Tony Silva🇺🇸

Self-taught since the BBS era. Fixed everything I ever touched.

Apple before it was cool. Now I watch what others miss.

AI & tech — from the other side.

[Tony Silva🇺🇸 @MacKellogs]

7 years ago, I chose this country as my home. Today the USA turns 250 — and I'm celebrating not just a nation's history, but the opportunities, the freedom to build, and the community that welcomed me. Happy birthday, America. Thank you for everything. 🇺🇸✨ #4thOfJuly #America250

[Tony Silva🇺🇸 @MacKellogs]

Apple hid a clickwheel animation in Music for the iPod's birthday.

Tap the top of your Now Playing screen ten times. Simple tribute to the device that saved the company and changed how we carried music.

The best easter eggs honor what mattered.

[Tony Silva🇺🇸 @MacKellogs]

OpenShell is NVIDIA’s attempt at a safe runtime for autonomous agents.
Each agent runs in a sandbox with file, network and process policies. It can run code without touching your actual system.
The hard part isn’t the sandbox. It’s deciding what agents should do unsupervised.

[Tony Silva🇺🇸 @MacKellogs]

Anthropic shipped Claude Duet for paired AI collaboration.

Two models working together on complex tasks. One researches, one synthesizes. Reminds me of pair programming in the BBS days.

Shipping beats theorizing.

[Tony Silva🇺🇸 @MacKellogs]

@sobanuven Deixa o Pé pra ela coitada!

[Tony Silva🇺🇸 @MacKellogs]

An $1,800 Xbox isn't really an Xbox anymore.

Microsoft reportedly exploring a high-end console tier. That's gaming PC territory. Different buyers. Different expectations.

The brand meant accessible power. This changes the contract.

[Tony Silva🇺🇸 @MacKellogs]

OpenBSD 8.0 shipped last week.

Theo and the team keep doing what they've done since 1996. Security first. Clean code. No compromises.

Some things should never change.

[Greg Joswiak @gregjoz]

Can’t wait to pro and tell next week.

[Tony Silva🇺🇸 @MacKellogs]

@neerajjj6785 Apple maps

[Tony Silva🇺🇸 @MacKellogs]

Latent space reasoning skips the token generation step entirely.

The model thinks in its own internal representations before committing to words. Faster inference, less verbose reasoning chains, closer to how we actually form thoughts.

Shipped code will tell us if it scales.

[Tony Silva🇺🇸 @MacKellogs]

@sobanuven Nossa me lembro q era um pouco maior! Mas é uma delicia!

[Tony Silva🇺🇸 @MacKellogs]

OpenAI hiring robotics engineers again after shutting down their team in 2021.

They're serious this time. Physical AI needs embodiment to truly learn. The simulators only get you so far.

Curious what form factor they choose.

[Tony Silva🇺🇸 @MacKellogs]

@SidharthVijay_ Apple!

[Tony Silva🇺🇸 @MacKellogs]

@saymon_jk Safari is S tier and so is Vivaldi !

[Tony Silva🇺🇸 @MacKellogs]

PS5 running llama.cpp is exactly the kind of hardware repurposing I love.

Someone got local LLM inference working on jailbroken PlayStation hardware. The APU has enough compute to make it worthwhile.

Consoles are just locked down PCs anyway.

[Tony Silva🇺🇸 @MacKellogs]

Needed a daily driver. Found a Roadster doing backflips.

Classic rabbit hole. You open the specs to check the range. Three hours later you’re watching a car flip through the air.

Love it or hate it, the Roadster split the world in two. The boring and the fun.

[Tony Silva🇺🇸 @MacKellogs]

Total Annihilation still holds up.

Chris Taylor's 1997 RTS introduced streaming 3D terrain, hundreds of units on screen, and a physics system that mattered. Cavedog shipped something ambitious.

The source code lives on through community projects now.

[Tony Silva🇺🇸 @MacKellogs]

@Tesla

[Tony Silva🇺🇸 @MacKellogs]

480GB of memory on a single AI accelerator card.

Intel's D270 pushes past NVIDIA's 192GB limit by roughly 2.5x. That changes what models you can load entirely in one place without sharding.

Memory bandwidth still matters but capacity unlocks new things.

[Tony Silva🇺🇸 @MacKellogs]

@MacRumors My $ 20 Walmart scale does all that and is better than Withings.

[Tony Silva🇺🇸 @MacKellogs]

LFM2.5-24B-A2B ships with 24 billion parameters optimized for agent tasks.

The A2B variant runs inference fast enough for real-time device interaction. Quantized down to 4-bit, it fits in consumer hardware without cloud dependency.

Local agents just got more practical.