AI Updates on 2025-10-26

DeepSeek-OCR demonstrates exceptional handwritten text recognition capabilities, accurately parsing extremely difficult handwritten letters including mathematical equations from 1913 @deedydas

OpenAI projects historically unprecedented growth to $100 billion in revenue, with IPO conditions requiring restructuring and going public by end of 2025 @a16z
Sam Altman's Neuralink competitor Merge Labs is preparing to announce after raising $250M at $850M valuation, with most capital coming directly from OpenAI, planning to alter neurons through gene therapy and interact via ultrasound @AndrewCurran_
Ohio House Bill 469 would prevent AI from being a solo founder and CEO, potentially blocking Sam Altman's concept of a 'zero person unicorn' company @AndrewCurran_
Perplexity launches Perplexity Finance with sufficient daily usage to warrant sidebar placement for easy access @AravSrinivas
Engineering managers at companies heavily using AI coding tools now look for software engineers who can navigate complexity and get things done, rather than just technical skills @GergelyOrosz

Stephen Wolfram suggests LLMs may have definitively shown that consciousness is not magical beyond physics, with awareness potentially originating as a simple decision-making mechanism for early animals @vitrupo
AI labs are characterized as rapidly growing startups with multiple entrepreneurial people making choices in uncertain environments, rather than coherently executing long-term strategies @emollick

Sora demonstrates improved capability in generating Magic: The Gathering gameplay videos, creating fake but appropriately colored cards and showing closer approximation to actual game mechanics @emollick
AI coding tools enable developers to perform advanced Git operations like rewriting commit sequences and recovering files from reflog, transforming previously monthly tasks into daily workflows @simonw
GenAI is expected to enable widespread game development by reducing the getting started phase from 30 minutes to 2 minutes, potentially creating 100 million new developers @OfficialLoganK

HRM and TRM approaches achieve state-of-the-art results on ARC-AGI using zero external knowledge, with TRM being the public leader for such approaches, suggesting potential superhuman capabilities on reasoning problems @fchollet
Base models like Llama 3.1 405B provide insight into fragmentary associative concepts beneath human writing, offering potential research opportunities for humanities scholars studying archetypes and collective unconscious @emollick
Technical debugging investigation reveals PyTorch MPS backend issues with non-contiguous output tensors, raising questions about when LLMs will be capable of such complex technical detective work @karpathy
Frontier AI labs are reportedly understaffed despite immense amounts of low-hanging fruit, leading to intense work schedules and ruthless prioritization due to insufficient people and compute resources @brianryhuang