Meta released Muse Spark 1.3, its most powerful AI model to date, which developers can access Wednesday. The company says it's narrowing the gap with leading competitors like OpenAI and Google.

• Developers get paid access to Muse Spark 1.3 starting Wednesday
• Meta plans to integrate the model into Instagram, Facebook, and Meta AI
• Meta's chief AI officer claims the model's capabilities are now closer to top competitors

Two mathematicians used large language models to discover counterexamples to long-standing conjectures, automating hypothesis generation and testing with minimal guidance. The article calls this “brute intelligence,” where AI runs iterative search loops to tackle any problem framed like a math exercise. It argues we’ll need to reshape tasks into testable, calculable formats for AI to industrialize discovery across fields.

• Tao got an LLM (Anthropic's Fable) to produce a counterexample to the Jacobian conjecture over a weekend, calling it "a massive miracle" unlikely via manual search.
• Rybin disproved another long-standing conjecture using ChatGPT alone with just four prompts and no expert steering.
• The pattern works because these problems admit short, verifiable counterexamples—AI can find them if it can check its own work.
• The real implication is reframing fields (code, drug discovery, finance) into testable, formalized "math-like" tasks so AI can brute-force solutions via fast, parallel iteration.

Karpathy outlines three software eras: human-written code, trained neural weights, and now natural language prompts. He claims LLMs treat English as code, making traditional coding steps optional. The linked talk will explain why this shift matters in 40 minutes.

• Karpathy frames software history as three eras: 1.0 (hand-written code), 2.0 (neural net weights), and now 3.0 (natural language prompts)
• In Software 3.0, English prompts function as code and LLMs act as the CPU running them
• This makes traditional coding infrastructure (compilers, libraries, frameworks) optional rather than mandatory
• A 40-minute talk/demo is promised to explain why this shift matters

Deep Think with Confidence (DeepConf) is introduced as a method to improve reasoning efficiency and performance in large language models by using internal confidence signals to filter out low-quality reasoning traces. It requires no additional training or tuning and can be easily integrated into existing systems. Evaluations show significant accuracy improvements and a reduction in generated tokens on various reasoning tasks.

• DeepConf uses internal confidence signals to filter out low-quality reasoning traces without any additional training or tuning
• It integrates easily into existing LLM systems
• Evaluations show significant accuracy improvements alongside a reduction in generated tokens across various reasoning tasks