Build Agents in 2025, RL tune them in 2026. GPT‑5 economics (~15x cheaper output, 2–4x faster than 4.5; per Epoch AI) signal a shift from naive scale to post‑training RL.

What does this mean for reliable, long‑horizon agents? My take: aimlbling-about.ninerealmlabs.com/blog/reinfor...

What do AI sycophancy, the Chatbot Arena, and the Pepsi Challenge have in common?

aimlbling-about.ninerealmlabs.com/blog/sycopha...

Consider the Pepsi Challenge as a parable for short-term preference (over)optimization that contributes to less reliable and trustworthy experiences.

Keeping up with AI feels like a treadmill on max speed.
I been sharing a biweekly newsletter to help my team stay sane—now I’m publishing it more broadly.
Artisanal, curated, human-written insights (no AI fluff).
Catch your breath:

aimlbling-about.ninerealmlabs.com/treadmill

Superintelligence Strategy proposes Mutual Assured AI Malfunction (MAIM) as a model for AI national security—borrowing from Mutual Assured Destruction (MAD). The problem:

AI is an arms race, not a deterrent, even though risk potential is real: aimlbling-about.ninerealmlabs.com/blog/maim-is...

The Bitter Lesson, somewhat confusingly, is a parable with an optimistic view of how machine learning and AI systems improve with scale, yet might leave us all with a bad taste in our mouths.

aimlbling-about.ninerealmlabs.com/blog/the-bit...

In 2024, I shared ~2000 AI resources (40% of which were arXiv papers!) in an internal newsletter I curate. 
No wonder staying up-to-date feels like running on a treadmill cranked all the way up!

Details in my blog: aimlbling-about.ninerealmlabs.com/blog/the-ai-...