umar

Coder, Writer, Opinion haver.

[umar @Umar_init_]

@yc989 They can’t be independent persons with opinions?

[umar @Umar_init_]

@_Felipe Congratulations 🙌 a breath of fresh air to see the results.

[umar @Umar_init_]

@gitshadan Acha

[Miguel de Icaza ᯅ🍉 @migueldeicaza]

These bans deputize random busybodies into hall monitors, replacing communities built by people who contribute with hierarchies ruled by people who have nothing to offer but the power to stop others.

[umar @Umar_init_]

@avrldotdev OIDC.

[umar @Umar_init_]

@JohnGreenDev Adversarial testing, honestly. Running everything that can go wrong will go wrong type of E2E penetration testing. Can give an LLM all the APIs and tell it to cause the program to fail. Failures can then be made deterministic tests.

[umar @Umar_init_]

@tanejagarvit_ What’s your activity manager showing?

[umar @Umar_init_]

@tanejagarvit_ That’s the killer I think. See if you can reduce dockerized containers in itself, they eat a buttload of RAM lol

[umar @Umar_init_]

@tanejagarvit_ are these dockerized backend services?

[smallest.ai @smallest_AI]

Eminem's fastest verse vs Smallest Pulse STT

The last verse of "Godzilla" is one of the fastest in mainstream rap. Pulse kept up with all of it: 296 words, 1.4% word error rate.

Just the raw audio, transcribed live.

If Pulse can handle Eminem, your customer calls are easy.

[umar @Umar_init_]

Don’t really understand European tech doomers

[solst/ICE of Astarte @IceSolst]

Many teams SUCK at making software, and “patch everything” is so short-sighted.

Vulnerabilities are a symptom of a larger problem. We’ve been getting away with a lack of engineering rigor for a long time (decades) in most industries.

We’re not making resilient software. We will continue to have issues until most orgs mature how they build, no matter how many times a day you patch.

It’s been a financial decision to prioritize shipping fast, over minimizing brittleness.
As exploitability gets cheaper, it drives up response and maintenance costs.

This may finally incentivize orgs to invest more in software resilience.

And we have the tools: each dev team can now spend tokens analyzing their own code and deployment for security and availability issues. Automate testing, automate dependency upgrades, hold design review and threat modeling sessions.

Slop is what you make it if you’re not investing in mature architectural decisions. And we’ll keep having trivial vulns until you do so.

[umar @Umar_init_]

@LewisCTech some nuggets of wisdom and/or recommended reads?

[umar @Umar_init_]

How will you implement HLD wo knowing LLD basics? Otherwise you’re gonna have lots of race conditions. Learnt this the hard way while working on a marketplace project myself. If I didn’t properly check the code, full gutter

[Gaurav Sen @gkcs_]

Low-Level Design is dead.

AI has made coding obsolete. No one cares about the programming language we use.

What matters is engineering: making a system reliable, cheap, and scalable.

In the past 3 months, I have worked on dozens of side projects and built custom frameworks for myself.

I hadn't written code professionally since 2019, and yet I was blazing through projects in 2026.

My system design fundamentals helped. My business fundamentals also helped.

Coding didn't. I haven't written a single line myself.
In 2026, code is all generated.

If you are a software engineer, you must transition to AI now.

All the best, my brethren.

[umar @Umar_init_]

@shubh6200 It’s both. You can only distribute so much horizontally until you’re hit with inference limits and models’ own capacity limits. What will you do then? 5 GPUs serving worse model < 3 GPUs serving better model w inf optimizations.

[umar @Umar_init_]

@gitshadan You’re getting hit on bruh

[umar @Umar_init_]

@ZackKorman hi Zack, can you check your DMs?

[f4mi ‼️ @f4micom]

BABY CRYING COUNTER inside a TV spying on you ffs this shit has to be regulated out of existence

[umar @Umar_init_]

GOATed

[lyra @_lyraaaa_]

if you have a qwen3.5 4b around, please try steering neurons L14 N8949 +1 and L13 N7956 -1
now you have evil qwen

recipe: dose × median(‖resid‖ leaving layer L, tokens 1:) / ‖W_down[:, n]‖` at the down_proj at every input