OpenAI Triples GPT-5.6's ARC-AGI-3 Score
On July 29, OpenAI announced that adjusting just two API settings for GPT-5.6 tripled its score on the ARC-AGI-3 benchmark. The keys were two settings: "reasoning retention" and "compaction".
ARC-AGI-3 is a highly challenging benchmark that tests abstract reasoning patterns, and it has drawn attention as a measure of an AI model's generalization ability. According to the announcement, the change brought substantial gains in both score and efficiency. That such an improvement came purely from API configuration tweaks—without modifying the model itself—suggests there is still significant headroom in operational optimization.
Meta Earnings: Zuckerberg Says "Billions of People Will Have AI Agents Within 5 Years"
Meta reported its Q2 earnings on July 29. During the announcement, CEO Mark Zuckerberg predicted that billions of people will have personal AI agents within five years. The company has poured billions of dollars into AI infrastructure and agents, and Zuckerberg continues to make the case to investors that the spending will pay off.
Zuckerberg also said Meta's enterprise AI opportunity extends beyond agents to APIs, compute, and internal software. Meanwhile, investor frustration over the massive AI spending is growing, and Meta's stock fell following the earnings report. The BBC reports that "frustration is mounting over AI spending plans".
Microsoft: $3.2 Billion Profit from Anthropic Investment, OpenAI a "Mixed Bag"
In its FY2026 Q4 earnings, Microsoft disclosed its AI investment returns in an interesting way. The company booked a $3.2 billion profit from its investment in Anthropic, while describing its stake in OpenAI as a "mixed bag" with both good and bad aspects.
Microsoft also became the only data center giant to hold its AI-related capital expenditure (capex) flat. With competitors ramping up AI investment across the board, Microsoft's stance is an outlier. At the same time, industry reactions to AI's influence remain divided—notably, the GCC compiler project has signaled a policy of generally rejecting significant contributions from AI/LLMs.
ChatGPT Expected to Be Designated an EU "Very Large Platform"
According to Bloomberg, OpenAI's ChatGPT and Roblox are expected to be designated "Very Large Online Platforms" (VLOPs) under the Digital Services Act (DSA) after exceeding 45 million monthly users in the European Union. The formal designation could come as soon as August.
The designation will subject ChatGPT to stricter content moderation requirements, monitoring obligations, and transparency reporting. An AI chat service coming under the DSA is a notable milestone, showing that generative AI has taken on the societal influence of a platform.
Anthropic's Mythos Finds Unknown Weakness in Cryptography
Anthropic announced that its AI security model, "Mythos", discovered a previously unknown weakness in the mathematical foundations of a widely used cryptographic algorithm. According to Ars Technica, the finding does not directly break existing cryptosystems at this point; what was found is a method that "moderately reduces" the computational effort needed to crack the encryption.
That said, some have noted that Anthropic's announcement may contain exaggeration, and careful verification will be needed to assess the results. Even so, AI bringing a new approach to cryptanalysis deserves attention as a possible sign of future security risks.
Lilian Weng Leaves Thinking Machines and Returns to OpenAI
Lilian Weng, co-founder of Thinking Machines Lab, has returned to OpenAI after leaving the company for health reasons. Weng previously served as VP of AI Safety Research at OpenAI.
The movements of prominent AI researchers are an important signal for reading each company's research priorities and organizational direction. Thinking Machines is a startup founded by Mira Murati (former OpenAI CTO), and Weng's departure is seen as a significant blow to the company.
Other Notable Developments
Trillion-dollar selloff in semiconductor stocks**: AI-related chip stocks including SK Hynix, Samsung, and SoftBank plunged. The AI investment frenzy may be starting to cool.
AI startups sharply cut paper publishing**: According to Science, top-tier AI startups have drastically reduced their research paper output, heightening tensions with open science principles.
AI security leaderboard**: FAR.AI released a leaderboard comparing AI models on cybersecurity safety and CBRN (chemical, biological, radiological, and nuclear) safeguards. GPT-5.6 Sol and Fable 5 came out as the most robust.
GGUF quantization of Kimi K3 (2.8T parameters)**: The LocalLLaMA community reported an experiment running Kimi K3 on CPU with Q3_K_S quantization (1.1TB).