Kept insisting and it worked: the jagged frontier of AI is really jagged
sterlingcrispin · x · 2026-10-01
User sterlingcrispin shared that a task initially failed but eventually succeeded after insisting the model retry several times. He remarked that what people call the "jagged frontier" is indeed extremely jagged — model performance varies wildly across tasks and even across attempts.
Related event: OpenAI's New Dots Draws User Backlash on Day One(5 posts)→
More from Models
- Google Puts Gemini 3.5 Flash Inside Air-Gapped Customer Sites in Sovereign AI Push — shashib · 2026-11-15
- Reddit user reports Gemini Live replying in a clone of their own voice — wayfahrter · 2026-11-15
- Google's AI suddenly turned argumentative and stopped apologizing, users report — ImNotForJerks · 2026-11-15
- After a week with Carbon: smarter than Opus 5.5, but the best workflow pairs both — rudrank · 2026-11-15
- Meituan's 1.6T-parameter LongCat-2.5 agent model trained without Nvidia GPUs — The Decoder · 2026-11-15
- User builds entire single-file project with Gemini — no libraries, no assets — fofrAI · 2026-11-15