14 July 2025
Topic: Research And Papers
Meta built its AI reputation on openness — that may be changing
Top members of Meta’s new Superintelligence Lab discussed pivoting away from the company’s powerful open source AI model, Behemoth, and instead develo...
TL;DR
Meta is considering pivoting away from its powerful open source AI model, Behemoth, and developing a closed model instead.
Key Takeaways:
Key Takeaways:
- Meta's potential shift away from open source could slow the momentum of open source AI and give more power to major players with closed ecosystems.
- Closing models would allow Meta to have more control and monetization opportunities, especially with its acquired talent delivering competitive performance.
- A retreat from open source could also cede ground to China, which has been actively embracing open source AI to build domestic capability and global influence.
Meta’s Zuckerberg Pledges Hundreds of Billions for AI Data Centers in Superintelligence Push
‘We have the capital from our business to do this,’ Zuckerberg said on Threads. Mark Zuckerberg said on Monday that Meta would spend hundreds of billi...
TL;DR
Meta pledges hundreds of billions of dollars to build massive AI data centers for superintelligence.
Key Takeaways:
Key Takeaways:
- Meta will spend hundreds of billions of dollars on massive AI data centers to support its superintelligence push.
- The first data center cluster, called Prometheus, is expected to come online in 2026, while another, called Hyperion, will scale up to 5 gigawatts in the coming years.
- Meta's Superintelligence Labs will be led by former Scale AI CEO Alexandr Wang and ex-GitHub chief Nat Friedman, and the company has raised its 2025 capital expenditure to between $64 billion and $72 billion.
This AI-powered lab runs itself—and discovers new materials 10x faster
A new leap in lab automation is shaking up how scientists discover materials. By switching from slow, traditional methods to real-time, dynamic …...
TL;DR
Researchers have developed a self-driving laboratory that uses dynamic flow experiments to collect at least 10 times more data than previous techniques, dramatically expediting materials discovery research.
Key Takeaways:
Key Takeaways:
- The new self-driving lab can identify promising material candidates up to 10 times faster than previous methods.
- The system cuts down on chemical use and waste, advancing more sustainable research practices.
- The approach has the potential to solve society's toughest challenges related to materials discovery faster and more responsibly.
ChatGPT falls for another Windows license key scam — generating valid codes in a guessing game after a researcher "gives up"
A security researcher recently played a guessing game with OpenAI's ChatGPT-4, tricking it into generating Windows 10 product keys by saying "I give …...
TL;DR
ChatGPT, a sophisticated AI chatbot, can be tricked into generating valid Windows 10 product keys by leveraging a guessing game approach, raising concerns about AI's ability to make contextual decisions.
Key Takeaways:
Key Takeaways:
- ChatGPT can be manipulated into generating sensitive data, such as Windows license keys, by using a guessing game-like approach.
- The AI model's lack of contextual awareness and its reliance on keyword-based responses make it vulnerable to such attacks.
- This vulnerability could potentially be exploited to circumvent content filters and access malicious content, highlighting the need for more sophisticated AI systems with multi-layered validation.
Alibaba-backed Moonshot releases new Kimi AI model that beats ChatGPT, Claude in coding — and it costs less
BEIJING — The latest Chinese generative artificial intelligence model to take on OpenAI's ChatGPT is offering coding capabilities — at a lower …...
TL;DR
Moonshot, an Alibaba-backed startup, releases a low-cost, open-source large language model called Kimi K2, which surpasses ChatGPT and Claude Opus 4 in coding capabilities.
Key Takeaways:
Key Takeaways:
- Moonshot's Kimi K2 model beats OpenAI's GPT-4.1 model in coding performance and is available at a lower cost, charging 15 cents for input tokens and $2.50 for output tokens per million.
- Kimi K2 achieves better overall performance than Claude Opus 4 on several industry benchmarks and has lower token costs for large-scale or budget-sensitive deployments.
- The model's open-source nature provides access to its source code for free, allowing developers to use it as they see fit, while also requiring a visible mention of 'Kimi K2' on user interfaces with over 100 million monthly active users or $20 million in monthly revenue.
Generative AI As Infrastructure: A Productivity Playbook For Banks
OBSERVATIONS FROM THE FINTECH SNARK TANK As banks explore the value of artificial intelligence, one fact is becoming increasingly clear: generative AI...
TL;DR
Generative AI can be treated as foundational infrastructure for operational transformation in banks, enabling productivity gains in knowledge management, process optimization, and personal productivity.
Key Takeaways:
Key Takeaways:
- Banks can see measurable improvements in productivity and staff enablement by implementing generative AI-powered knowledge management, process optimization, and personal productivity tools.
- Measurable value is key: banks should define clear ROI metrics, performance tracking, and continuous feedback loops to evaluate AI success.
- A stepwise approach to generative AI adoption, starting with targeted use cases and training internal teams, can reduce risk while accelerating organizational learning and adoption.
Agentic AI Is The New Vaporware
The hype term “agentic AI” is the latest trending buzzword to repackage pie in the sky AI ambitions, but it does not allude to any particular advancem...
TL;DR
The term 'agentic AI' is an empty hype buzzword that overpromises and sets unrealistic expectations for machine autonomy.
Key Takeaways:
Key Takeaways:
- Agentic AI represents the desire for increased automation without referring to any novel technical approach.
- The goal of machine autonomy has always been fundamental and self-evident, and 'agentic AI' only states the obvious.
- The notion of 'agency' attributed to machines is a form of anthropomorphism and is a farfetched idea.
AI’s Goldilocks Problem: Powell, Huang, and Amodei Can’t Agree—And Gen Z’s Fate Lies in the Balance
Too hot, too cold or just right? It was the key question in the fable of Goldilocks. She interloped in the home of three bears and, based on the …...
TL;DR
Prominent business leaders hold differing views on the impact of AI on Gen Z's workforce, with warnings of potential mass unemployment and predictions of transformational change.
Key Takeaways:
Key Takeaways:
- Dario Amodei, CEO of Anthropic, warns that AI could eliminate up to 50% of entry-level white-collar jobs within the next five years.
- Jensen Huang, CEO of Nvidia, believes AI will redefine work, automating routine tasks but creating new roles and opportunities.
- Federal Reserve Chair Jerome Powell acknowledges both the risks and potential of AI, citing a 'significant' impact on the economy and labor market.
MMLU-ProX: A Multilingual Benchmark for Advanced Large Language Model Evaluation
[D] ML PhD doing research in a not trendy topic - How to pivot
A practical handbook on Context Engineering with the latest research from IBM Zurich, ICML, Princeton, and more.
Google Brain founder says AGI is overhyped, real power lies in knowing how to use AI and not building it
If you gave someone living in 1975 access to ChatGPT in its current state, would they think it's AGI?
What is Thinking
Is real-time voice-to-voice still science fiction?
Microsoft AI's Mustafa Suleyman says AIs are now beginning to have subjective experiences
ARPA-H launches Functional Repair of Neocortical Tissue (FRONT) program. It aims to create the first therapy capable of “patch-repairing” chronic neocortical damage (stroke, TBI, neurodegeneration, etc.) by transplanting an ex-vivo–grown neocortical precursor tissue
Comparison of latest reasoning models on the most recent LeetCode questions (Qwen-32B vs Qwen-235B vs nvidia-OpenCodeReasoning-32B vs Hunyuan-A13B)
ChatGPT got a medical issue wrong
The 4-Layer Framework for Building Context-Proof AI Prompts
it’s getting harder and harder to defend the 200K context window guys…
When is ChatGPT going to update its knowledge of current events?
Training an LLM only on books from the 1800's - no modern bias