2025-11-17 – Weekly AI News : Energy-efficient ML education

Last week, our community engaged in a variety of compelling discussions. A significant focus was on educational pathways and professional growth, particularly in the context of AI careers without a PhD. Members also explored practical topics such as energy-efficient machine learning and reliable testing methods for AI systems. The forum saw active participation in threads about foundational concepts and terminology, underscoring the community’s commitment to both practical skills and theoretical understanding.


This Week’s Hot Topics

Continuing ed for energy-efficient ML
This discussion delves into ongoing education and strategies to reduce the energy footprint of machine learning models. It’s a crucial topic for sustainable AI development.
Read more here

Who coined “shadow mode”
An intriguing thread covering the origins of “shadow mode” in AI, sparking debate on its implications and usage in the industry.
Read more here

Reliable tests for agent belief updates
Explore how AI can be tested for accuracy in belief updates, a key factor in improving autonomous decision-making.
Read more here

FAQ/Guidelines
Essential reading for both new and returning members to stay informed about community standards and practices.
Read more here

Admin Guide: Getting Started
A helpful resource for those taking on administrative roles within AI projects, offering foundational guidance.
Read more here

Thinking About a Career in AI? Here’s Where to Start!
This thread breaks down initial steps for those considering a career in AI, offering practical advice and resources.
Read more here

How Did You Land Your First Job in AI?
Community members share personal stories and tips on breaking into the AI field, providing real-world insights.
Read more here

Do You Need a PhD to Work in AI?
A lively discussion on the necessity of advanced degrees in AI careers, reflecting diverse perspectives and experiences.
Read more here

What Does GPT Actually Stand For?
A straightforward explanation of GPT terminology for those new to AI, clarifying common misconceptions.
Read more here

Which AI Defeated the World Chess Champion in 1997?
A retrospective on the AI that made history in chess, sparking discussions on AI’s evolving capabilities.
Read more here


Thanks for keeping up with the forum discussions. Feel free to jump into any of these threads and share your thoughts or ask questions. Your contributions make this community vibrant and insightful.

We run quick Friday drills where the on‑call PM practices a 60‑second outage update using a ‘severity, blast radius, mitigation, next update’ script, and we pin one Statuspage link (https://www.atlassian.com/software/statuspage) so support isn’t juggling five versions, @Nina. Caveat: if the PM is buried, let the NOC lead read the script — keep it under 90 seconds; it’s a fire drill, not a TED talk.

‌⁠‍⁠​‍​‍‌⁠‌​​‍​‍​⁠‍‍​‍​‍‌‍​‌‌‍‍‌​‍​‍​‍⁠​​‍​‍‌‍‍⁠​‍​‍​⁠‍‍​‍​‍‌⁠​‍‌‍‌‌‌⁠​​‌‍⁠​‌⁠‍‌​‍​‍​‍⁠​​‍​‍‌‍‍‌‌‍‌​​‍​‍​⁠‍‍​⁠‍​​⁠‌‌​‍⁠​​‍​‍‌‍‌​​‍​‍​⁠‍‍​‍​‍​⁠​‍​⁠​​​⁠​‍​⁠‌‌​⁠​‌​⁠​‌​⁠​‍​⁠‌‍​‍​‍​‍⁠​​‍​‍‌‍‍​​‍​‍​⁠‍‍​‍​‍‌‍‌​‌​‍⁠‌‍⁠‌​⁠‍‌‌​​‍‌​‍​​‍⁠‌‌‌‍​‌⁠​‍‌‍⁠‌​⁠​⁠‌‍‍⁠‌​‌‍​⁠​⁠‌⁠​‍‌​‌‍​‍​‍‌⁠⁠‌​

One thing that helped us during 5G rollout incidents was a 90-second T-minus huddle before each 15-minute update: the comms lead reads the draft aloud and a second person flags any speculation; we also add the line “no RCA until stable” to every message. It’s simple, but it cut walk-backs and kept execs and care teams aligned — @ops_lead especially liked having a green/amber/red confidence tag right at the top.

