Rani

ML Engineer

[Rani @RaniBaghezza]

@kimmonismus Very cool. Just missing cheap gold and scams spams in General to make it feel more real

[Rani @RaniBaghezza]

@yacineMTB Very complete playbook from HF if that helps, linking to hyperparams section https://t.co/B6y2rl1YR7

[Rani @RaniBaghezza]

Precisely why I'd tell my kids "you're smart, persistent, resilient AND hard working".

I'd go as far as remove the "smart".

Being told I was so smart that things should require no effort was destructive for a long time for me.
I saw work as something shameful. Something only stupid people did.
Took a while to understand that hard work is always necessary, no matter how smart you are.

That simple belief capped my potential for a long time, made me look for shortcuts, made me avoid thinking deeply and pushing my abilities to their limit.
You end up afraid of facing a problem you cannot solve, seeing your limits, and having that self-identity of the "smart effortless kid" shattered.

Still wonder where I'd be if I had started pushing my abilities to their limit earlier.

Be very careful with the seeds you plant in a developping brain, they have a huge impact.

[Rani @RaniBaghezza]

"This requires a new architectural approach where every business is able to build agentic systems that improve over time, while still retaining control over their IP."

AI embedded in the company from the ground up is a cool vision.
Few points I'm curious about:

- Which companies can actually do it quickly/cheaply enough to justify the investment (relative to their size/profit/time horizon)?
It requires knowing where you are going, what you are optimizing for, having clean data architecture, and clear processes.
Without a clear direction, you can't build any of this.

-How quickly can the architecture adapt to the company changing or pivoting?
Data and process evolution, restructuration, change of strategy/objectives, change of landscape.

-Seems like only very mature companies could even pretend to that kind of AI architecture,

-Seems like there is an opportunity for service companies/ESNs to provide package services for similar companies through standardized approaches (restaurants, hotels, retail, manufacturing)... Anything where the goal is very clear, and the processes are known, with very little differentiation except for branding, marketing, distribution, timing, positioning, etc...

I like this vision of the future though, always believed intelligence should be injected in all business processes for maximum efficiency. But as a freelance, I also learned that 90% of companies' internal data is a chaotic mess, so I expect data engineers to actually have a lot of work to do for this to happen.

[Rani @RaniBaghezza]

@nguyenwahed I appreciate this type of art a lot

[Rani @RaniBaghezza]

@lichthauch Noticed that too.
It's like poker: variance catches up.

Short term exploits bring immediate results, but long-term downfall.

Sticking to honesty and good character often means taking short-term Ls.
But it steers you towards long-term alignment with like-minded people.

[Rani @RaniBaghezza]

I like this training style except for bench, squat, deadlift.
More sets (3-4) help practice the pattern, and sets to failure aren't ideal.

You do hit a wall at a point where adding one more set and lowering intensity can be useful to keep making progress (or eating more)

[Rani @RaniBaghezza]

Last workout in Montreal so I did a body scan
7 months between these scans - both at exactly 191.4lbs
+7.8lbs lean mass
-7.7lbs fat mass
Body fat went from 19.8 to 15.8%

[Rani @RaniBaghezza]

Basically a 7 month experiment of high intensity/low volume training after watching Jeff Nippard's video.

Cycles of 6-7 weeks slow cut
2-3 weeks maintenance/slow bulk to recover
Ate more this past month.

4x weight lifting a week, 1-2 zwift sessions/walking

[Rani @RaniBaghezza]

It's never about "avoiding failure".
This is impossible.

It's about engineering systems where failure is never fatal.

[Rani @RaniBaghezza]

What I learned orchestrating 5 Local SLMs for NPC dialogue

[Rani @RaniBaghezza]

What I learned:
The system will always break in ways you don't expect
You can't let the game crash
You can't let NPCs get into corrupted states

Solution:
Test again and again.
Add guardrails
Think of the unthinkable

Ask yourself:
-How can it break?
-How do you recover cleanly?

[Rani @RaniBaghezza]

some kind of elvish

[Rani @RaniBaghezza]

Not sure how much IQ that is

[Rani @RaniBaghezza]

Fitting 5 agents into 3GB of RAM
-Qwen 3.5 0.8B for dialogue
-Falcon-H1-Tiny 90M for emotion/intent
Weight sharing and llama-server for inference
LoRA for fine-tuning

Showing UI, explaining current challenges and next directions
(Fine-tuning, stuttering/orchestration)

[Rani @RaniBaghezza]

Here I thought I was committing too often, so I skipped committing transient plan changes.
Turns out you really do have to commit often to make sure you don't lose too much work when this happens

[Rani @RaniBaghezza]

At least the 4th bad mistake this week from Opus 4.7 - high.
Even with medium effort, 4.6 is much better.
I really don't know what Anthropic did, first time seeing such a noticeable/critical regression.