/Tech12h ago

Google DeepMind's Susan Zhang warns that routine ML tasks like hyperparameter tuning and basic scripting face automation within a year

Post-training techniques will also increasingly automate post-training workflows.

315911922157.1K

#264

Original post

Susan Zhang@suchenzang#306inTech

if your bread-and-butter consists solely of:

- tuning hyperparams/config files

- fitting points on a log-log plot

- tweaking a few lines in http://model.py, http://transformer.py, http://optimizer.py, http://train.py

- waiting a week for <= 512 chips to free up and then another week for loss curves to converge

it is completely understandable to be stressed about becoming automated into irrelevance within the next year or so.

question is, do you wait for that to happen, or do you start doing something differently now?

6:00 AM · Jun 7, 2026 · 48.5K Views

/Tech12h ago

Google DeepMind's Susan Zhang warns that routine ML tasks like hyperparameter tuning and basic scripting face automation within a year

Post-training techniques will also increasingly automate post-training workflows.

315911922157.1K

#264

Original post

Susan Zhang@suchenzang#306inTech

if your bread-and-butter consists solely of:

- tuning hyperparams/config files

- fitting points on a log-log plot

- tweaking a few lines in http://model.py, http://transformer.py, http://optimizer.py, http://train.py

- waiting a week for <= 512 chips to free up and then another week for loss curves to converge

it is completely understandable to be stressed about becoming automated into irrelevance within the next year or so.

question is, do you wait for that to happen, or do you start doing something differently now?

6:00 AM · Jun 7, 2026 · 48.5K Views

Sentiment

Many users reacted with distress to warnings that AI advances will automate routine ML engineering tasks, expressing pessimism about job impacts like ads ranking while one noted relief at doing more advanced work.

Pos

12.5%

Neg

87.5%

8 comments with sentiment.

Cluster Engagement

Posts from X

Most Activity

VIEWS1.8K

Adam@aschiff19

@suchenzang When people let AI make their decisions for them, solutions continue to get built for yesterday’s problems…

here in the Triangle I find AI uptake to be at a slower clip, but companies here are focused on problem-founder fit and using AI as a tool, not racing against AI progress

9h1.8K

BOOKMARKS1LIKES7

Elena Dyachkova@ElenaRunsNYC

@suchenzang This is most everyone in meta ads ranking ai 😭

11h78571

REPLIES2

Shane Gu@shaneguML

imo the metric isn't impact, it's impact / compute — however GPU-rich you are.

Susan Zhang@suchenzang

if your bread-and-butter consists solely of:

- tuning hyperparams/config files

- fitting points on a log-log plot

- tweaking a few lines in http://model.py, http://transformer.py, http://optimizer.py, http://train.py

- waiting a week for <= 512 chips to free up and then another week for loss curves to converge

it is completely understandable to be stressed about becoming automated into irrelevance within the next year or so.

question is, do you wait for that to happen, or do you start doing something differently now?

5h61030

Aeron@aeronxbt

@madiator so the output is just. more post training input?

8h147

Draven@notdrvx

@madiator double post training for that extra loop feel

is there a word for this or did we just make it up?

7h121

ëugene kharitonov 🏴‍☠️@n0mad_0

@suchenzang Ah, the http://transformer.py

9h6621

Travis Davies 🇦🇺@travisddavies

@suchenzang Winter was always coming for those that atrophy

11h890

Yash@yash1_

@suchenzang Just enjoying the " beginning of the end "

12h794

RBA@benalt

@suchenzang Ouch.

9h620

neon@neonnbits

@suchenzang I’m happy i do much more than this

11h503

Patrick Baitman@taofanqq

@suchenzang Recursive self-improvement will fit those points better than u anon

7h481

phil@big_algocracy

@suchenzang aggressively deepmind coded post (no shade)

7h1211

Roman Vorushin@vorushin

@suchenzang No LLMs can drink Flat White as I can!

10h249

gandamu@gandamu_ml

@suchenzang do it across several domains simultaneously so that more things magically make sense over time and it feels like progress. The delusion helps until I find out there's no edge because everyone else is doing the same thing.

4h196