Meta Llama 4 Delay, Scale AI Challenges, Midjourney Partnership

Meta is navigating a complex AI strategy involving partnerships, internal restructuring, and significant investments. Despite investing heavily in Scale AI, challenges have emerged, including executive departures and questions about data quality. Meta's AI division is undergoing reorganization, now known as Meta Superintelligence Lab (MSL), with new hires from OpenAI and Scale AI, though some staff have left after short periods. Meta is partnering with Midjourney for image generation and has open-sourced its Llama models, which have seen over 600 million downloads. However, Meta delayed its next AI model, Llama 4 Behemoth, due to engineering problems. Other companies are also making strides in AI. Alibaba continues to invest in AI, with revenue growing over 100% each year for seven quarters, focusing on natural language processing and computer vision. Palantir AI will be used by Coventry council in social services, raising ethical concerns. The Air Force is testing AI tools to speed up battle decisions, while UC Riverside scientists have developed a method to remove private data from AI models. However, there are also concerns about over-reliance on AI. A study indicates doctors may lose skills if they depend too much on AI, and electricity prices are rising due to the energy demands of AI data centers. Andrej Karpathy is skeptical about using reinforcement learning for training large language models, suggesting AI should learn from interactive environments instead. The importance of traditional learning methods and human intelligence remains critical in the age of AI.

Key Takeaways

  • Meta's partnership with Scale AI faces challenges despite Meta's investment, with executive departures and concerns about data quality.
  • Meta's AI division is undergoing its fourth reorganization in six months, now called Meta Superintelligence Lab (MSL).
  • Meta's open-source Llama models have over 600 million downloads, but Llama 4 Behemoth was delayed due to engineering issues.
  • Alibaba's AI revenue has grown over 100% each year for seven quarters, focusing on NLP and computer vision.
  • UC Riverside developed a method to remove private data from AI models without needing original training data.
  • The Air Force is testing AI tools like DASH 2 to help battle managers make faster decisions.
  • Coventry council will use Palantir AI in social services, raising ethical concerns about data privacy.
  • A study suggests doctors may lose skills if they rely too much on AI for tasks like cancer detection.
  • Electricity prices are rising, partly due to the energy demands of AI data centers.
  • Andrej Karpathy is skeptical about using reinforcement learning for LLM training, advocating for AI to learn from interactive environments.

Meta and Scale AI partnership showing cracks after investment

Meta's partnership with Scale AI is facing challenges despite Meta's recent investment. Ruben Mayer, a Scale AI executive brought to Meta, left after two months. Meta's TBD Labs is using data vendors like Mercor and Surge, Scale AI's competitors. Some researchers find Scale AI's data quality low. Meta invested in Scale AI to attract talent like CEO Alexandr Wang, but the value of Scale AI to Meta is now questioned.

Meta and Scale AI face partnership challenges after big investment

Meta's partnership with Scale AI is having problems even after Meta invested in the data company. Scale AI's former executive, Ruben Mayer, left Meta after only two months. Meta's AI lab, TBD Labs, is using other data companies like Mercor and Surge. Some people at Meta think Scale AI's data isn't very good. Meta invested billions in Scale AI, possibly to get CEO Alexandr Wang, but Scale AI's value to Meta is now uncertain.

Meta's AI team faces disruption with new hires and departures

Meta's AI division is experiencing upheaval with new hires from OpenAI and Scale AI. Shengjia Zhao, a ChatGPT co-creator, almost quit Meta shortly after joining. Some new AI staff have left after short periods, while veteran employees are also departing. Mark Zuckerberg is reorganizing Meta's AI group, now called Meta Superintelligence Lab (MSL), for the fourth time in six months. Alexandr Wang leads a secretive department called TBD, focused on building cutting-edge AI models.

