OpenAI o3

122 HackerNews comments reveal a deeply technical, skeptical audience debating whether o3/o3-mini represents genuine reasoning progress or sophisticated pattern matching. On the positive side, one user reports o3-mini is the FIRST model to correctly solve a twisted wolf/goat/cabbage riddle where even o1 and o1-pro reasoned correctly then still concluded 'goat' — suggesting incremental reasoning improvements. Users discuss ARC-style visual pattern tasks that o3 still cannot solve, debating whether textual training data alone can teach spatial reasoning. On coding, a Cursor user who tested o3-mini after claude-3.5-sonnet, o1, o1-pro, and deepseek-r1 reports ALL reasoning models suffer from 'ADHD' — losing the plot in multi-step coding tasks (e.g., a 15-LOC NextJS page.tsx edit spiraling into unnecessary changes). Another user with 25+ years of software experience notes LLM coding works best in small teams with clear build understanding, but every line still needs review because models 'spit out working code' that's architecturally leaky. Hallucination remains a top concern: one user reports even the best models produce BS in 3-4 paragraphs on well-documented topics (specific example: incorrect guidance on running N VPN servers on N IPs with IP binding). Broader sentiment is skeptical about AGI claims — users distinguish between benchmark improvements and real societal transformation, noting that training costs have become 'prohibitively expensive even for companies with the resources of OpenAI.' Economic concerns about data labelers, job displacement, and who benefits from AI advances are repeatedly raised.

No aggregate ratings were found for this product during the last harvest.