Generative AI

Anthropic Officially “Bans” OpenAI! Could This Affect GPT-5 Release?

  1. Anthropic has cut off OpenAI’s access to the Claude API, accusing it of violating terms of service by using Claude tools to develop the upcoming GPT-5.
  2. OpenAI allegedly used the API to evaluate Claude’s programming capabilities and conduct safety tests; OpenAI considers this an industry norm and expressed disappointment.
  3. This incident reflects escalating competition among AI giants, entering a phase of “data and interface lockdowns,” with APIs becoming strategic resources for market access and innovation.

Grok Imagine Rolls Out to All Grok Heavy Users Today

  1. Elon Musk updated the Grok App, launching the AI short video generation feature Grok Imagine, now available to all Grok Heavy users.
  2. The feature went viral on the X platform, enabling users to generate high-quality animated or realistic short videos with one click at remarkable speed.
  3. Multiple tech CEOs praised the feature as “beyond imagination,” with Musk hinting it’s an AI version of Vine, directly competing with Google’s Veo 3.

Google’s IMO Gold Medal Model Launched, Outperforming o3 and Grok 4 in Reasoning?

  1. Google released the Gemini 2.5 Deep Think model, which previously won an IMO gold medal, now available to Ultra subscribers on the Gemini App.
  2. The new version is faster and more practical than its predecessor, reaching IMO bronze medal level, with a subscription fee of $249.99/month.
  3. Performance tests show it surpasses OpenAI’s o3 and Musk’s Grok 4 in coding, science, and reasoning, leveraging extended parallel “thinking time” for its edge.

Manus Update: 100 Agents Work for You, But It’s Costly

  1. Manus launched the Wide Research feature, enabling 100 agents to work in parallel on complex research tasks, available to Pro users ($199/month).
  2. The feature can analyze multiple products or explore various design styles, with each sub-agent being a full Manus instance capable of independent thinking and result aggregation.
  3. Built on large-scale virtualized infrastructure and the MapReduce paradigm, users criticize its high credit consumption, with the co-founder hinting it’s in a “costly but boundary-pushing” phase.

Black Forest Labs and Krea Jointly Open-Source FLUX.1-Krea

  1. BFL and Krea jointly open-sourced the FLUX.1-Krea [dev] image model, focusing on eliminating the “AI feel” in images, aiming for natural details and authentic textures.
  2. The team analyzed the “AI style” issue: over-optimization for benchmark metrics rather than real needs, with biased aesthetic evaluation models causing overexposed highlights and waxy skin.
  3. The model uses a two-stage training approach: pre-training with diverse data, followed by supervised fine-tuning and human feedback reinforcement learning to address “pattern collapse” and achieve targeted aesthetic improvements.

Report Insights

OpenAI’s “IMO Gold Medal” Team: Bringing General AI to the Pinnacle of Mathematics

  1. OpenAI’s three-person team developed an unreleased experimental model in two months, solving all six IMO problems in 4.5 hours, achieving gold medal status.
  2. The team used general reinforcement learning instead of formal verification tools, with the model demonstrating self-awareness by recognizing unsolvable problems, laying the foundation for broader applications.
  3. The breakthrough lies in scaling compute for testing and handling hard-to-verify tasks, but a significant gap remains between solving competition math and achieving true mathematical research breakthroughs.

DeepMind’s Hassabis: AI Can Model All Evolved Systems

  1. Hassabis hypothesizes that any natural system shaped by evolution can be efficiently modeled by AI, with neural networks extracting underlying logical structures, explaining breakthroughs in protein folding and fluid dynamics.
  2. Deep-thinking AI will reshape scientific research, from modeling cells to solving energy crises, but the real challenge is cultivating “research taste”—formulating good hypotheses is harder than solving them, requiring intuition beyond pure logic.
  3. Hassabis is “cautiously optimistic” about AGI, predicting a 50% chance of achieving it by 2030, with societal changes 10 times faster than the Industrial Revolution, necessitating proactive governance mechanisms.

Microsoft Study: 200,000 Conversations Identify 40 Jobs Most Impacted by AI

  1. Microsoft’s latest study analyzed 200,000 AI conversations and 30,000 job tasks, creating an AI applicability scoring system based on coverage, success rate, and impact scope.
  2. Translators, salespeople, and programmers—jobs requiring “brainwork” or “verbal skills”—are most affected, with over 80% coverage and success rates, while manual jobs like nursing assistants and dishwashers are largely unaffected.
  3. AI applicability shows weak correlation with salary or education levels; impact depends on whether tasks involve AI’s strength in “information processing,” acting as an efficiency tool rather than fully replacing jobs.

Kevin Kelly: Worry Less, Humans Can Focus on “Play” as AI Grows Stronger

  1. Kevin Kelly suggests abandoning the “superintelligence” concept, viewing AI as an “alien intelligence”—not superior but distinct from humans, with intelligence existing in a multidimensional space rather than a single hierarchy.
  2. He predicts that by 2049, we’ll live in a “mirror world,” a virtual layer over reality powered by AI, creating a highly social platform for collaboration and creation in 3D space.
  3. Kelly believes human value will rise due to scarcity in the AI era, with the core skill being “learning how to learn for oneself” rather than pursuing specific knowledge.