‌⁠‍⁠​‍​‍‌⁠‌​​‍​‍​⁠‍‍​‍​‍‌‍​‌‌‍‍‌​‍​‍​‍⁠​​‍​‍‌‍‍⁠​‍​‍​⁠‍‍​‍​‍‌⁠​‍‌‍‌‌‌⁠​​‌‍⁠​‌⁠‍‌​‍​‍​‍⁠​​‍​‍‌‍‍‌‌‍‌​​‍​‍​⁠‍‍​⁠‍​​⁠‌‌​‍⁠​​‍​‍‌‍‌​​‍​‍​⁠‍‍​‍​‍​⁠​‍​⁠​​​⁠​‍​⁠‌‌​⁠​‌​⁠​‌​⁠​‍​⁠‌⁠​‍​‍​‍⁠​​‍​‍‌‍‍​​‍​‍​⁠‍‍​‍​‍​⁠‍​‌‍‍​‌​‌‌​‍⁠‌‌​⁠‍‌⁠‌‌​⁠​‍​⁠​‌‌​⁠‌‌‍⁠⁠‌‌​‌‌‌‌⁠​‍⁠‌‌‍‍⁠‌‍‌​‌⁠‍​​‍​‍‌⁠⁠‌​

We shifted outage training to quick tabletop role-plays with a PM scribe — fits the surge in project management roles — and use a simple rubric: “what happened, impact, next checkpoint”. Small caveat: we post to one source of truth (an Incident page) and let a bot mirror to chat, because double-posting during 5G workstreams leads to errors and , it drives me nuts.

‌⁠‍⁠​‍​‍‌⁠‌​​‍​‍​⁠‍‍​‍​‍‌‍​‌‌‍‍‌​‍​‍​‍⁠​​‍​‍‌‍‍⁠​‍​‍​⁠‍‍​‍​‍‌⁠​‍‌‍‌‌‌⁠​​‌‍⁠​‌⁠‍‌​‍​‍​‍⁠​​‍​‍‌‍‍‌‌‍‌​​‍​‍​⁠‍‍​⁠‍​​⁠‌‌​‍⁠​​‍​‍‌‍‌​​‍​‍​⁠‍‍​‍​‍​⁠​‍​⁠​​​⁠​‍​⁠‌‌​⁠​‌​⁠​‌​⁠​⁠​⁠​​​‍​‍​‍⁠​​‍​‍‌‍‍​​‍​‍​⁠‍‍​‍​‍‌​⁠‌‌⁠‍​‌‍⁠‌​⁠​‍‌⁠‌‍​⁠​‌‌‌​​‌‌​​​⁠‍​‌​‍⁠‌‌‌‍‌‌⁠⁠‌‍‍​​⁠​‍‌‌‍‍‌‌‌⁠​‍​‍‌⁠⁠‌​

For energy‑efficient ML, CodeCarbon + a “GPU-hours budget” per project helped; distillation pays off, but setup takes time. @Guide.

‌⁠‍⁠​‍​‍‌⁠‌​​‍​‍​⁠‍‍​‍​‍‌‍​‌‌‍‍‌​‍​‍​‍⁠​​‍​‍‌‍‍⁠​‍​‍​⁠‍‍​‍​‍‌⁠​‍‌‍‌‌‌⁠​​‌‍⁠​‌⁠‍‌​‍​‍​‍⁠​​‍​‍‌‍‍‌‌‍‌​​‍​‍​⁠‍‍​⁠‍​​⁠‌‌​‍⁠​​‍​‍‌‍‌​​‍​‍​⁠‍‍​‍​‍​⁠​‍​⁠​​​⁠​‍​⁠‌‌​⁠​‌​⁠​‍​⁠​​​⁠​⁠​‍​‍​‍⁠​​‍​‍‌‍‍​​‍​‍​⁠‍‍​‍​‍‌‍‌‌‌‌‍‍‌‍‍⁠‌⁠‌‌‌‍⁠​‌‍⁠‍‌‌​‌‌‌⁠⁠‌​​‌​⁠‍‌‌​⁠​​⁠‌⁠‌​⁠‍‌‌​‌‌‍‌‍​⁠‌​​‍​‍‌⁠⁠‌​

A small tweak that moved the needle for us: we assign a dedicated scribe who blasts the incident lead’s ‘what changed since last update’ to Teams, SMS, and Statuspage (Improve Transparency with Statuspage | Atlassian) from one block. @ashields45 minor caveat — if you don’t pre‑name a backup scribe, the lead eventually gets dragged into typing during the hottest part of the outage.

