From: aidotengineer
This article summarizes insights from a survey of 100 AI Engineers and discussions during a “Frontier Feud” event, highlighting current trends and considerations in AI engineering [00:01:07].
Current Industry Perspectives
Model Deployment and Architecture
There is a predicted trend towards the majority of deployed models being on-device within a year and a half [00:03:34]. This shift suggests a move towards smaller models that are orchestrated together and potentially hyperspecialized for specific tasks, rather than solely relying on larger cloud-based models [00:03:40].
One participant’s “hot take” suggests that Transformers may not be the final architecture, with a future possibility of models being built using biological materials [00:04:19].
Model Selection Considerations
When AI engineers choose models, the top considerations identified by the survey are:
- Cost [00:10:47]
- Latency [00:11:04]
- Accuracy/Performance [00:11:28]
- Open Source vs. Closed Source [00:12:45]
Workforce and Disruption
Predictions suggest a significant shift in the AI workforce, with at least one of the leading five individuals currently training large models potentially no longer doing so by the end of the year [00:01:55].
Jobs perceived to be most at risk of AI disruption include:
- Software Engineers [00:17:44]
- Content creation/Writing [00:20:14]
- Data Entry [00:20:43]
AI Tools and Preferences
Popular AI tools favored by engineers include:
- Cursor [00:17:34] [00:18:08]
- Model APIs [00:20:04]
Buzzwords and Areas of Fatigue
AI Engineers are reportedly tired of hearing the following buzzwords:
- Agents [00:13:56]
- AGI (Artificial General Intelligence) [00:15:30]
- RAG (Retrieval-Augmented Generation) [00:16:21]
- Prompt Engineering [00:16:25]
Challenges and Nightmares
The biggest nightmares for an AI engineer at 2 AM include:
- Hardware failure [00:17:55]
- An outage of the model [00:20:52]
- Cuda errors [00:20:32]
Future Directions for Software Architecture
It is suggested that the future of conversations will involve more than just one-on-one interactions with a bot, with most conversations potentially having at least two other bots present [00:04:40]. This points to a trend of increased multi-agent interaction and orchestration. There’s also an emphasis on human cognitive involvement, with a “tech hot take” suggesting that people should spend time with paper and pencil for “non-AI interrupted thoughts” [00:02:27], perhaps to generate better content for AI models that will eventually represent individuals [00:02:50].
An upcoming large-scale survey on the state of AI engineering is planned to delve deeper into tools and workflows used by AI Engineers, aiming for greater industry transparency [00:21:39].