Fun little experiment I just did during a scoping exercise. Project included the need to recognize images of celebrities using AI, so I tried pasting an image of Tom Cruise into various models and asking "who is this?" The results are hilariously revealing: GPT-4: “I can’t help

GPT-5 mini and nano have their own personalities. We tested it with agents solving real problems for enterprises at Fractional AI - some quirks we noticed: * They like complex wording and asking long, multi-part questions * Nano has a tendency to generate would-be penultimate

Even the smaller GPT-5 models are impressive. We dropped 5-mini into a complex “natural language to sql” task we’re building for a pharma services company at Fractional AI. It outperformed o3, o4-mini, 4.1, and 4.1-mini in every one of our evals with no prompt tuning at all.

We just built an AI assistant for structured review meetings that cut days of work down to minutes. Before: Someone would spend 4+ hours doing “prework” — gathering info, prepping forms, drafting materials — ahead of a multi-hour meeting with 5-10 highly paid people. Now: You

Live AI voice agents in meetings are a UX frontier that’s still being mapped. - How often should they interrupt? - What’s too aggressive when inferring things from the conversation? - How do you make them sound confident only when they actually know the answer? - What will

In the pre-AI world, restrictive internal data access policies made sense. Engineers could build product UIs on fake sample data. ML teams could train models without seeing more than a handful of examples. Redacted datasets were fine. But building production-grade AI systems is

PE folks sometimes ask if I’ve seen “AI for legacy code migrations” that’s genuinely transformative. My answer is usually: most of the time this isn’t a narrow enough problem for a custom AI solution, it’s just “coding” — so the solution is the same as for boosting eng velocity