Survey: AIEWF 2025

“AI-native” is still forming. Past ML cycles show the trail:

• Data is moat. Narrow eval sets build deeper moat (see 3DBench or ShapeNet-Core55).
• Close the loop. Grab 👍/👎 or edits on every output (ChatGPT buttons, Copilot inline edit).
• Show work. Stream tokens or rough grids so waiting feels alive (ChatGPT typing, Midjourney first grid). Cheap silicon like the $8 Sipeed K210 keeps shrinking the wait.
• Blend model tiers. Jump across rules, GPT-4o, and LoRAs to hit latency/cost (Instagram now serves 1 000 + models in one feed).
• Ship on every surface. Text, voice, mobile, desktop, watch, AR.
• Textbox + mic are default UI. Offer context nudges before the user types (“Ask about this PDF?”).
• Expose background agents. Status chips or logs show what cron-style jobs are doing.
• Skip cute personas. Nobody needs “Alice the CMO bot.” They want invoices sent—with trace.
• Price on delivered value. Bill by answers, leads, or bugs caught (Perplexity Pro, Snyk AI, Copilot Chat).
• Be AI-native inside, too. One agent per user watches analytics and fires campaigns through Resend—proofs like Retool Agents or GrowthBook AI hint at the future.

“AI Native” applications and patterns are still too early to commit, but we know tons from previous ML/DL cycles and can extrapolate the rest based on patterns being pushed by the top consumer apps: Data is a moat. Niche evaluation datasets are a bigger moat. Collect user feedback on AI output, e.g., thumbs up/down (ala ChatGPT) Compute will improve over time like CNN can now train/run on $9 ASIC but in the interim, give the user a sense of progress, i.e., “work is being done,” e.g., text streaming (ala ChatGPT) and partial renders (ala Midjourney) Route b/w varying levels of intelligence for use-cases; make it all look seamless to the user, e.g., FB deploys 20+ RL/other models per user The user should be able to use the product across modalities (text, voice), devices (mobile, desktop, watch, AR) and form factors (mobile/web/desktop app, phone call, sms) Big Tech, OAI, and others will build user behavior around the textbox and voice. Align with this design affordance. Chat isn’t limited to completion; agents will run jobs (complex) in the background. Users stare at blank text boxes. They need to be context-aware and provide ‘nudges’ to direct the initial action or interaction. Agents working in the background need to demonstrate progress and allow users to see what’s happening under the hood. The earliest example of this in software was the use of batch or cron jobs. Anthropomorphizing intelligence is not great UX; users don’t care for Alice, the CMO AI, but instead want job X done. Do the job, do it well, do it with traceability. Value-based pricing is more straightforward when selling intelligence; still, aligning cost and pricing models is essential. AI-native means not just that the product is AI-native but also that all internal workflows and tooling are AI-native. Facebook built its own servers and infrastructure. AI-native companies should reconsider their tooling, for example, by deploying one agent per user that analyzes user analytics, journeys, and features and runs marketing campaigns via Resend rather than using Iterable or Customer.io.