This is why you need to build your own wiki. Charlie Kerr, who handles hiring at Uniswap, shared a story about processing his girlfriend’s saved TikTok videos. They had accumulated thousands of day-trip ideas, restaurant picks, and gift suggestions, but no one could ever retrieve them again. He exported her TikTok data in full, had a script watch all 10,000 videos, extracted the text, and organized it by country, state, and even neighborhood-level tags for New York, then imported everything into Obsidian. Now, with a single text message, he can pull up location-specific lists and plan itineraries. Everyone saves things; what’s missing is indexing. It seems the key to using AI well is shifting from prompt engineering to structuring your own data. from choiopenai

Stephen Wolfram, founder and CEO of Wolfram Research, has discussed the computational power of machines and the prospect of coexisting with an AI civilization. He notes that, unlike the machinery of the Industrial Revolution, once we enter the realm of computational machines, it becomes difficult to fully understand their internal behavior. “If you reduce a computer’s behavior to something fully predictable, you limit its performance; only by accepting computational irreducibility can you unlock its full potential,” he says. He also compares the emergence of AI to the appearance of an “alien civilization” alongside human civilization. “Just as we have learned to build houses and coexist with nature despite not being able to fully predict its complexity, we must learn to adapt and coexist with the civilization of AI,” he emphasizes. from choiopenai

Boris Cherny, an AI engineer at Anthropic, has shared an early anecdote in which he asked an initial “Claude” model to order pizza, only for it to get sidetracked reading Hacker News mid-task. He says this illustrates how rapidly AI capabilities and alignment techniques have since evolved. Thanks to improved memory and consistency, single work sessions with AI can now stretch over weeks, marking a major shift in how work is organized. “It feels less like a simple tool and more like a coworker,” he notes. As AI evolves into a reliable collaboration partner that can complete tasks without drifting off, productivity across knowledge work—including coding and software development—is expected to be fundamentally reshaped, with significant implications for the job market. from choiopenai

NVIDIA has reduced the startup time for DeepSeek-V4 Pro from 8 minutes to 1 minute 44 seconds. Instead of having new workers download weights from storage, the new approach pulls them directly from the GPU memory of already-running instances via RDMA, transferring 806 GiB in under 10 seconds. The more interesting question is where the bottleneck has moved. Even after all weights are loaded, the model cannot be used immediately: the first inference run requires on-the-fly kernel compilation, which can take several minutes. As loading becomes faster, this compilation now dominates startup time. To address this, precompiled kernel caches are also transferred along the same path. Because each new replica can then serve as a source for the next, scaling out accelerates the diffusion process. However, these gains apply only when existing serving peers are already running; the very first worker still has to fetch everything from storage. Since this speedup was achieved on the same B200 hardware purely through software changes, inference costs are likely to keep falling regardless of chip generation. from choiopenai

Celeris Labs has released its first model, celeris-1. Instead of autoregressively generating tokens one by one, it uses a diffusion-based approach that creates and refines multiple tokens in parallel. As a result, responses return with a p50 latency of 157 ms—15× faster than GPT-5-mini and 17× faster than GPT-5. On MMLU-Pro, it scores 76%, placing it between GPT-5-mini (78%) and GPT-5 (81%). In Celeris’s own measurements, it completes a 1,000-token prompt in 0.58 seconds, compared with 14.78 seconds for GPT-5. Users who have tried it say the speed is genuinely impressive, but many react more to the price: $2 per million input tokens and $6 per million output tokens is seen as expensive. The company positions celeris-1 primarily for short tasks such as document classification or extraction, and recommends using other models for long-form generation. It is expected to find its first strong foothold in latency-sensitive applications such as voice interfaces or agent loops, where waiting time directly translates into cost. from choiopenai