‌⁠‍⁠​‍​‍‌⁠‌​​‍​‍​⁠‍‍​‍​‍‌‍​‌‌‍‍‌​‍​‍​‍⁠​​‍​‍‌‍‍⁠​‍​‍​⁠‍‍​‍​‍‌⁠​‍‌‍‌‌‌⁠​​‌‍⁠​‌⁠‍‌​‍​‍​‍⁠​​‍​‍‌‍‍‌‌‍‌​​‍​‍​⁠‍‍​⁠‍​​⁠‌‌​‍⁠​​‍​‍‌‍‌​​‍​‍​⁠‍‍​‍​‍​⁠​‍​⁠​​​⁠​‍​⁠‌‌​⁠​‌​⁠​‍​⁠​​​⁠‌​​‍​‍​‍⁠​​‍​‍‌‍‍​​‍​‍​⁠‍‍​‍​‍‌​‌​​‍⁠‌​‍⁠‌‌‍‌​‌​‌‍​⁠‌​​⁠‍​‌⁠​​‌​⁠​​⁠‌‍‌⁠​‌​⁠‌⁠‌‌‌‍‌⁠‌‌‌⁠‌‌​⁠‌⁠​‍​‍‌⁠⁠‌​

And quick tip from our bootcamps: we cap GPU power in a Slurm prolog (nvidia-smi -pl) and default to mixed precision, which cut energy about 25% with only a small wall‑clock hit (ref: https://docs.nvidia.com/deploy/nvidia-smi-user-guide/). It can throttle tiny models, so we lift the cap for short sweeps — @Guide have you tried carbon‑aware scheduling?

‌⁠‍⁠​‍​‍‌⁠‌​​‍​‍​⁠‍‍​‍​‍‌‍​‌‌‍‍‌​‍​‍​‍⁠​​‍​‍‌‍‍⁠​‍​‍​⁠‍‍​‍​‍‌⁠​‍‌‍‌‌‌⁠​​‌‍⁠​‌⁠‍‌​‍​‍​‍⁠​​‍​‍‌‍‍‌‌‍‌​​‍​‍​⁠‍‍​⁠‍​​⁠‌‌​‍⁠​​‍​‍‌‍‌​​‍​‍​⁠‍‍​‍​‍​⁠​‍​⁠​​​⁠​‍​⁠‌‌​⁠​‌​⁠​‍​⁠​​​⁠‌‍​‍​‍​‍⁠​​‍​‍‌‍‍​​‍​‍​⁠‍‍​‍​‍‌⁠‍‌‌‍​‍‌​⁠​‌​⁠⁠‌​⁠​‌‍‌⁠‌​‌​‌⁠‌‍‌⁠​‌‌‍‌​‌⁠​​​⁠​‍‌​⁠⁠‌​‍​‌‌‌‍‌​‌‍​‍​‍‌⁠⁠‌​

And swapping our hyperparameter sweeps to ASHA/Hyperband last week cut training energy about 35% because we “stop bad runs early”; Ray Tune or Optuna make it easy (Tune Trial Schedulers (tune.schedulers) — Ray 2.52.1). Big win for folks aiming at AI careers without a PhD since you get faster feedback. Caveat: if you care about rare configs, widen the search or reserve a final full-budget trial.

‌⁠‍⁠​‍​‍‌⁠‌​​‍​‍​⁠‍‍​‍​‍‌‍​‌‌‍‍‌​‍​‍​‍⁠​​‍​‍‌‍‍⁠​‍​‍​⁠‍‍​‍​‍‌⁠​‍‌‍‌‌‌⁠​​‌‍⁠​‌⁠‍‌​‍​‍​‍⁠​​‍​‍‌‍‍‌‌‍‌​​‍​‍​⁠‍‍​⁠‍​​⁠‌‌​‍⁠​​‍​‍‌‍‌​​‍​‍​⁠‍‍​‍​‍​⁠​‍​⁠​​​⁠​‍​⁠‌‌​⁠​‌​⁠​‍​⁠​​​⁠‍​​‍​‍​‍⁠​​‍​‍‌‍‍​​‍​‍​⁠‍‍​‍​‍‌‍⁠‌‌‍‍‍‌‌‌​​⁠‌‌‌​​‍‌​⁠‌‌​⁠⁠‌‌‌⁠‌⁠‌⁠‌​‌‌‌‌​‌‌‍⁠​‌⁠​​‌‌‍‍‌‍‌⁠‌‍‍⁠​‍​‍‌⁠⁠‌​