Meta's AI strategy partnerships and competition in the tech world

Meta is using partnerships to boost its AI, working with Midjourney for image generation and investing in Scale AI for model training. Meta spent $14.8 billion on Scale AI and $12.94 billion on AI research in Q2 2025. Meta's open-source Llama models have over 600 million downloads. Meta is also part of the AI Alliance, focusing on safe and responsible AI. Despite challenges, Meta's strong finances support its AI investments and growth.

Alibaba continues investing in AI for long-term growth

Alibaba is still investing heavily in AI products and services, even without immediate big profits. The company's AI revenue has grown over 100% each year for seven quarters. Alibaba is focusing on natural language processing, computer vision, and machine learning. Challenges include geopolitical issues and AI regulations in China. Alibaba is optimistic about AI's long-term potential, focusing on AI integration and open-source models.

UC Riverside develops method to remove private data from AI models

UC Riverside scientists created a way to remove private and copyrighted data from AI models. The method works without needing the original training data. This addresses concerns about personal information staying in AI models. The innovation makes AI models forget specific information while keeping their functionality. The technique uses a substitute dataset to adjust the model and add noise, ensuring data is erased.

Why your electric bill is going up now

Electricity prices are rising because of old equipment, high demand, and new AI data centers. The average electricity price is up 7% from last year and 32% from five years ago. An outdated power grid can't handle the load, causing higher costs. Data centers use a lot of energy, increasing demand. Red tape delays energy projects, and extreme weather requires stronger, more expensive equipment.

Air Force uses AI to speed up battle decisions

The Air Force tested AI tools that help battle managers make faster decisions. The tools give a better view of the battlefield. The experiment, called DASH 2, involved AI microservices for choosing the best weapons for targets. Companies developed code to rank available weapons based on battlefield data. The AI tools helped battle managers make decisions more quickly and confidently.

Why human intelligence is still important in the age of AI

An article discusses the importance of traditional learning methods in the age of AI. It encourages using libraries, taking notes, and critical thinking instead of relying solely on AI. Meta Platforms delayed its next AI model, Llama 4 Behemoth, due to engineering problems. A psychology professor notes AI requires enormous hardware and power, unlike the human brain. The article emphasizes the unique value and potential of human intelligence.

Coventry to use Palantir AI in social services raising ethical concerns

Coventry council will use Palantir AI in social work and children's services under a £500,000 contract. Palantir supplies AI to the Israel Defense Forces. Workers are concerned about ethical issues and data privacy. The AI will help with case notes and support for children with special needs. Unions question the use of AI in surveillance and data collection.

AI reliance may cause doctors to lose skills

Doctors may lose cancer detection skills if they rely too much on artificial intelligence. A study showed that doctors' ability to find adenomas decreased after using AI for colonoscopies. Researchers suggest doctors become less focused without AI assistance. The study raises concerns about deskilling in medicine due to AI use. More research is needed to track patient outcomes and minimize deskilling.

AI expert skeptical of reinforcement learning for LLM training

Andrej Karpathy, an AI researcher, is doubtful about using reinforcement learning (RL) to train large language models (LLMs). He believes RL reward functions are unreliable and not good for teaching problem-solving. Karpathy suggests training AI in interactive environments where they learn from their actions. He argues future AI should learn from experience rather than copying human language.

Sources

Meta Scale AI Partnership Investment Data quality TBD Labs AI talent Alexandr Wang AI division OpenAI ChatGPT Meta Superintelligence Lab (MSL) Mark Zuckerberg AI strategy Midjourney Image generation Llama models AI Alliance Open-source Alibaba AI products AI services Natural language processing Computer vision Machine learning AI regulations UC Riverside Private data removal Copyrighted data AI models Data privacy Electricity prices AI data centers Power grid Air Force AI tools Battle decisions DASH 2 Human intelligence Learning methods Libraries Critical thinking Llama 4 Behemoth Coventry council Palantir AI Social services Ethical concerns Data collection Doctors Cancer detection Deskilling Reinforcement learning Large language models (LLMs) Andrej Karpathy