AI Updates on 2025-06-29

Tech industry experiencing deep malaise with new grads unable to find jobs, middle managers justifying existence, and everyone not in AI wanting to transition to AI, while compensation insecurity reaches all-time highs @deedydas
Enterprise AI spending report reveals OpenAI remains the top model provider with Claude as second choice among 300 software startup executives at companies with $10M-$1B+ revenue @deedydas
Companies spend more on data storage, processing and AI infrastructure than inference and training, with AI talent being the most expensive line item @deedydas
Scaling companies at ~$500M median revenue spend approximately $100M per year across training, inference, data storage and processing @deedydas
90% of high-growth startups are either actively deploying or experimenting with AI agents @deedydas
Subscription pricing models failing for AI companies due to power users creating negative margins from LLM API costs while light users risk churning @deedydas
Coding assistance tools like Cursor and Claude lead internal productivity applications, with AI writing 33% of total code at high-growth startups @deedydas

For practical AI agent applications, problems like drift, hallucination, and compounding errors are more solvable than theoretical concerns suggest through clever prompting, tool use, constrained topics, LLM judges and organizational processes @emollick
Complex AI agent workflows can often be made to work effectively despite studies showing failures of out-of-the-box LLMs in complex use cases @emollick

Hugging Face releases FineWeb2, a new 20TB multilingual dataset supporting 1000+ languages with an adaptable data processing pipeline for any language @HuggingPapers
Open AI ecosystem analysis shows 141 different organizations contributing models and datasets, highlighting the collaborative nature of open AI development @interconnectsai
Neural network optimization success remains empirically proven despite lack of theoretical guarantees, with no mathematical reasons for why non-convex objective functions succeed in practice @Shalev_lif