to train their models, and he was working with the New York Times to file a lawsuit against OpenAI.
To add to this, the death was ruled as a suicide by police within 40 minutes of examining the body. Full details on that can be seen here - https://fortune.com/2025/02/08/openai-whistleblower-suchir-balaji-death-police-investigation-san-francisco-family-questions/
In other news,
- OpenAI updates the GPT-4o model, surpassing DeepSeek capabilities.
The cat-and-mouse game continues with who is on top of the LLM leaderboard. As of now, OpenAI leads all LLMs in creative writing, programming, instruction following, long text queries, and multi-turn dialogue. However, its mathematical abilities remain weak, finishing up sixth on the leader board.
- Mistral AI released its Saba model focused on Middle Eastern and Southeast Asian Languages.
While most LLMs compete mostly in English, Mistral is focusing on other language-speaking countries. One of the problems AI, Generative AI, in particular, is that it often struggles with cultural and linguistic-specific contexts, especially in non-native English regions. Tailored models focused on specific regional languages have emerged to capture this market. Saba is a 24B parameter model trained in datasets from the Middle East and South Asia. In addition to the API access, it also comes as an option to deploy in local environments. The model can run on a single GPU system with a decent speed of up to 150 tokens per second.
- The New York Times recently allowed their product and editorial staff to use AI officially.
This is especially interesting as NYT was involved in multiple lawsuits against some of the major LLM providers accusing them of IP theft of its proprietary data. In an internal email, the company introduced its new AI summarization tool called Echo. While the email encouraged employees to use AI for brainstorming, the employees were told not to use it to write articles or input any confidential source of information. NYT is also planning to use AI to digitally voiced articles and translate content into other languages.
- Netflix uses AI to clone murder a victim's voice creating a public outrage.
Netflix recently released a true crime documentary "American Murder: The Gabby Petito Story.” Social media influencer Gabby was murdered by her fiancé Brian Laundrie in August 2021. Netflix used AI to recreate her voice for the documentary. Her parents approved this and supported the creation of this documentary so her story could be told to the public.
- Elon Musk's new AI model Grok 3 has serious security vulnerabilities that hackers can exploit!
This AI model is susceptible to "simple jailbreak attacks. The "prompt leakage" flaw can expose the full system prompts of the Grok model. While Grok3 performed well on large language model (LLM) rankings, it has failed to meet expectations in terms of cybersecurity.
- UK startup Humanoid launches versatile humanoid robot HMND 01.
Last year, Tesla launched its Optimus robot, which is set to be sold next year for $25,000 to $30,000. Meanwhile, Chinese electric vehicle manufacturer Xpeng has also introduced its humanoid robot Iron, which has already been deployed in the company's factory. Apple and Meta have been developing their own AI robots.
- Meta Announces Inaugural LlamaCon for Generative AI Developers, Focusing on the Llama Series Models.
Meta said it will hold the first LlamaCon developer conference on April 29. This conference is specifically aimed at developers to share the latest achievements in their open-source AI, helping them better build applications and products to create more traction for their Llama series open-source models. Meta said they will showcase Llama’s several new features and technological advancements. Llama models were quite popular until DeepSeek disrupted them recently. Meta claims their models have been downloaded hundreds of millions of times and also hosted by 25+ partners. Meta is facing a lot of backlash to their models, especially when it comes to ethical practices. The company is dealing with copyright-related lawsuits, accused of using copyrighted book materials without permission during the training of its models. They are also facing challenges from EU countries related to data privacy issues.
- xAI Launches Grok 3.
xAI utilized around 200,000 GPUs at their Memphis data center to train Grok3. Grok3 is a family of models, including Grok3Reasoning, Grok3mini reasoning, and Grok3mini, which can respond to questions faster but with slightly reduced accuracy. xAI claims that Grok3 outperformed GPT-4o in multiple benchmark tests including math assessments and physics problems. Grok3Reasoning and Grok3mini Reasoning, can "think" more carefully about problems to solve them rather than just responding with answers.
- By 2027, Generative AI Will Lead to Over 40% of Data Breaches.
Gartner claims that by 2027 over 40% of data breaches related to artificial intelligence will stem from the misuse of generative AI (GenAI). This is scary as enterprises are struggling with structured data and years-old technologies, This new set of technologies will overwhelm the IT Ops personnel who may not understand the technology or the attack surface and security vulnerabilities fully.
- Microsoft Prepares for the Release of OpenAI's GPT-4.5 and GPT-5 Models.
Microsoft is gearing up for the upcoming launch of OpenAI's new models, GPT-4.5 and GPT-5. According to OpenAI CEO Sam Altman, GPT-4.5 is set to be officially released in a few weeks, with Microsoft likely to host this new AI model as soon as next week. Codenamed "Orion," GPT-4.5 will be OpenAI's next-generation model and will be the last to operate without a chain-of-thought reasoning capability.
Reports suggest that GPT-5 is expected to be released by the end of May, following Altman's previous commitment to launch this next-generation model within just a few months. Notably, GPT-5 is anticipated to undergo the most significant changes among the two models, as Altman describes it as "a system that integrates many of our technologies." Furthermore, GPT-5 will incorporate OpenAI's newly introduced o3 reasoning model, which was previously hinted at during the Christmas event in December.
- Gemini 2.0 Flash has the lowest hallucination rate among large language model (LLM) providers.
Vectara released a report titled "Hallucination Leaderboard," which compares the performance of various LLMs in generating hallucinations while summarizing short documents. This leaderboard employs Vectara's Hughes Hallucination Evaluation Model (HHEM-2.1), which is regularly updated to assess how frequently these models introduce false information in their summaries.
According to the latest data from the report, key metrics highlighted include the hallucination rate, factual consistency rate, response rate, and average summary length for several popular models. In the most recent ranking, Google's Gemini 2.0 series performed exceptionally well, especially the Gemini-2.0-Flash-001 model, which achieved the top position with a low hallucination rate of 0.7%. This indicates its ability to introduce almost no false information when processing documents. Following closely behind were the Gemini-2.0-Pro-Exp and OpenAI's o3-mini-high-reasoning models, both demonstrating commendable performance with hallucination rates of 0.8%.
Full details of the hallucination leader board can be seen here - https://github.com/vectara/hallucination-leaderboard
More worthy news, no analysis 🙂
♦ DeepSeek App Surpasses 100 Million Downloads in One Month
♦ Musk announces Grok will be completely free making premium+ users furious
♦ ChatGPT active weekly users reach 400 Million. Paid enterprise users reach 2 Million.
♦ India's Largest Payment Platform Paytm Launches Perplexity AI Smart Search Feature Within App
#GenerativeAI #GenAI #AI #LLM #OpenAI #Google #Gemini #Grok3
